Chapter 2 REVIEW OF PREVIOUS WORK

Chapter 2 REVIEW OF PREVIOUS WORK

2.1 Introduction Automatic Speech Recognition (ASR) is a critical core technology in the field of intelligence communication between human and machine. Despite the long history of research on the acoustic characteristics of Vowel / Consonant-Vowel (V/CV) unit waveforms [Sakai and Doshita(1963)], [Schaffer and Rabiner(1970)], [D. Dutta Majumder and Pal(1976)], [Broad(1972)], current state-of-the-art ASR systems are still incapable of performing accurate recognition for these class of sounds . Beginning 1910, Campell and Crandall from AT & T and Western Electric Engineering initiated a series of experiments to explore the nature of human speech perception. After this in 1918, these experiments were continued by Fletcher and his colleagues at The Bell Telephone Laboratories (Western Electric Engineering until 1925). These studies lead to a speech recognition measure called articulation index, which accurately characterizes speech intelligibly under condition of filtering and noise. All these experiments began with normal conversational speech over a modified telephone channel [Fletcher(1922)] [Fletcher and Munson(1937)] [Fletcher and Galt(1950)]. In the 1930’s Homer

14

Dudley, influenced greatly by Fletcher’s research, developed a speech synthesizer called the VODER (Voice Operating Demonstrator), which was an electrical equivalent (with mechanical control) of Wheatstone’s mechanical speaking machine [H. Dudley and Watkins(1939)]. These two speech poineers thoroughly established the importance of the signal spectrum for reliable identification of the phonetic nature of a speech sound. In 1940’s Dudley had developed mathematical models for speech based on linguistic research that viewed spoken language with the impulses from larynx and the vocal folds as the input to the system, the shape of the vocal tract representing the filter parameter and the speech waveform as the system output [Dudley(1940)].

In 1952,Davis et al., of Bell Laboratories built a system for isolated digit recognition for a single speaker , using the formant frequencies measured (or estimated) from vowel regions of each digit. These trajectories served as the reference pattern for determining the identity of an unknown digit utterance as the best matching digit [K. H. Davis and Balashek(1952)]. In the 1960’s, several Japanese laboratories established their capability of building special purpose hardware to achieve a speech recognition task. Most important were the vowel recognizer of Suzuki and Nakata at the Radio Research Lab in Tokyo [Suzuki and Nakata(1961)], the phoneme recognizer of Sakai and Doshita at Kyoto [Sakai and Doshita(1962)], and the digit recognizer of NEC Laboratories [K. Nagata and Chiba(1963)]. The work of Sakai and Doshita involved the first use of a speech segmenter for analysis and recognition of speech in different portions of the input utterance.

15

In another early recognition system Fry and Denes, at University College in England, built a phoneme recognizer to recognize 4 vowels and 9 consonants [Fry and Denes(1959)]. Including statistical information about allowable phoneme sequences in English, they increased the overall phoneme recognition accuracy for words consisting of two or more phonemes. This work marked the first use of statistical syntax (at the phoneme level) in automatic speech recognition. An alternative to the use of a speech segmenter was the concept of adopting a nonuniform time scale for aligning speech patterns. This concept started to gain acceptance in the 1960’s through the work of Tom Martin at RCA Laboratories [T. B. Martin and Zadell(1964)] and Vintsyuk in the Soviet Union [Vintsyuk(1968)]. Martin recognized the need to deal with the temporal non-uniformity in repeated speech events and suggested a range of solutions, including detection of utterance endpoints, which greatly enhanced the reliability of the recognizer performance. Vintsyuk proposed the use of dynamic programming for time alignment between two utterances in order to derive a meaningful assessment of their similarity [Vintsyuk(1968)]. His work, though largely unknown in the West, appears to have preceded that of Sakoe and Chiba as well as others who proposed more formal methods [Sakoe and Chiba(1978)], generally known as dynamic time warping, in speech pattern matching. Since the late 1970’s,mainly due to the publication by Sakoe and Chiba, dynamic programming, in numerous variant forms (including the Viterbi algorithm which came from the communication theory community), has become an indispensable technique in automatic speech recognition [Viterbi(1967)].

16

In the late 1960’s, Atal and Itakura independently formulated the fundamental concepts of Linear Predictive Coding (LPC) , which greatly simplified the estimation of the vocal tract response from speech waveforms [Atal and Hanauer(1971)] [Itakura and Saito(1970)]. By the mid 1970’s, the basic ideas of applying fundamental pattern recognition technology to speech recognition, based on LPC methods, were proposed by Itakura [Itakura(1975)], Rabiner and Levinson and others [L. R. Rabiner and Wilpon(1979)]. In the early 1980’s at Bell Laboratories, the theory of HMM was extended to mixture densities which have since proven vitally important in ensuring satisfactory recognition accuracy, particularly for speaker independent, large vocabulary speech recognition tasks [Juang(1985)] [B. H. Juang and Sondhi(1986)] . Another technology that was (re)introduced in the late 1980’s was the idea of artificial neural networks (ANN). Neural networks were first introduced in the 1950’s, but failed to produce notable results initially [McCullough and Pitts(1943)]. The advent, in the 1980’s, of a parallel distributed processing (PDP) model, which was a dense interconnection of simple computational elements, and a corresponding training method, called error backpropagation, revived interest around the old idea of mimicking the human neural processing mechanism [Lippmann(1987)] [Kohonen(1988)] [Pal and Mitra(1988)].

In the 1990’s, a number of inventions took place in the field of pattern recognition. The problem of pattern recognition, which traditionally followed the framework of Bayes and required estimation of distributions for the data, was transformed into an optimization problem involving minimization of the empirical recognition error [Juang(1985)]. This fundamental change of paradigm was caused by the recognition of the fact that the distribution functions for the speech 17

signal could not be accurately chosen or defined, and that Baye’s decision theory would become inapplicable under these circumstances. After all, the objective of a recognizer design should be to achieve the least recognition error rather than the best fitting of a distribution function to the given (known) data set as advocated by the Bayes criterion. The concept of minimum classification or empirical error subsequently spawned a number of techniques, among which discriminative training and kernel-based methods such as the support vector machines (SVM) have become popular subjects of study [B.H. Juang and Chou(1997)] [Vapnik(1998)].

This chapter presents a review of previous works in the area of linear and nonlinear speech processing, multi resolution analysis and wavelet transform, neural networks and statistical learning algorithms and is organized as follows. Section 2.2, section 2.3 and section 2.4 provides a summary of research findings in the area of traditional speech processing, wavelet transform and nonlinear speech processing respectively. Section 2.5 gives a review of previous works in the applications of neural network for speech recognition and section 2.6 contains review of previous work in the application of k-Nearest Neighborhood (k-NN) and SVM. Finally section 2.7 concludes this review.

2.2 Review on Traditional Features for Speech Recognition Linear Predictive Coding (LPC) and Mel Frequency Cepstral Coefficient (MFCC) features are known to be traditional basic speech features in the sense that the speech model used in many of these applications is the source-filter model which represent the vocal characteristics of the speech signal. The linear prediction (LP)

18

model for speech analysis and synthesis was first introduced by Saito and Itakura and Atal and Schroeder [Itakura and Saito(1970)] [Atal and Schroeder(1967)]. Saito and Itakura at NTT, Japan, developed a statistical approach for the estimation speech spectral density using a maximum likelihood method [Itakura and Saito(1970)]. Their work was originally presented at conferences in Japan and therefore, was not known worldwide. The theoretical analysis behind their statistical approach were slightly different than that of linear prediction, but the overall results were identical. Based on their statistical approach, Itakura and Saito introduced new speech parameters such as the partial autocorrelation (PARCOR) coefficients for efficient encoding of linear prediction coefficients. Later, Itakura discovered the line spectrum pairs, which are now widely used in speech coding applications.

In 1975, John Makhoul presented a tutorial review on Linear Predictive Coding [Makhoul(1975)], which gives an exposition of linear prediction in the analysis of discrete signals. The major part of this paper is devoted to all-pole models. The model parameters are obtained by a least squares analysis in the time domain. Two methods resulted, depending on whether the signal is assumed to be stationary or non stationary. The same results are then derived in the frequency domain also. The resulting spectral matching formulation allows for the modeling of selected portions of a spectrum for arbitrary spectral shaping in the frequency domain, and for the modeling of continuous as well as discrete spectra.

The method of linear prediction has proved quite popular and successful for use in speech compression system [Markel and A.H.Gray(1974)] [Itakura(1972)] 19

[Atal and Hanauer(1971)]. An efficient method for transmitting the linear prediction parameters has been found by Sambur using the techniques of differential PCM [Sambur(1975)]. Using this technique, speech transmission is employing fewer than 1500 bits/second. Further reduction in the linear prediction storage requirements can be realized at a cost of higher system complexity by transmission of the most significant eigenvectors of the parameters. It has been found that this technique in combination with differential PCM can lower the bitrate to 1000 bits/sec. Sambur and Jayant discusses several manipulations of LPC parameters for providing speech encryption [Sambur and Jayant(1976)]. They considers temporal rearrangement or scrambling of the LPC code sequence, as well as the alternative of perturbing individual samples in the sequence by means of pseudorandom additive or multiplicative noise. The latter approach is believed to have greater encryption potential than the temporal scrambling technique, in terms of time needed to break the security code. The encryption technique are assessed on the basis of perceptual experiments, as well as by means of qualitative assessment of speech spectrum distortion, as given by an appropriate distance measure.

2.3 Review on Wavelet Transform for Speech recognition Over the last decades, wavelet analysis have become very popular and new interests are emerging in this topic. It has turned to be a standard technique in the ‘area of geophysics, meteorology, audio signal processing and image compression [Hongyu Liao and Cockburn(2004)], [Soman and Ramachandran(2005)], [Mallat(2009)]. Wavelet Transform is a tool for Multi Resolution Analysis which can be used to efficiently represent the speech signal in the time-frequency plane. 20

Martin Vetterli [Vetterli(1992)] had compared the wavelet transform with the more classical short-time Fourier transform approach to signal analysis. In addition he also pointed out the strong similarities between the details of these techniques. Gianpaolo [Evangelista(1993)] explored a new wavelet representation using the transform based on a pitch-synchronous vector representation and its adaptation to the oscillatory or aperiodic characteristics of signals. Pseudo-periodic signals are represented in terms of an asymptotically periodic trend and aperiodic fluctuations at several scales. The transform reverts to the ordinary wavelet transform over totally aperiodic signal segments. The pitch-synchronous wavelet transform is particularly suitable to the analysis, rate-reduction in coding and synthesis of speech signals and it may serve as a preprocessing block in automatic speech recognition systems. Separation of voice from noise in voiced consonants is easily performed by means of partial wavelet expansions.

In [Xia and Zhang(1993)], the authors studied the properties of Cardinal Orthogonal Scaling Functions (COSF), which provide the standard sampling theorem in multiresolution spaces with scaling functions as interpolants. They presented a family of COSF with exponential decay, which are generalizations of the Haar functions. With these COSF, an application is the computation of Wavelet Series Transform (WST) coefficients of a signal by the Mallat algorithm. They also presented some numerical comparisons for different scaling functions to illustrate the advantage of COSF. For signals which are not in multiresolution spaces, they estimated the aliasing error in the sampling theorem by using uniform samples.

21

In [M Lang and R.O.Wells(1996)] Lang et al., had presented a new nonlinear noise reduction method using discrete wavelet transform.They employed thresholding in the wavelet transform domain following a suggestion by Coifman, using undecimated, shift-invariant, nonorthogonal wavelet transform instead of the usual orthogonal one. This approach can be interpreted as a repeated application of the original Donoho and Johnstone method for different shifts. The main feature of this algorithm is a significantly improved noise reduction compared to the original wavelet based approach. This holds for a large class of signals, and is shown theoretically as well as by experimental results.

A multi wavelet design criterion known as omnidirectional balancing using wavelet transform is introduced by James E. Fowler and Li Hua to extend to vector transforms the balancing philosophy previously proposed for multiwavelet based scalar-signal expansion [Fowler and Hua(2002)]. It is shown that the straightforward implementation of a vector wavelet transform, namely, the application of a scalar transform to each vector component independently, is a special case of an omnidirectionally balanced vector wavelet transform in which filter-coefficient matrices are constrained to be diagonal. Additionally, a family of symmetricantisymmetric multiwavelets is designed according to the omnidirectional balancing criterion. In empirical results for a vector-field compression system, it is observed that the performance of vector wavelet transforms derived from these omnidirectionally balanced symmetric-antisymmetric multiwavelets is far superior to that of transforms implemented via other multiwavelets and can exceed that of diagonal transforms derived from popular scalar wavelets.

22

O Farooq and S Dutta [Farooq and S.Datta(2004)]had proposed a subband feature extraction technique based on an admissible wavelet transform and the features are modified to make them robust to Additive White Gaussian Noise (AWGN). The performance of this system is compared with the conventional mel frequency cepstral coefficients (MFCC) under various signal to noise ratios. The recognition performance based on the eight sub-band features is found to be superior under the noisy conditions compared with MFCC features using this approach.

Elawakdy et al., proposed a speech recognition algorithm using wavelet transform. This paper discussed the combination of a feature extraction by wavelet transform, subtractive clustering and adaptive neuro-fuzzy inference system (ANFIS). The feature extraction is used as input of the subtractive clustering to put the data in a group of clusters. Also it is used as an input of the neural network in ANFIS. The initial fuzzy inference system is trained by the neural network to obtain the least possible error between the desired output (target) and the fuzzy inference system (FIS) output to get the final FIS. The performance of the proposed speech recognition algorithm (SRA) using a wavelet transform and ANFIS is evaluated by different samples of speech signals of isolated words with added background noise. The proposed speech recognition algorithm is tested using different isolated words obtaining a recognition ratio about 99%.

A multi resolution hidden markov model using class specific features is proposed by Baggenstoss [Baggenstoss(2010)]. He applied the PDF projection theorem to generalize the hidden Markov model (HMM) to accommodate multiple 23

simultaneous segmentations of the raw data and multiple feature extraction transformations. Different segment sizes and feature transformations are assigned to each state. The algorithm averages over all allowable segmentations by mapping the segmentations to a "proxy" HMM and using the forward procedure. A by-product of the algorithm is the set of a posteriori state probability estimates that serve as a description of the input data. These probabilities have simultaneously both the temporal resolution of the smallest processing windows and the processing gain and frequency resolution of the largest processing windows. The method is demonstrated on the problem of precisely modeling the consonant "T" in order to detect the presence of a distinct "burst" component. He compared the algorithm against standard speech analysis methods using data from the TIMIT Speech Database.

In short, from the literature survey it is shown that there is an emerging research trends in the study of application of Multi Resolution Analysis using Wavelet Transform for the human speech recognition for the fast few years. Thus Malayalam V/CV speech unit recognition using MRA based Wavelet Transform is of great importance to capture the non-stationary nature of the speech signal.

2.4 Review on Non-Linear Dynamical System Models for Speech Recognition Nonlinear speech processing is a rapidly growing area of research. Naturally, it is difficult to define a precise date for the origin of the field, but it is clear that there was a rapid growth in this area, which started in the mid-nineteen eighties. 24

Since that time, numerous techniques were introduced for nonlinear time series analysis, which are ultimately aimed at engineering applications.

Among the nonlinear dynamics community, a budding interest has emerged in the application of theoretical results to experimental time series data analysis in 1980’s. One of the profound results established in chaos theory is the celebrated Takens’ embedding theorem. Takens’ theorem states that under certain assumptions, phase space of a dynamical system can be reconstructed through the use of time-delayed versions of the original scalar measurements. This new state space is commonly referred to in the literature as Reconstructed State Space (RSS), and has been proven to be topologically equivalent to the original state space of the dynamical system.

Packard et al., first proposed the concept of phase space reconstruction in 1980 [Packard.N.H and Shaw.R.S(1980)]. Soon after, Takens showed that a delay coordinate mapping from a generic state space to a space of higher dimension preserves topology [Takens(1980)]. Sauer and Yorke have modified Taken’s theorem to apply for experimental time series data analysis [Sauer.T and Casdagli.M(1991)].

Conventional linear digital signal processing techniques often utilize the frequency domain as the primary processing space, which is obtained through the Discrete Fourier Transform (DFT) of a time series. For a linear dynamical system, representation of the signal appears in the frequency domain that takes the form of sharp resonant peaks in the spectrum. However for a nonlinear or chaotic system, the signal representation does not appear in the frequency domain, because the 25

spectrum is usually broadband and resembles noise. In the RPS, a signal representation emerges in the form of complex, dense orbits that form patterns known as attractors. These attractors contain the information about the time evolution of the system, which means that features derived from a RPS can potentially contain more or different information.

The majority of literature that utilizes a RSS for signal processing applications revolves around its use for control, prediction, and noise reduction, reporting both positive and negative results. There is only scattered research using RPS features for classification and /or recognition experiments.

In contrast to the linear source-filter model for speech production process, a large number of research works are reported in the literature to show the nonlinear effects in the physical process. Koizumi.T, Taniguchi.S et al., in 1985 showed that the vocal tract and the vocal folds do not function independently of each other, but that, there is in fact some form of coupling between them when the glottis is open [Koizumi.T and Hiromitsu.S(1985)]. This can cause significant changes in formant characteristics between open and closed glottis cycles [Brookes.D.M and Naylor.P.A(1988)].

Teager and Teager [Teager.S.M(1989)] have claimed that voiced sounds are characterised by highly complex airflows in the vocal tract, rather than well behaved laminar flow. Turbulent flow of this nature is also accepted to occur during unvoiced speech, where the generation of sound is due to a constriction at some point in the vocal tract. In addition, the vocal folds will themselves be responsible 26

for further nonlinear behaviour, since the muscle and cartilage, which comprise the larynx, have nonlinear stretching qualities [Fletcher and Munson(1937)].

Such non-linearities are routinely included in attempts to model the physical process of vocal fold vibration, which have focused on two or more mass models [Fletcher and Galt(1950)], [H. Dudley and Watkins(1939)], [Dudley(1940)] in which the movement of the vocal folds is modeled by masses connected by springs, with nonlinear coupling. Observations of the glottal waveform reinforce this evidence, where it has been shown that this waveform can change shape at different amplitudes [Schoentgen(1990)]. Such a change would not be possible in a strictly linear system where the waveform shape is unaffected by amplitude changes.

Extraction of invariant parameters from speech signal has attracted researchers for designing speech and speaker recognition systems. In 1988, Narayanan.N.K. and Sridhar C.S. [Narayanan.N.K and Sridhar.C.S(1988)] used the dynamical system technique mentioned in the nonlinear dynamics to extract invariant parameters from speech signal. The dynamics of speech signal is experimentally investigated by extracting the second order dimension of the attractor D2 and the second order Kolmogorov entropy K2 of speech signal. The fractal dimension of D2 and non-zero value of K2 confirms the contribution of deterministic chaos to the behavior of speech signal. The attractor dimension D2 and Kolmogorov entropy K2 are then used as a powerful tool for voiced / unvoiced classification of speech signals.

27

The dimension of the trajectories, or the dimension of the attractor is an important characteristic of the dynamic systems. The estimation of the dimension gives a lower bound of the number of parameters needed in order to model the system. The goal is to find if the system under study occupies all the state space or if it is most of the time in a subset of the space, called attractor. The correlation dimension [Tishby.N(1990)] is a practical method to estimate the dimension of an empirical temporal series.

There are a large variety of techniques found in the literature of nonlinear methods and it is difficult to predict which techniques ultimately will be more successful in speech processing. However, commonly observed methods in the speech processing literature are various forms of oscillators and nonlinear predictors, the latter being part of the more general class of nonlinear autoregressive methods. The oscillator and autoregressive techniques themselves are also closely related since a nonlinear autoregressive model in its synthesis form, forms a nonlinear oscillator if no input is applied. For the practical design of a nonlinear autoregressive model, various approximations have been proposed [Farmer.J.D and Sidorowich.J.D(1988)] [Casdagli.M and Gibson.J(1991)] [Abarbanel.H.D.I and Tsimring.L.S(1993)] [Kubin.G(1995)]. These can be split into two main categories: parametric and non parametric methods.

Phase space reconstruction is usually the first step in the analysis of dynamical systems. An experimenter obtains a scalar time series from one observable of a multidimensional system. State-space reconstruction is then needed for the indirect measurement of the system’s invariant parameters like, dimension, Lyapunov 28

exponent etc. Takens’ theorem gives little guidance, about practical considerations for reconstructing a good state space. It is silent on the choice of time delay (τ) to use in constructing m-dimensional data vectors. Indeed, it allows any time delay as long as one has an infinite amount of infinitely accurate data. However, for reconstructing state spaces from real-world, finite, noisy data, it gives no direction [Casdagli.M and Gibson.J(1991)]. Two heuristics have been developed in the literature for establishing a time lag [Kantz and Schreiber.T(2003)]. First one is the first zero of the autocorrelation function and the second one is the first minimum of the auto mutual information curve [Fraser.A.M and Swinney.H.L(1986)]. Andrew M Fraser and Harry L Swinney reported in their work that the mutual information is examined for a model dynamical system and for chaotic data from an experiment on the Belousov-Zhabotinskii reaction. An N log N algorithm for calculating mutual information (I) is presented. A minimum in ’I’ is found to be a good criterion for the choice of time delay in Phase Space Reconstruction from time series data. This criterion is shown to be far superior than choosing a zero of the autocorrelation function.

There have been many discussions on how to determine the optimal embedding dimension from a scalar time series based on Taken’s theorem or its extensions [Sauer.T and Casdagli.M(1991)]. Among different geometrical criteria, the most popular seems to be the method of False Nearest Neighbors. This criterion concerns the fundamental condition of no self-intersections of the reconstructed attractor [Kennel.M.B and Abarbanel.H.D.I(1992)].

29

Work by Banbrook, McLaughlin et al., [Banbrook and McLaughlin(1994)] Kumar et al.,

[Kumar.A and Mullick.S.K(1996)]

and Narayanan et al.,

[Narayanan.S.S and Alwan.A.A(1995)] has attempted to use nonlinear dynamical methods to answer the question: "Is speech chaotic?" These papers focused on calculating theoretical quantities such as Lyapunov exponents and Correlation dimension. Their results are largely inconclusive and even contradictory. A synthesis technique for voiced sounds is developed by Banbrook et al., inspired by the technique for estimating the Lyapunov exponents.

In a work presented by Langi and Kinsner [Langi.A and Kinsner.W(1995)], speech consonants are characterised by using a fractal model for speech recognition systems . Characterization of consonants has been a difficult problem because consonant waveforms may be indistinguishable in time or frequency domain. The approach views consonant waveforms as coming from a turbulent constriction in a human speech production system, and thus exhibiting turbulent and noise like time domain appearance. However, it departs from the usual approach by modeling consonant excitation using chaotic dynamical systems capable of generating turbulent and noise-like excitations. The scheme employs correlation fractal dimension and Takens embedding theorem to measure fractal dimension from time series observation of the dynamical systems. It uses linear predictive coding (LPC) excitation of twenty-two consonant waveforms as the time series. Furthermore, the correlation fractal dimension is calculated using a fast Grassberger algorithm [Grassberger and Procaccia(1983)].

30

The criterion in the False Nearest Neighbor approach for determining optimal embedding dimension is subjective in some sense that, different values of parameters may lead to different results [Cao.L(1997)]. He proposed in his work a practical method to determine the minimum embedding dimension from a scalar time series. It does not contain any subjective parameters except for the time delay for the embedding. It does not strongly depend on how many data points are available and it is computationally efficient. Several time series are tested to show the above advantages of the method. For real time series data, different optimal embedding dimensions are obtained for different values of the threshold value. Also with noisy data this method gives spurious results [Kantz and Schreiber.T(2003)].

Narayanan presented an algorithm for voiced/unvoiced speech signal classification using second order attractor dimension and second order kolmogorov entropy of the speech signals. The non-linear dynamics of the speech signal is experimentally analyzed using this approach. The proposed techniques were further used as a powerful tool for the classification of voiced/unvoiced speech signals in many applications [Narayanan(1999)].

In [N K Narayanan and Sasindran(2000)] Narayanan et al., investigated on the applications of phase space map and phase space point distribution parameter for the recognition of Malayalam vowel units. The presented features were extracted by utilizing the non-linear/chaotic signal processing techniques. Andrew et al., presented phase space feature for the classification of TIMIT corpus and demonstrated that the proposed technique outperform compared with frequency domain based MFCC feature parameters [Andre C Lingren and Povinelli(2003)]. 31

Petry et al.,

[A Petry and Barone.C(2002)] and Pitsikalis et al.,

[Pitsikalis.V and Maragos.P(2003)] have used Lyapunov exponents and Correlation dimension in unison with traditional features (cepstral coefficients) and have shown minor improvements over baseline speech recognition systems. Central to both sets of these papers is the importance of Lyapunov exponents and Correlation dimension, because they are invariant metrics that are the same regardless of initial conditions in both the original and reconstructed phase space. Despite their significance, there are several issues that exist in the measuring of these quantities on real experimental data. The most important issue is that these measurements are very sensitive to noise. Secondarily, the automatic computation of these quantities through a numerical algorithm is not well established and this can lead to drastically differing results. The overall performance of these quantities as salient features remains an open research question.

In [P Prajith and Narayanan(2004)], P Prajith et al., proposed a feature parameter by utilizing nonlinear or chaotic signal processing techniques to extract time domain based phase space features.Two sets of experiments are presented. In the first, exploiting the theoretical results derived in nonlinear dynamics, a processing space called phase space is generated and a recognition parameter called Phase Space Point Distribution (PSPD) is extracted. In the second experiment Phase Space Map at a phase angle p/2 is reconstructed and PSPD is calculated. The output of a neural network with error back propagation algorithm demonstrate that phase space features contain substantial discriminatory power.

32

Kevin M Lindrebo et al., introduced a method for calculating speech features from third-order statistics of sub band filtered speech signals which are used for robust speech recognition [Kevin M. Indrebo(2005)]. These features have the potential to capture nonlinear information not represented by cepstral coefficients. Also, because the features presented in this method are based on the third-order moments, they may be more immune to Gaussian noise than cepstrals, as Gaussian distributions have zero third-order moments.

Richard J Povinelli et al., introduced a novel approach to the analysis and classification of time series signals using statistical models of reconstructed phase spaces [Povinelli.R.J(2006)]. With sufficient dimension, such reconstructed phase spaces are, with probability one, guaranteed to be topologically equivalent to the state dynamics of the generating system, and, therefore, may contain information that is absent in analysis and classification methods rooted in linear assumptions. Parametric and non parametric distributions are introduced as statistical representations over the multidimensional reconstructed phase space, with classification accomplished through methods such as Bayes maximum likelihood and artificial neural networks (ANNs). The technique is demonstrated on heart arrhythmia classification and speech recognition. This new approach is shown to be a viable and effective alternative to traditional signal classification approaches, particularly for signals with strong nonlinear characteristics.

In [Prajith and Narayanan(2006)] P Prajith and N K Narayanan had introduced a flexible algorithm for pitch calculation by utilizing the methodologies developed for analyzing chaotic time series. The experimental result showed that the pitch 33

estimated using reconstructed phase space feature agrees with that obtained using conventional pitch detection algorithm.

Marcos Faundez-Zanuy compared the identification rates of a speaker recognition system using several parameterizations, with special emphasis on the residual signal obtained from linear and nonlinear predictive analysis [Zanuy(2007)]. It is found that the residual signal is still useful even when using a high dimensional linear predictive analysis. If instead of using the residual signal of a linear analysis a nonlinear analysis is used, both combined signals are more uncorrelated and although the discriminating power of the nonlinear residual signal is lower, the combined scheme outperforms the linear one for several analysis orders.

P Prajith introduced in his thesis the applications of nonlinear dynamical theory. As an alternate to traditional model of speech production a nonlinear system has been proposed. The problem of whether speech (especially vowel sounds) is chaotic has been examined through discussion of previous studies and experiments. Nonlinear invariant parameters for Malayalam vowels are calculated. The major invariant features include attractor dimensions and Kolmogorov entropy. The non-integer attractor dimension and non-zero value of Kolmogorov entropy confirm the contribution of deterministic chaos to the behavior of speech signal [Prajith(2008)].

In a recent study Kar proposed a novel criterion for the global asymptotic stability of fixed-point state-space digital filters under various combinations of quantization and overflow non linearities [Kar(2011)]. Yucel Ozbek et al., pro34

posed a systematic framework for accurate estimation of articulatory trajectories from acoustic data based on multiple-model dynamic systems via state-space representation [Yucel Ozbek and Demirekler(2012)]. The acoustic measurements and articulatory positions are considered as observable (measurement) and hidden (state) quantities of the system, respectively. To improve the performance of state space-based articulatory inversion they have used jump-Markov linear system (JMLS). Comparison of the performance of their method with the reported ones given in the literature shows that the proposed method improves the performance of the state-space approaches.

It is seen that the majority of literature that utilizes the nonlinear techniques for signal processing applications revolves around its use for control, prediction and noise reduction, reporting both positive and negative results. There is only scattered research using these methods for classification or recognition experiments. It is also important to notice that very less works are reported yet in nonlinear speech processing for Malayalam and no such work has been reported in other Indian languages. The succeeding session of this chapter is focused on the review of the applications of artificial neural network for speech recognition.

2.5 Review on Applications of ANN for Speech Recognition Artificial neural net (ANN) algorithms have been designed and implemented for speech pattern recognition by a number of researchers. ANNs are of interest because algorithms used in many speech recognizers can be implemented using highly parallel neural net architectures and also because new parallel algorithms

35

are being developed making use of the newly acquired knowledge of the working of biological nervous systems. Hutton.L.V compares neural network and statistical pattern comparison method for pattern recognition purpose [Hutton.L.V(1992)]. Neural network approaches to pattern classification problems complement and compete with statistical approaches. Each approach has unique strengths that can be exploited in the design and evaluation of classifier systems. Classical (statistical) techniques can be used to evaluate the performance of neural net classifiers, which often outperform them. Neural net classifiers may have advantages even when their ultimate performance on a training set can be shown to be no better than the classical. It is possible to be implemented in real time using special purpose hardware.

Personnaz L et al., presents an elementary introduction to networks of formal neurons [Personnaz.L and Dreyfus.G(1990)]. The state of the art regarding basic research and the applications are presented in this work. First, the most usual models of formal neurons are described, together with the most currently used network architectures: static (feedforward) nets and dynamic (feedback) nets. Secondly, the main potential applications of neural networks are reviewed: pattern recognition (vision, speech), signal processing and automatic control. Finally, the main achievements (simulation software, simulation machines, integrated circuits) are presented.

Willian Huang et al., presents some neural net approaches for the problem of static pattern classification and time alignment. For static pattern classification multi layer perceptron classifiers trained with back propagation can form arbitrary 36

decision regions, are robust, and are trained rapidly for convex decision regions. For time alignment, the Viterbi net is a neural net implementation of the Viterbi decoder used very effectively in recognition systems based on Hidden Markov Models (HMMs) [William Huang and Gold(1988)].

Waibel.A et al., [Weibel.A and Lang.K(1988)] proposed a time delay neural network (TDNN) approach to phoneme recognition, which is characterized by two important properties. Using a three level arrangement of simple computing units, it can represent arbitrary non-linear decision surface. The TDNN learns these decision surfaces automatically using error back propagation. The time delay arrangement enables the network to discover acoustic phonetic features and temporal relationships between them independent of position in time and hence not blurred by temporal shifts in the input. For comparison, several discrete Hidden Markov Models (HMM) were trained to perform the same task, i.e. the speaker dependent recognition of the phonemes "B", "D" and "G" extracted from varying phonetic contexts. The TDNN achieved a recognition rate of 98.5% correct compared to 93.7% for the best of HMMs. They showed that the TDNN has well known acoustic-phonetic features (e.g., F2-rise, F2-fall, vowel-onset) as useful abstractions. It also developed alternate internal representations to link different acoustic realizations to the same concept.

Yoshua Bengio and Renato De Mori used The Boltzmann machine algorithm and the error back propagation algorithm to learn to recognize the place of articulation of vowels(front, center or back), represented by a static description of spectral lines [Bengio and Mori(1988)]. The error rate is shown to depend on the 37

coding. Results are comparable or better than those obtained by them on the same data using hidden Markov Models. They also show a fault tolerant property of the neural nets, i.e. that the error on the test set increases slowly and gradually when an increasing number of nodes fail.

Mah. R.S.H and Chakravarthy.V examined the key features of simple networks and their application to pattern recognition [Mah.R.S.H and Chakravarthy.V(1992)]. Beginning with a three-layer back propagation network, the authors examine the mechanisms of pattern classification. They relate the number of input, output and hidden nodes to the problem features and parameters. In particular, each hidden neuron corresponds to a discriminant in the input space. They point out that the interactions between number of discriminants, the size and distribution of the training set, and numerical magnitudes make it very difficult to provide precise guidelines. They found that the shape of the threshold function plays a major role in both pattern recognition, and quantitative prediction and interpolation. Tuning the sharpness parameter could have a significant effect on neural network performance. This feature is currently under-utilized in many applications. For some applications linear discriminant is a poor choice.

Janssen et al., developed a phonetic front-end for speaker-independent recognition of continuous letter strings [Janssen.R.D.T and Cole.R.A(1991)]. A feedforward neutral network is trained to classify 3 msec speech frames as one of the 30 phonemes in the English alphabet. Phonetic context is used in two ways: first, by providing spectral and waveform information before and after the frame to be classified, and second, by a second-pass network that uses both acoustic features 38

and the phonetic outputs of the first-pass network. This use of context reduced the error rate by 50%. The effectiveness of the DFT and the more compact PLP (perceptual linear predictive) analysis is compared, and several other features, such as zerocrossing rate, are investigated. A frame-based phonetic classification performance of 75.7% was achieved.

Ki-Seok-Kim and Hee-Yeung-Hwang present the result of the study on the speech recognition of Korean phonemes using recurrent neural network models conducted by them [Ki-Seok-Kim and Hee-Yeung-Hwang(1991)]. The results of applying the recurrent multi layer perceptron model for learning temporal characteristics of speech phoneme recognition is presented. The test data consist of 144 vowel+consonant+vowel (V+CV) speech chains made up of 4 Korean monothongs and 9 Korean plosive consonants. The input parameters of the artificial neural network model used are the FFT coefficients, residual error and zero crossing rates. The baseline model showed a recognition rate of 91% for vowels and 71% for plosive consonants of one male speaker. The authors obtained better recognition rates from various other experiments compared to the existing multilayer perceptron model, thus showing the recurrent model to be better suited to speech recognition. The possibility of using the recurrent models for speech recognition was experimented upon by changing the configuration of this baseline model.

Ahn.R and Holmes.W.H propose a voiced / unvoiced / silence classification algorithm of speech using 2-stage neural networks with delayed decision input [Ahn.R and Holmes.W.H(1996)]. This feed forward neural network classifier is 39

capable of determining voiced, unvoiced and silence in the first stage and refining unvoiced and silence decisions in the second stage. Delayed decision from the previous frame’s classification along with preliminary decision by the first stage network, zero crossing rate and energy ratio enable the second stage to correct the mistakes made by the first stage in classifying unvoiced and silence frames. Comparisons with a single stage classifier demonstrate the necessity of two-stage classification techniques. It also shows that the proposed classifier performs excellently.

Sunilkumar and Narayanan investigated the potential use of zerocrossing based information of the signal for Malayalam vowel recognition. A vowel recognition system using artificial neural network is developed. The highest recognition accuracy obtained for normal speech is 90.62% [Sunilkumar(2002)], [R K Sunilkumar and Narayanan(2004)].

Dhananjaya et al., proposed a method for detecting speaker changes in a multi speaker speech signal [Dhananjaya.N and Yagnanarayana.B(2004)]. The statistical approach to a point phenomenon (speaker change) fails when the given conversation involves short speaker turns (< 5 sec duration). They used auto associative neural network (AANN) models to capture the characteristics of the excitation source that present in the linear prediction (LP) residue of speech signal. The AANN models are then used to detect the speaker changes.

In [P Prajith and Narayanan(2004)] P Prajith et al., presented the implementation of a neural network with error back propagation algorithm for the speech 40

recognition application with Phase Space Point Distribution as the input parameter. A method is suggested for speech recognition by utilizing nonlinear or chaotic signal processing techniques to extract time domain based phase space features. Two sets of experiments are presented in this paper. In the first, exploiting the theoretical results derived in nonlinear dynamics, a processing space called phase space is generated and a recognition parameter called Phase Space Point Distribution (PSPD) is extracted. In the second experiment Phase Space Map at a phase angle p/2 is reconstructed and PSPD is calculated. The output of a neural network with error back propagation algorithm demonstrate that phase space features contain substantial discriminatory power.

In [R K Sunilkumar and Narayanan(2004)] the speech signal is modeled using the zerocrossing base features of the signal. This feature is used for recognizing the Malayalam vowels and Consonant Vowel Units using Kolmogorov- Smirnov statistical test and multilayer feed forward artificial neural network. The average vowel recognition accuracy for single speaker using ANN base method is 92.62 % and for three female speaker is 91.48%. The average Consonant vowel recognition accuracy for single speaker is 73.8%. The advantage of this method is that the network shows better performance than the other conventional techniques and it takes less computation than the other conventional techniques of parameterization of speech signal like FFT, and Cepstral methods.

Xavier Domont et al., proposed a feed forward neural network for syllable recognition [Xavier Domont and Goerick(2007)]. The core of the recognition system is based on a hierarchical architecture initially developed for visual object 41

recognition. In this work, they showed that, given the similarities between the primary auditory and visual cortexes, such a system can successfully be used for speech recognition. Syllables are used as basic units for the recognition. Their spectrograms, computed using a Gammatone filter bank, are interpreted as images and subsequently feed into the neural network after a preprocessing step that enhances the formant frequencies and normalizes the length of the syllables.

P Prajith investigated in his work the application of Multi Layer Feed Forward Neural Network (MLFFNN) with error back propagation algorithm for the classification of Malayalam vowel units. To evaluate the credibility of the classifier he used reconstructed phase approach in combination with Mel Frequency Cepstral Coefficient(MFCC). An overall recognition accuracy of 96.24% is obtained from the simulation experiment and reported that a significant boost in recognition accuracy is obtained using ANN with the hybrid features [Prajith(2008)].

Anupkumar et al., [Anupkumar Paul and Kamal(2009)] studied Linear Predictive Coding Coefficients (LPCC) and Artificial Neural Network (ANN) for the recognition of Bangala speech. They presented the different neural network architecture design for the pattern at hand. It is concluded that neural networks having more hidden layers are able to solve the problems very easily. By comparing error curves and recognition accuracy of digits it is concluded that Multi Layer Perceptron with 5 layers is a more generic approach rather than Multi Layer Perceptron with 3 hidden layers.

42

Hanchate et al., investigated the application of Neural Networks with one hidden layer with sigmoid functions and the output layer with linear functions. There are 10 output neurons for all the networks while the numbers of hidden neurons vary from 10 to 70. The inputs of the network are the features of 4 selected frames with 256 samples per frame. Each frame is represented by 12 MFCC coefficients of the signal in the frame. A comparatively good recognition accuracy is obtained using this approach [D B Hanchate and Mourya(2010)].

In [Yong and Ting(2011)], Yong and Ting investigated the speaker independent vowel recognition for Malay children using the Time Delay Neural Network (TDNN). Due to less research done on the children speech recognition, the temporal structure of the children speech was not fully understood. Two hidden layers TDNN was proposed to discriminate 6 Malay vowels: /a/, /e/, /i/, /o/ and /u/. The speech database consisted of vowel sounds from 360 children speakers. Cepstral coefficient was normalized for the input of TDNN. The frame rate of the TDNN was tested with 10ms, 20ms, and 30ms. It was found out that the 30ms frame rate produced the highest vowel recognition accuracy with 81.92%. The TDNN also showed higher speech recognition rate compared to the previous studies that used Multilayer Perceptron.

The zerocrossing interval distribution of vowel speech signal is studied by Sunilkumar and Lajish [Sunilkumar and Lajish(2012)] using 5 Malayalam short vowel units. The classification of these sounds are carried out using multilayer feed forward artificial neural network. After analyzing the distribution patterns and the vowel recognition results, they reported that the zerocrossing interval dis43

tribution parameters can be effectively used for the speech phone classification and recognition. The noise adaptness of this parameter is also studied by adding additive white Gaussian noise at different signal to noise ratio. The computational complexity of the proposed technique is also less compared to the conventional spectral techniques which includes FFT and Cepstral methods, used in the parameterization of speech signal.

In [Battacharjee(2012)] Bhattacharjee discussed a novel technique for the recognition of Assamese phonemes using Recurrent Neural Network (RNN) based phoneme recognizer. A Multi-Layer Perceptron (MLP) has been used as phoneme segmenter for the segmentation of phonemes from isolated Assamese words. Two different RNN based approaches have been considered for recognition of the phonemes and their performances have been evaluated. MFCC has been used as the feature vector for both segmentation and recognition. With RNN based phoneme recognizer, a recognition accuracy of 91% has been achieved. The RNN based phoneme recognizer has been tested for speaker mismatch and channel mismatch conditions. It has been observed that the recognizer is robust to any unseen speaker. However, its performance degrades in channel mismatch condition. Cepstral Mean Normalization (CMN) has been used to overcome the problem of performance degradation effectively.

In this thesis the application of linear and non-linear dynamical system models and multi resolution analysis using wavelet transform features of V/CV speech units for the recognition using brain like computing algorithm namely Artificial Neural Networks is explored in detail. 44

2.6 Review on Statistical Learning Algorithms for Speech Recognition Support Vector Machines (SVM) are learning techniques that is considered as an effective method for general purpose pattern recognition because of its high generalization performance without the need of domain specific knowledge [Vapnik(1995)]. Intuitively, given a set of points belonging to two classes, a SVM finds a hyperplane that separates the largest possible fraction of points of the same class on the same side, while maximizing the distance from either class to the hyperplane. This is the optimal separating hyperplane which minimizes the risk of misclassifying not only the examples in the training set, but also the unseen example of the test set.

The main characteristics of SVM are that they minimize a formally proven upper bound on the generalization error. They work on high dimensional feature space by means of a dual formulation in terms of kernels. The prediction is based on hyperplanes in these feature spaces, which may correspond to quite involved classification criteria on the input data. The layer in the training data set can be handled by means of soft margins.

In a work done [Clarkson and Moreno(1997)] by Clarkson and Moreno, authors explores the issues involved in applying SVMs to phonetic classification as a first step to speech recognition. They presented results on several standard vowel and phonetic classification tasks and show better performance than Gaussian mixture classifiers. They also presented an analysis of the difficulties they

45

foresee in applying SVMs to continuous speech recognition problems. This paper represents a preliminary step in understanding the problems of applying SVMs to speech recognition.

As a preliminary analysis on speech signal analysis Anil K Jain et al., [Anil k Jain and Mao(2000)] presented a robust speech recognizer based on features obtained from the speech signal. The authors explored the issues involved in applying SVMs to phonetic classification. They presented results on several standard vowel and phonetic classification tasks and showed better performance than Gaussian mixture classifiers. They also present an analysis of the difficulties in applying SVMs to continuous speech recognition problems.

In a work presented by Aravindh Ganapathiraju et al., they addressed the use of a support vector machine as a classifier in a continuous speech recognition system. The technology has been successfully applied to two speech recognition tasks. A hybrid SVM/HMM system has been developed that uses SVMs to postprocess data generated by a conventional HMM system. The results obtained in the experiments clearly indicate the classification power of SVMs and affirm the use of SVMs for acoustic modeling. The oracle experiments reported in their work clearly show the potential of this hybrid system while highlighting the need for further research into the segmentation issue [Aravind Ganapathiraju and Picone(2004)].

In a work done by Tsang-Long et al., [Tsang-Long Pao and Li(2006)] SVM & NN classifiers and feature selection algorithm were used to classify five emotions from Mandarin emotional speech and compared their experimental results. The 46

overall experimental results reveal that the SVM classifier (84.2%) outperforms the NN classifier (80.8%) and detects anger perfectly, but confuses happiness with sadness, boredom and neutral. The NN classifier achieves better performance in recognizing sadness and neutral and differentiates happiness and boredom perfectly.

In [Jing Bai(2006)], to improve the learning and generalization ability of the machine-learning model, a new compound kernel that may pay attention to the similar degree between sample space and feature space is proposed. The author used the new compound kernel support vector machine to a speech recognition system for Chinese isolated words, non-specific person and middle glossary quantity, and compared the speech recognition results with the SVM using traditional kernels and RBF network. Experiments showed that the SVM performance with the new compound kernel is much better than traditional kernels and has higher recognition rates than ones of using RBF network in different SNRs, and is of shorter training time.

Sandhya Arora et al., [Sandhya Arora and Basu(2010)] discussed the characteristics of some classification methods that have been successfully applied to handwritten Devnagari character recognition and results of SVM and ANNs classification method, applied on Handwritten Devnagari characters. After preprocessing the character image, they extracted shadow features, chain code histogram features, view based features and longest run features. These features are then fed to Neural classifier and in Support Vector Machine for classification. In neural classifier, they explored three ways of combining decisions of four MLP’s, de47

signed for four different features.

In a work done by Zhuo-ming Chen et al., [Zhuo-ming Chen and tao Yao(2011)] they extracted a new feature(DWTMFC-CT) of the consonants by employing wavelet transformation, and explains that the difference of similar consonants can be described more accurately by this feature. The algorithm used for classification was multi-class fuzzy support vector machine(FSVM). In order to reduce the computation complexity caused by using the standard fuzzy support vector machine for multi-class classification, this paper propose an algorithm based on two stages. Experimental results shows that the proposed algorithm could get better classification results while reducing the training time greatly.

In [Ruben Solera-urena and de MariA(2012)], authors suggest the use of a weighted least squares (WLS) training procedure that facilitates the possibility of imposing a compact semiparametric model on the SVM, which results in a dramatic complexity reduction. Such a complexity reduction with respect to conventional SVM, which is between two and three orders of magnitude, allows the hybrid WLS-SVC/HMM system to perform real-time speech decoding on a connected-digit recognition task (Spanish database namely SpeechDat). The experimental evaluation of the proposed system shows encouraging performance levels in clean and noisy conditions.

In short, SVM have been widely used for speech recognition application for the last few years. It is because of its high generalization performance even without the domain specific knowledge. In this thesis application of non-linear and 48

wavelet based feature of V/CV speech units for recognition using SVM is studied in detail

2.7 Conclusion In summary, the current stage in the evalution of speech recognition research result from a combination of several elements, such as versatility of the database used, credibility of the different strategies of the feature selection, environmental conditions, the performance of different classifiers and their combinations etc. It is clear that no much well known attempts are reported towards in the recognition of Malayalam V/CV speech units and hence more research is needed to improve the recognition rates of V/CV units in both clean and noisy conditions. The studies performed in the following chapter of this thesis reveal that, the multi resolution analysis and non-linear dynamical system approach have a very good role in providing flexible information processing capability by devising methodologies and algorithms on a massively parallel system capable of handling infinite intra-class variations for representation and recognition of V/CV speech units.

49

❘❡❢❡r❡♥❝❡s ❬❆✳ ❉❡✈ ❛♥❞ ❈❤♦✉❞❤✉r②✭✷✵✵✸✮❪ ❙✳❙✳ ❆❣r❛✇❛❧ ❆✳ ❉❡✈ ❛♥❞ ❉✳❘✳ ❈❤♦✉❞❤✉r②✳ ❈❛t❡❣♦r✐③❛t✐♦♥ ♦❢ ❤✐♥❞✐ ♣❤♦♥❡♠❡s ❜② ♥❡✉r❛❧ ♥❡t✇♦r❦s✳ ❆■ ❙♦❝✐❡t②✱ ✶✼ ✭✹✮✿✸✼✺✕✸✽✷✱ ✷✵✵✸✳ ❬❆ ●r♦ss♠❛♥ ❛♥❞ ●❛♦✉♣✐❧❧❛✉❞✭✶✾✽✹✮❪ ❏ ▼♦r❧❡t ❆ ●r♦ss♠❛♥ ❛♥❞ P ●❛♦✉♣✐❧✲ ❧❛✉❞✳ ❈②❝❧❡ ♦❝t❛✈❡ ❛♥❞ r❡❧❛t❡❞ tr❛♥s❢♦r♠s ✐♥ s❡✐s♠✐❝ s✐❣♥❛❧ ❛♥❛❧②s✐s✳ ●❡♦❡①♣❧❛r❛t✐♦♥✱ ✷✸✿✽✺✕✶✵✷✱ ✶✾✽✹✳ ❬❆✳ ▼✳ ❊❧✇❛❦❞② ❛♥❞ ❊❧❤❡♥♥❛✇②✭✷✵✵✽✮❪ ❈✳▼✳ ❊❧t♦❦❤② ❆✳ ▼✳ ❊❧✇❛❦❞②✱ ❇✳❊✳❊❧s❡❡❤❡❧② ❛♥❞ ❉✳❆✳ ❊❧❤❡♥♥❛✇②✳ ❙♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ❛ ✇❛✈❡❧❡t tr❛♥s❢♦r♠ t♦ ❡st❛❜❧✐s❤ ❢✉③③② ✐♥❢❡r❡♥❝❡ s②st❡♠ t❤r♦✉❣❤ s✉❜tr❛❝t✐✈❡ ❝❧✉s✲ t❡r✐♥❣ ❛♥❞ ♥❡✉r❛❧ ♥❡t✇♦r❦ ✭❛♥✜s✮✳ ■♥t❡r♥❛t✐♦♥❛❧ ❏♦✉r♥❛❧ ♦❢ ❈✐❡❝✉✐ts ❙②st❡♠s ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✹✭✷✮✿✷✻✹✕✷✼✸✱ ✷✵✵✽✳ ❬❆ P❡tr② ❛♥❞ ❇❛r♦♥❡✳❈✭✷✵✵✷✮❪ ❆✉❣♦st♦ ❉ ❆ P❡tr② ❛♥❞ ❇❛r♦♥❡✳❈✳ ❙♣❡❛❦❡r ✐❞❡♥t✐✜❝❛t✐♦♥ ✉s✐♥❣ ♥♦♥❧✐♥❡❛r ❞②♥❛♠✐❝❛❧ ❢❡❛t✉r❡s✳ ❈❤❛♦s✱❙♦❧✉t✐♦♥s ❛♥❞ ❋r❛❝t❛❧s✱ ✶✸✿✷✷✶✕✷✸✶✱ ✷✵✵✷✳ ❬❆❜❛r❜❛♥❡❧✳❍✳❉✳■ ❛♥❞ ❚s✐♠r✐♥❣✳▲✳❙✭✶✾✾✸✮❪ ❙✐❞r♦r♦✇✐❝❤✳❏✳❏ ❆❜❛r✲ ❜❛♥❡❧✳❍✳❉✳■✱ ❇r♦✇♥✳❘ ❛♥❞ ❚s✐♠r✐♥❣✳▲✳❙✳ ❚❤❡ ❛♥❛❧②s✐s ♦❢ ♦❜s❡r✈❡❞ ❝❤❛♦t✐❝ ❞❛t❛ ✐♥ ♣❤②s✐❝❛❧ s②st❡♠s✳ ❚❡❝✳ ▼♦❞✳ P❤②s✳✱✱ ✻✺✱ ✶✾✾✸✳ ❬❆❤♥✳❘ ❛♥❞ ❍♦❧♠❡s✳❲✳❍✭✶✾✾✻✮❪ ❆❤♥✳❘ ❛♥❞ ❍♦❧♠❡s✳❲✳❍✳ ❱♦✐❝❡❞✴✉♥✈♦✐❝❡❞✴s✐❧❡♥❝❡ ❝❧❛ss✐✜❝❛t✐♦♥ ♦❢ s♣❡❡❝❤ ✉s✐♥❣ ✷✲st❛❣❡ ♥❡✉r❛❧ ♥❡t✇♦r❦s ✇✐t❤ ❞❡❧❛②❡❞ ❞❡❝✐s✐♦♥ ✐♥♣✉t✳ Pr♦❝✳ ❋♦✉rt❤ ■♥t❡r♥❛t✐♦♥❛❧ ❙②♠✲ ♣♦s✐✉♠ ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣ ❛♥❞ ✐ts ❆♣♣❧✐❝❛t✐♦♥s✳ ■❙❙P❆ ✾✻✱❆✉str❛❧✐❛✱ ✶✾✾✻✳ ❬❆✳❑✳●❤♦s❤ ❛♥❞ ❈✳❆✳▼✉rt❤②✭✷✵✵✺✮❪ P✳❈❤❛✉❞❤✉r ❆✳❑✳●❤♦s❤ ❛♥❞ ❈✳❆✳▼✉rt❤②✳ ❖♥ ✈✐s✉❛❧✐③❛t✐♦♥ ❛♥❞ ❛❣❣r❡❣❛t✐♦♥ ♦❢ ♥❡❛r❡st ♥❡✐❣❤✲ ❜♦r ❝❧❛ss✐✜❡rs✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ P❛tt❡r♥ ❆♥❛❧②s✐s ❛♥❞ ▼❛❝❤✐♥❡ ■♥t❡❧❧✐❣❡♥❝❡✱ ✷✼✿✶✺✾✷✕✶✻✵✷✱ ✷✵✵✺✳ ❬❆❧❧✐❣♦♦❞✳❑ ❛♥❞ ❨♦r❦❡✳❏✭✶✾✾✼✮❪ ❙❛✉❡r✳❚ ❆❧❧✐❣♦♦❞✳❑ ❛♥❞ ❨♦r❦❡✳❏✳ ❈❤❛♦s✿ ❆♥ ■♥tr♦❞✉❝t✐♦♥ t♦ ❉②♥❛♠✐❝❛❧ ❙②st❡♠s✳ ❙♣✐♥❣❡râ⑨➇❱❡r❧❛❣✱ ◆❡✇ ❨♦r❦✱ ✶✾✾✼✳ ❬❆♥❞r❡ ❈ ▲✐♥❣r❡♥ ❛♥❞ P♦✈✐♥❡❧❧✐✭✷✵✵✸✮❪ ▼✐❝❤❛❡❧ ❚ ❏❤♦♥s♦♥ ❆♥❞r❡ ❈ ▲✐♥❣r❡♥ ❛♥❞ ❘✐❝❤❛r❞ ❏ P♦✈✐♥❡❧❧✐✳ ❙♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ r❡❝♦♥str✉❝t❡❞ ♣❤❛s❡ s♣❛❝❡ ❢❡❛t✉r❡s✳ ■♥ ■♥ Pr♦❝✳ ■❊❊❊ ✐♥t✳ ❈♦♥❢✳ ♦♥ ❆❝♦✉st❝st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✷✵✵✸✱ ✷✵✵✸✳ ✷✺✺

❬❆♥❞r❡♦♣♦✉❧♦s ❛♥❞ ✈❛♥ ❞❡r ❙❝❤❛❛r✭✷✵✵✼✮❪ ❨✐❛♥♥✐s ❆♥❞r❡♦♣♦✉❧♦s ❛♥❞ ▼✐✲ ❤❛❡❧❛ ✈❛♥ ❞❡r ❙❝❤❛❛r✳ ●❡♥❡r❛❧✐③❡❞ ♣❤❛s❡ s❤✐❢t✐♥❣ ❢♦r ♠✲❜❛♥❞ ❞✐s❝r❡t❡ ✇❛✈❡❧❡t ♣❛❝❦❡t tr❛♥s❢♦r♠s✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✺✺✭✷✮✿✼✹✷✕✼✹✼✱ ✷✵✵✼✳ ❬❆♥✐❧ ❦ ❏❛✐♥ ❛♥❞ ▼❛♦✭✷✵✵✵✮❪ ❘♦❜❡rt P ❲ ❉✉✐♥ ❆♥✐❧ ❦ ❏❛✐♥ ❛♥❞ ❏✐❛♥❝❤❛♥❣ ▼❛♦✳ ❙t❛✐st✐❝❛❧ ♣❛tt❡r♥ r❡❝♦❣♥✐t✐♦♥✿ ❆ r❡✈✐❡✇✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ P❛tt❡r♥ ❆♥❛❧②s✐s ❛♥❞ ▼❛❝❤✐♥❡ ■♥t❡❧❧✐❣❡♥❝❡✱ ✷✷✭✶✮✿✹✕✸✼✱ ✷✵✵✵✳ ❬❆♥✉♣❦✉♠❛r P❛✉❧ ❛♥❞ ❑❛♠❛❧✭✷✵✵✾✮❪ ❉✐♣❛♥❦❛r ❉❛s ❆♥✉♣❦✉♠❛r P❛✉❧ ❛♥❞ ▼❞✳ ▼✉st❛❢❛ ❑❛♠❛❧✳ ❇❛♥❣❧❛ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ s②st❡♠ ✉s✐♥❣ ❧♣❝ ❛♥❞ ❛♥♥✳ ■♥ ■♥ Pr♦❝✳ ■❊❊❊ ✐♥t✳ ❈♦♥❢✳ ♦♥ ❆❞✈❛♥❝❡s ✐♥ P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥✱ ♣❛❣❡s ✶✼✶✕✶✼✹✱ ✷✵✵✾✳ ❬❆r❛✈✐♥❞ ●❛♥❛♣❛t❤✐r❛❥✉ ❛♥❞ P✐❝♦♥❡✭✷✵✵✹✮❪ ❏♦♥❛t❤❛♥ ❊✳ ❍❛♠❛❦❡r ❆r✲ ❛✈✐♥❞ ●❛♥❛♣❛t❤✐r❛❥✉ ❛♥❞ ❏♦s❡♣❤ P✐❝♦♥❡✳ ❆♣♣❧✐❝❛t✐♦♥s ♦❢ s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡s t♦ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■❊❊❊✳ ❚r❛♥s✳ ♦♥ ❙✐❣♥❛❧ ♣r♦❝❡s✐♥❣✱ ✺✷✭✽✮✿✷✸✹✽✕✷✸✺✺✱ ✷✵✵✹✳ ❬❆t❛❧ ❛♥❞ ❍❛♥❛✉❡r✭✶✾✼✶✮❪ ❇✳ ❙✳ ❆t❛❧ ❛♥❞ ❙✳ ▲✳ ❍❛♥❛✉❡r✳ ❙♣❡❡❝❤ ❛♥❛❧②s✐s ❛♥❞ s②♥t❤❡s✐s ❜② ❧✐♥❡❛r ♣r❡❞✐❝t✐♦♥ ♦❢ t❤❡ s♣❡❡❝❤ ✇❛✈❡✳ ❏✳ ❆❝♦✉st✳ ❙♦❝✳ ❆♠❡r✐❝❛✱ ✺✵✭✷✮✿✻✸✼✕✻✺✺✱ ✶✾✼✶✳ ❬❆t❛❧ ❛♥❞ ❙❝❤r♦❡❞❡r✭✶✾✻✼✮❪ ❇✳ ❙✳ ❆t❛❧ ❛♥❞ ▼✳ ❘✳ ❙❝❤r♦❡❞❡r✳ Pr❡❞✐❝t✐✈❡ ❝♦❞✲ ✐♥❣ ♦❢ s♣❡❡❝❤ s✐❣♥❛❧s✳ ■♥ ✐♥ Pr♦❝✳ ■✾✻✼ ❈♦♥ ❛♥❞ ❈♦♠♠✉♥✳ ❛♥❞ Pr♦❝❡ss✲ ✐♥❣✱ ✶✾✻✼✳ ❬❆t❛❧ ❛♥❞ ❍❛♥s✉❡r✭✶✾✼✶✮❪ ❇✳❙✳ ❆t❛❧ ❛♥❞ ❙✳ ▲✳ ❍❛♥s✉❡r✳ ❙♣❡❡❝❤ ❛♥❛❧②s✐s ❛♥❞ s②♥t❤❡s✐s ❜② ❧✐♥❡❛r ♣r❡❞✐❝t✐♦♥ ♦❢ t❤❡ s♣❡❡❝❤ ✇❛✈❡✳ ❏♦✉r♥❛❧ ♦❢ ❆❝♦✉st✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✐❝❛✱ ✺✵✿✻✸✼✕✻✺✺✱ ✶✾✼✶✳ ❬❇ ❛♥❞ ◆✭✷✵✵✹✮❪ ❩❤❛♥❣✳ ❇ ❛♥❞ ❙r✐❤❛r✐ ❙ ◆✳ ❋❛st ❦ â⑨➇ ♥❡❛r❡st ♥❡✐❣❤❜♦r ✉s✐♥❣ ❝❧✉st❡r ❜❛s❡❞ tr❡❡s✳ ■❊❊❊ tr❛♥s✳ ♦♥ P❛tt❡r♥ ❆♥❛❧②s✐s ❛♥❞ ▼❛❝❤✐♥❡ ■♥t❡❧❧✐❣❡♥❝❡✱ ✷✻✭✹✮✿✺✷✺✕✺✷✽✱ ✷✵✵✹✳ ❬❇✳ ❍✳ ❏✉❛♥❣ ❛♥❞ ❙♦♥❞❤✐✭✶✾✽✻✮❪ ❙✳ ❊✳ ▲❡✈✐♥s♦♥ ❇✳ ❍✳ ❏✉❛♥❣ ❛♥❞ ▼✳ ▼✳ ❙♦♥❞❤✐✳ ▼❛①✐♠✉♠ ❧✐❦❡❧✐❤♦♦❞ ❡st✐♠❛t✐♦♥ ❢♦r ♠✉❧t✐✈❛r✐❛t❡ ♠✐①t✉r❡ ♦❜✲ s❡r✈❛t✐♦♥s ♦❢ ♠❛r❦♦✈ ❝❤❛✐♥s✳ ■❊❊❊ ❚r❛♥s✳ ■♥❢♦r♠❛t✐♦♥ ❚❤❡♦r②✱ ■❚✲✸✷ ✭✷✮✿✸✵✼✕✸✵✾✱ ✶✾✽✻✳ ❬❇✳ ❲✐tt❡♥♠❛r❦ ❛♥❞ ❆r③❡♥✭✷✵✵✷✮❪ ❑✳❏✳❆str♦♠ ❇✳ ❲✐tt❡♥♠❛r❦ ❛♥❞ ❑ ❊ ❆r③❡♥✳ ❈♦♠♣✉t❡r ❈♦♥tr♦❧✿ ❆♥ ❖✈❡r✈✐❡✇✳ ■❋❆❈ Pr♦❢❡ss✐♦♥❛❧ ❇r✐❡❢✱ ✷✵✵✷✳ ✷✺✻

❬❇❛❣❣❡♥st♦ss✭✷✵✶✵✮❪ P❛✉❧ ▼✳ ❇❛❣❣❡♥st♦ss✳ ❆ ♠✉❧t✐✲r❡s♦❧✉t✐♦♥ ❤✐❞❞❡♥ ♠❛r❦♦✈ ♠♦❞❡❧ ✉s✐♥❣ ❝❧❛ss✲s♣❡❝✐✜❝ ❢❡❛t✉r❡s✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✺✽✭✶✵✮✿✺✶✻✺✕✺✶✼✽✱ ✷✵✶✵✳ ❬❇❛❦❡r ❛♥❞ ●♦❧❧✉❜✭✶✾✾✻✮❪ ●✳ ▲✳ ❇❛❦❡r ❛♥❞ ❏ ●♦❧❧✉❜✳ ❈❤❛♦t✐❝ ❆♥ ■♥tr♦❞✉❝t✐♦♥✳ ❈❛♠❜r✐❞❣❡ ❯♥✐✈❡rs✐t② Pr❡ss✱ ✶✾✾✻✳

❉②♥❛♠✐❝s ✿

❬❇❛♥❜r♦♦❦ ❛♥❞ ▼❝▲❛✉❣❤❧✐♥✭✶✾✾✹✮❪ ▼ ❇❛♥❜r♦♦❦ ❛♥❞ ❙ ▼❝▲❛✉❣❤❧✐♥✳ ■s s♣❡❡❝❤ ❝❤❛♦t✐❝❄ ■♥ ✐♥ Pr♦❝❡❡❞✐♥❣s✳ ■❊❊ ❈♦❧❧♦q✳ ❊①♣❧♦✐t✐♥❣ ❈❤❛♦s ✐♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✶✾✾✹✳ ❬❇❛ss❛♠ ❆✳ ◗✳ ❆❧✲◗❛t❛❜✭✷✵✶✵✮❪ ❘❛❥❛ ◆✳ ❆✐♥♦♥ ❇❛ss❛♠ ❆✳ ◗✳ ❆❧✲◗❛t❛❜✳ ❆r❛❜✐❝ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ❤✐❞❞❡♥ ♠❛r❦♦✈ ♠♦❞❡❧ t♦♦❧❦✐t✭❤t❦✮✳ ■♥ ■♥ ♣r♦❝✳ ■♥t❡r♥❛t✐♦♥❛❧ ❙②♠♣♦s✐✉♠ ✐♥ ■♥❢♦r♠❛t✐♦♥ ❚❡❝❤♥♦❧♦❣②✱ ♣❛❣❡s ✺✺✼✕✺✻✷✱ ✷✵✶✵✳ ❬❇❛tt❛❝❤❛r❥❡❡✭✷✵✶✷✮❪ ❯t♣❛❧ ❇❛tt❛❝❤❛r❥❡❡✳ ❘❡❝♦❣♥✐t✐♦♥ ♦❢ ❛ss❛♠❡s❡ ♣❤♦♥❡♠❡s ✉s✐♥❣ r♥♥ ❜❛s❡❞ r❡❝♦❣♥✐③❡r✳ ❙♣❡❡❝❤✱ ❙♦✉♥❞ ❛♥❞ ▼✉s✐❝ Pr♦❝❡ss✐♥❣✿ ❊♠✲ ❜r❛❝✐♥❣ ❘❡s❡❛r❝❤ ■♥ ■◆❉■❆✱ ✼✶✷✿✶✽✼✕✶✾✻✱ ✷✵✶✷✳ ❬❇❛②r❛♠ ❛♥❞ ❙❡❧❡s♥✐❝❦✭✷✵✵✾✮❪ ■❧❦❡r ❇❛②r❛♠ ❛♥❞ ❊✈❛♥✳ ❲✳ ❙❡❧❡s♥✐❝❦✳ ❋r❡q✉❡♥❝②✲❞♦♠❛✐♥ ❞❡s✐❣♥ ♦❢ ♦✈❡r❝♦♠♣❧❡t❡ r❛t✐♦♥❛❧✲❞✐❧❛t✐♦♥ ✇❛✈❡❧❡t tr❛♥s❢♦r♠s✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✺✼✭✽✮✿✷✾✺✼✕✷✾✼✷✱ ✷✵✵✾✳ ❬❇❡♥❣✐♦ ❛♥❞ ▼♦r✐✭✶✾✽✽✮❪ ❨♦s❤✉❛ ❇❡♥❣✐♦ ❛♥❞ ❘❡♥❛t♦ ❉❡ ▼♦r✐✳ ❯s❡ ♦❢ ♥❡✉r❛❧ ♥❡t✇♦r❦ ❢♦r t❤❡ r❡❝♦❣♥✐t✐♦♥ ♦❢ ♣❧❛❝❡ ♦❢ ❛rt✐❝✉❧❛t✐♦♥✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ◆❡✉r❛❧ ◆❡t✇♦r❦s✱ ✽✭✷✮✱ ✶✾✽✽✳ ❬❇❡rts❡❦❛s✭✶✾✾✺✮❪ ❉ P ❇❡rts❡❦❛s✳ ❉②♥❛♠✐❝❛❧ Pr♦❣r❛♠♠✐♥❣ ❛♥❞ ❖♣t✐♠❛❧ ❈♦♥✲ tr♦❧✱ ❱♦❧✉♠❡ ■ ❛♥❞ ■■✳ ▼❆✿❆t❤❡♥❛s ❙❝✐❡♥t✐✜❝✱ ✶✾✾✺✳ ❬❇✳❍✳ ❏✉❛♥❣ ❛♥❞ ❈❤♦✉✭✶✾✾✼✮❪ ❈✳❍✳ ▲❡❡ ❇✳❍✳ ❏✉❛♥❣ ❛♥❞ ❲✉ ❈❤♦✉✳ ▼✐♥✐♠✉♠ ❝❧❛ss✐✜❝❛t✐♦♥ ❡rr♦r r❛t❡ ♠❡t❤♦❞s ❢♦r s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■❊❊❊ ❚r❛♥s✳ ❙♣❡❡❝❤ ❛♥❞ ❆✉❞✐♦ Pr♦❝❡ss✐♥❣✱ ✺✭✸✮✿✷✺✼✕✷✻✺✱ ✶✾✾✼✳ ❬❇♦✉r❧❛r❞ ❛♥❞ ❉✉♣♦♥t✭✶✾✾✼✮❪ ❍ ❇♦✉r❧❛r❞ ❛♥❞ ❙ ❉✉♣♦♥t✳ ❙✉❜❜❛♥❞ ❜❛s❡❞ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■♥ Pr♦❝ ♦❢ t❤❡ ✶✶t❤ ■❊❊❊ ✐♥t❡r♥❛t✐♦♥❛❧ ❝♦♥❢❡r❡♥❝❡ ♦♥ ❆❝♦✉st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ♣❛❣❡s ✶✷✺✶✕✶✷✺✹✱ ✶✾✾✼✳ ❬❇r♦❛❞✭✶✾✼✷✮❪ ❉✳ ❏✳ ❇r♦❛❞✳ ❋♦r♠❛♥ts ✐♥ ❛✉t♦♠❛t✐❝ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ❏✳ ▼❛♥✲▼◗❝❤✐♥❡ ❙t✉❞✐❡s✱ ✹✿✹✶✶✕✹✷✵✱ ✶✾✼✷✳

✷✺✼

■♥t✳

❬❇r♦♦❦❡s✳❉✳▼ ❛♥❞ ◆❛②❧♦r✳P✳❆✭✶✾✽✽❛✮❪ ❇r♦♦❦❡s✳❉✳▼ ❛♥❞ ◆❛②❧♦r✳P✳❆✳ ❙♣❡❡❝❤ ♣r♦❞✉❝t✐♦♥ ♠♦❞❡❧❧✐♥❣ ✇✐t❤ ✈❛r✐❛❜❧❡ ❣❧♦tt❛❧ r❡✢❡❝t✐♦♥ ❝♦❡✣✲ ❝✐❡♥t✳ ■♥ Pr♦❝✳ ■♥t✳ ❈♦♥❢✳ ♦♥ ❆❝♦✉st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✶✾✽✽❛✳ ❬❇r♦♦❦❡s✳❉✳▼ ❛♥❞ ◆❛②❧♦r✳P✳❆✭✶✾✽✽❜✮❪ ❇r♦♦❦❡s✳❉✳▼ ❛♥❞ ◆❛②❧♦r✳P✳❆✳ ❙♣❡❡❝❤ ♣r♦❞✉❝t✐♦♥ ♠♦❞❡❧❧✐♥❣ ✇✐t❤ ✈❛r✐❛❜❧❡ ❣❧♦tt❛❧ r❡✢❡❝t✐♦♥ ❝♦❡✣✲ ❝✐❡♥t✳ ■♥ Pr♦❝✳ ❆❝♦✉st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ♣❛❣❡s ✻✼✶✕✻✼✹✱ ✶✾✽✽❜✳ ❬❇r♦♦♠❤❡❛❞ ❛♥❞ ❑✐♥❣✭✶✾✽✻✮❪ ❉✳ ❙✳ ❇r♦♦♠❤❡❛❞ ❛♥❞ ●✳ P✳ ❑✐♥❣✳ ❊①tr❛❝t✐♥❣ q✉❛❧✐t❛t✐✈❡ ❞②♥❛♠✐❝s ❢r♦♠ ❡①♣❡r✐♠❡♥t❛❧ ❞❛t❛✳ P❤②s✐❝❛ ❉✱ ♣❛❣❡s ✷✶✼✕ ✷✸✻✱ ✶✾✽✻✳ ❬❇✉rt ❛♥❞ ❊❞❡❧s♦♥✭✶✾✽✸✮❪ P✳ ❏✳ ❇✉rt ❛♥❞ ❊✳ ❍✳ ❊❞❡❧s♦♥✳ ❚❤❡ ❧❛♣❛❧❛❝✐❛♠ ♣②r❛♠✐❞ ❛s ❛ ❝♦♠♣❛❝t ✐♠❛❣❡ ❝♦❞❡✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❈♦♠♠✉♥✐❝❛t✐♦♥s✱ ❈❖▼✲✸✶✿✺✸✷✕✺✹✵✱ ✶✾✽✸✳ ❬❈❛♦✳▲✭✶✾✾✼✮❪ ❈❛♦✳▲✳ Pr❛❝t✐❝❛❧ ♠❡t❤♦❞ ❢♦r ❞❡t❡r♠✐♥✐♥❣ t❤❡ ♠✐♥✐♠✉♠ ❡♠✲ ❜❡❞❞✐♥❣ ❞✐♠❡♥s✐♦♥ ♦❢ ❛ s❝❛❧❛r t✐♠❡ s❡r✐❡s✳ P❤②s✐❝❛ ❉✱ ✶✶✵✿✹✸✶✕✺✵✱ ✶✾✾✼✳ ❬❈❛s❞❛❣❧✐✭✶✾✾✶✮❪ ▼ ❈❛s❞❛❣❧✐✳ ❈❤❛♦s ❛♥❞ ❞❡t❡r♠✐♥✐st✐❝ ✈❡rs✉s st♦❝❤❛st✐❝ ♥♦♥✲ ❧✐♥❡❛r ♠♦❞❡❧✐♥❣✳ ❏✳ ❘✳ ❙t❛t✐st✳ ❙♦❝✱ ✺✹✿✸✵✸✕✸✷✽✱ ✶✾✾✶✳ ❬❈❛s❞❛❣❧✐✳▼ ❛♥❞ ●✐❜s♦♥✳❏✭✶✾✾✶✮❪ ❋❛r♠❡r✳❏✳❉ ❈❛s❞❛❣❧✐✳▼✱ ❊✉❜❛♥❦✳❙ ❛♥❞ ●✐❜s♦♥✳❏✳ ❙t❛t❡ s♣❛❝❡ r❡❝♦♥str✉❝t✐♦♥ ✐♥ t❤❡ ♣r❡s❡♥❝❡ ♦❢ ♥♦✐s❡✳ P❤②s✐❝❛ ❉✱ ✺✶✱ ✶✾✾✶✳ ❬❈✳❈❤❛♥❞r❛ ❙❡❦❤❛r ❛♥❞ ❋✳■t❛❦✉r❛✭✷✵✵✶✮❪ ❑✳❚❛❦❡❞❛ ❈✳❈❤❛♥❞r❛ ❙❡❦❤❛r ❛♥❞ ❋✳■t❛❦✉r❛✳ ❘❡❝♦❣♥✐t✐♦♥ ♦❢ ❝♦♥s♦♥❛♥t✲✈♦✇❡❧ ✉tt❡r❛♥❝❡s ✉s✐♥❣ s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡s✳ ■♥ ■♥ ♣r♦❝✳ ❊✉r♦♣❡❛♥ ❙②♠♣♦s✐✉♠ ♦♥ ❆rt✐✜❝✐❛❧ ◆❡✉r❛❧ ◆❡t✇♦r❦s ❇r✉❣❡s ✭❇❡❧❣✐✉♠✮✱ ♣❛❣❡s ✼✕✶✷✱ ✷✵✵✶✳ ❬❈❤❛♥❞r❛s❤❡❦❤❛r ❛♥❞ ❨❡❣♥❛♥❛r❛②❛♥❛✭✷✵✵✷✮❪ ❈ ❈❤❛♥❞r❛s❤❡❦❤❛r ❛♥❞ ❇ ❨❡❣✲ ♥❛♥❛r❛②❛♥❛✳ ❆ ❝♦♥str❛✐♥t s❛t✐s❢❛❝t✐♦♥ ♠♦❞❡❧ ❢♦r r❡❝♦❣♥✐t✐♦♥ ♦❢ st♦♣ ❝♦♥s♦♥❛♥t✲✈♦✇❡❧ ✉tt❡r❛♥❝❡s✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙♣❡❡❝❤ ❛♥❞ ❆✉❞✐♦ Pr♦❝❡ss✐♥❣✱ ✶✵✭✼✮✿✹✼✷✕✹✽✵✱ ✷✵✵✷✳ ❬❈✳❏✳❈✳❇✉r❣❡s✭✶✾✾✽✮❪ ❈✳❏✳❈✳❇✉r❣❡s✳ ❆ t✉t♦r✐❛❧ ♦♥ ❙✉♣♣♦rt ❱❡❝t♦r ▼❛❝❤✐♥❡s ❢♦r P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥✳ ❉❛t❛ ♠✐♥✐♥❣ ❛♥❞ ❦♥♦✇❧❡❞❣❡ ❞✐s❝♦✈❡r②✱ ✶✾✾✽✳ ❬❈❧❛r❦s♦♥ ❛♥❞ ▼♦r❡♥♦✭✶✾✾✼✮❪ P❤✐❧✐♣ ❈❧❛r❦s♦♥ ❛♥❞ P❡❞r♦ ❏ ▼♦r❡♥♦✳ ❖♥ t❤❡ ✉s❡ ♦❢ s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡ ❢♦r ♣❤♦♥❡t✐❝ ❝❧❛ss✐✜❝❛t✐♦♥✳ ■❊❊❊ ■♥t✳ ❈♦♥❢✳ ❆❝♦✉st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ♣❛❣❡s ✶✷✺✶✕✶✷✺✹✱ ✶✾✾✼✳ ✷✺✽

❬❈♦✐❢♠❛♥ ❛♥❞ ▼❛❣❣✐♦♥✐✭✷✵✵✻✮❪ ❘ ❘ ❈♦✐❢♠❛♥ ❛♥❞ ▼ ▼❛❣❣✐♦♥✐✳ ❉✐✛✉s✐♦♥ ✇❛✈❡❧❡ts✳ ❆♣♣❧✳ ❈♦♠♣✉t❛t✳ ❍❛r♠♦♥✳ ❆♥❛❧✱ ✷✶✭✶✮✿✺✸✕✾✹✱ ✷✵✵✻✳ ❬❈♦✈❡r ❛♥❞ ❍❛rt✭✶✾✻✼✮❪ ❚ ▼ ❈♦✈❡r ❛♥❞ P ❊ ❍❛rt✳ ◆❡❛r❡st ♥❡✐❣❤❜♦r ♣❛tt❡r♥ ❝❧❛ss✐✜❝❛t✐♦♥✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ■♥❢♦r♠❛t✐♦♥ ❚❤❡♦r②✱ ✶✸✭✶✮✿✷✶✕✷✼✱ ✶✾✻✼✳ ❬❈r♦✇❧❡②✭✶✾✽✼✮❪ ❏✳ ❈r♦✇❧❡②✳ ❆ r❡♣r❡s❡♥t❛t✐♦♥ ❢♦r ✈✐s✉❛❧ ✐♥❢♦r♠❛t✐♦♥✳ ❘♦❜♦t✐❝✳ ■♥st✳ ❈❛r♥❡❣✐❡✲▼❡❧❧♦♥ ❯♥✐✈❡rs✐t②✱ ❚❡❝❤✳ ❘❡♣✳ ❈▼❯✲❘■✲❚❘✱ ♣❛❣❡s ✽✷✕ ✽✼✱ ✶✾✽✼✳ ❬❈r✉t❝❤✜❡❧❞✳❏ ❛♥❞ ❙❤❛✇✳❘✭✶✾✽✻✮❪ P❛❝❦❛r❞✳◆ ❈r✉t❝❤✜❡❧❞✳❏✱ ❋❛r♠❡r✳❏ ❛♥❞ ❙❤❛✇✳❘✳ ❈❤❛♦s✳ ❙❝✐❡♥t✐✜❝ ❆♠❡r✐❝❛♥✱ ✷✺✺✿✸✽✕✹✾✱ ✶✾✽✻✳ ❬❉ ❇ ❍❛♥❝❤❛t❡ ❛♥❞ ▼♦✉r②❛✭✷✵✶✵✮❪ ▼❛♥♦❥ P❛✇❛r ❱✐❥❛② P♦♣❤❛❧❡ ❉ ❇ ❍❛♥✲ ❝❤❛t❡✱ ▼♦❤✐♥✐ ◆❛❧❛✇❛❞❡ ❛♥❞ Pr❛❜❤❛t❤ ❑✉♠❛r ▼♦✉r②❛✳ ❱♦❝❛❧ ❞✐❣✐t r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ❛rt✐✜❝❛✐❧ ♥❡✉r❛❧ ♥❡t✇♦r❦✳ ■♥ ■♥ Pr♦❝✳ ■❊❊❊✳ ✐♥t✳ ❈♦♥❢✳ ♦♥ ❈♦♠♣✉t❡r ❊♥❣✐♥♥❡r✐♥❣ ❛♥❞ ❚❡❝❤♥♦❧♦❣②✱ ♣❛❣❡s ✽✽✕✾✶✱ ✷✵✶✵✳ ❬❉✳ ❉✉tt❛ ▼❛❥✉♠❞❡r ❛♥❞ P❛❧✭✶✾✼✻✮❪ ❆✳ ❑✳ ❉❛tt❛ ❉✳ ❉✉tt❛ ▼❛❥✉♠❞❡r ❛♥❞ ❙✳ ❑✳ P❛❧✳ ❈♦♠♣✉t❡r r❡❝♦❣♥✐t✐♦♥ ♦❢ t❡❧✉❣✉ ✈♦✇❡❧ s♦✉♥❞s✳ ❏✳ ❈♦r♥♣✉t✳ ❙❛❝✳ ■♥❞✐❛✱ ✼✿✶✹✕✷✵✱ ✶✾✼✻✳ ❬❉ ❑ ❍❛♠♠♦♥❞ ❛♥❞ ●r✐❜♦♥✈❛❧✭✷✵✶✵✮❪ P ❱❛♥❞❡r❣❤❡②♥st ❉ ❑ ❍❛♠♠♦♥❞ ❛♥❞ ❘ ●r✐❜♦♥✈❛❧✳ ❲❛✈❡❧❡ts ♦♥ ❣r❛♣❤ ✈✐❛ s♣❡❝tr❛❧ ❣r❛♣❤ t❤❡♦r②✳ ❆♣♣❧✳ ❈♦♠♣✉t❛t✳❍❛r♠♦♥✳ ❆♥❛❧✱ ✷✵✶✵✳ ❬❉❛✉❜❡❝❤✐❡s✭✶✾✾✵✮❪ ■♥❣r✐❞ ❉❛✉❜❡❝❤✐❡s✳ ❚❤❡ ✇❛✈❡❧❡t tr❛♥s❢♦r♠✱ t✐♠❡✲ ❢r❡q✉❡♥❝② ❧♦❝❛❧✐③❛t✐♦♥ ❛♥❞ s✐❣♥❛❧ ❛♥❛❧②s✐s✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ■♥❢♦r♠❛t✐♦♥ ❚❤❡♦r②✱ ✸✻✭✺✮✿✾✻✶✕✶✵✵✺✱ ✶✾✾✵✳ ❬❉❡❧❧❡r✳❏✳❘ ❛♥❞ Pr♦❛❦✐s✳❏✳●✭✷✵✵✵✮❪ ❍❛♥s❡♥✳❏✳❍✳▲ ❉❡❧❧❡r✳❏✳❘ Pr♦❛❦✐s✳❏✳●✳ ❉✐s❝r❡t❡✲t✐♠❡ ♣r♦❝❡ss✐♥❣ ♦❢ s♣❡❡❝❤ s✐❣♥❛❧s✳ ❡❞✳✱ ■❊❊❊ Pr❡ss✿◆❡✇②♦r❦✱ ✷✵✵✵✳

❛♥❞ ❙❡❝♦♥❞

❬❉❤❛♥❛♥❥❛②❛✳◆ ❛♥❞ ❨❛❣♥❛♥❛r❛②❛♥❛✳❇✭✷✵✵✹✮❪ ●✉r✉♣r❛s❛❞✳❙ ❉❤❛♥❛♥❥❛②❛✳◆ ❛♥❞ ❨❛❣♥❛♥❛r❛②❛♥❛✳❇✳ ❙♣❡❛❦❡r s❡❣♠❡♥t❛t✐♦♥ ❢❡❛t✉r❡s ❛♥❞ ♥❡✉r❛❧ ♥❡t✲ ✇♦r❦ ♠♦❞❡❧s✳ ■♥ ✐♥ Pr♦❝✳ ✶✶t❤ ■♥t❡r♥❛t✐♦♥❛❧ ❈♦♥❢❡r❡♥❝❡ ♦♥ ◆❡✉r❛❧ ■♥❢♦r♠❛t✐♦♥ Pr♦❝❡ss✐♥❣✱ ■❈❖◆■P✱ ■◆❉■❆✱ ✷✵✵✹✳ ❬❉▼■✭✮❪ ❉▼■✳ ❚■▼■❚✳ ❆❝♦✉st✐❝✲P❤♦♥❡t✐❝ ❈♦♥t✐♥✉♦✉s ❙♣❡❡❝❤ ❈♦r♣✉s✱ ✶✾✾✵✳ ❬❉✉❞❛ ❛♥❞ ❊✳❍❛rt✭✶✾✼✸✮❪ ❘✐❝❤❛r❞ ❖✳ ❉✉❞❛ ❛♥❞ P❡t❡r ❊✳❍❛rt✳ P❛tt❡r♥ ❈❧❛s✲ s✐✜❝❛t✐♦♥ ❛♥❞ ❙❝❡♥❡ ❆♥❛❧②s✐s✳ ❲✐❧❡② ■♥t❡rs❝✐❡♥❝❡✱ ✶✾✼✸✳ ✷✺✾

❬❉✉❞❧❡②✭✶✾✹✵✮❪ ❍✳ ❉✉❞❧❡②✳ ❚❤❡ ❝❛rr✐❡r ♥❛t✉r❡ ♦❢ s♣❡❡❝❤✳ ❇❡❧❧ ❙②st❡♠s ❚❡❝❤✲ ♥✐❝❛❧ ❏♦✉r♥❛❧✱ ✶✾✿✹✾✺✕✺✶✸✱ ✶✾✹✵✳ ❬❉✉t❝❤✭✮❪ ❉✉t❝❤✳ ❚❤❡ ❙♣♦❦❡♥ ❤tt♣✿✴✴✇✇✇✳❡❧✐s✳r✉❣✳❛❝✳❜❡✴❝❣♥✴✳

❉✉t❝❤

❈♦r♣✉s✳

✿

❬❊❝❦♠❛♥♥✳❏✳P ❛♥❞ ❘✉❡❧❧❡✳❉✭✶✾✽✺✮❪ ❊❝❦♠❛♥♥✳❏✳P ❛♥❞ ❘✉❡❧❧❡✳❉✳ ❊r❣♦❞✐❝ t❤❡♦r② ♦❢ ❝❤❛♦s ❛♥❞ str❛♥❣❡ ❛ttr❛❝t♦rs✳ ❘❡✈✐❡✇ ♦❢ ▼♦❞❡r♥ P❤②s✐❝s✱ ✺✼✿✻✶✼✕✻✺✻✱ ✶✾✽✺✳ ❬❊✈❛♥❣❡❧✐st❛✭✶✾✾✸✮❪ ●✐❛♥♣❛♦❧♦ ❊✈❛♥❣❡❧✐st❛✳ P✐t❝❤ s②♥❝❤r♦♥♦✉s ✇❛✈❡❧❡t r❡♣✲ r❡s❡♥t❛t✐♦♥ ♦❢ s♣❡❡❝❤ ❛♥❞ ♠✉s✐❝ s✐❣♥❛❧s✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✹✶✭✶✷✮✿✸✸✶✸✕✸✸✸✵✱ ✶✾✾✸✳ ❬❋❛r♠❡r✳❏✳❉ ❛♥❞ ❙✐❞♦r♦✇✐❝❤✳❏✳❉✭✶✾✽✽✮❪ ❋❛r♠❡r✳❏✳❉ ❛♥❞ ❙✐❞♦r♦✇✐❝❤✳❏✳❉✳ ❊①♣❧♦✐t✐♥❣ ❝❤❛♦s t♦ ♣r❡❞✐❝t t❤❡ ❢✉t✉r❡ ❛♥❞ r❡❞✉❝❡ ♥♦✐s❡✳ ❊✈♦❧✉t✐♦♥✱ ▲❡❛r♥✐♥❣ ❛♥❞ ❈♦❣♥✐t✐♦♥ ✭❨✳▲❡❡ ❡❞✳✮✱ ✶✾✽✽✳ ❬❋❛r♦♦q ❛♥❞ ❙✳❉❛tt❛✭✷✵✵✹✮❪ ❖✳ ❋❛r♦♦q ❛♥❞ ❙✳❉❛tt❛✳ ❲❛✈❡❧❡t ❜❛s❡❞ r♦❜✉st s✉❜✲❜❛♥❞ ❢❡❛t✉r❡s ❢♦r ♣❤♦♥❡♠❡ r❡❝♦❣♥✐t✐♦♥✳ ■❊❊ Pr♦❝✳✲❱✐s✳ ■♠❛❣❡ ❙✐❣✲ ♥❛❧ Pr♦❝❡ss✱ ✶✺✶✭✸✮✿✶✽✼✕✶✾✸✱ ✷✵✵✹✳ ❬❋❡✉❡r ❛♥❞ ●♦♦❞✇✐♥✭✶✾✾✻✮❪ ❆✳ ❋❡✉❡r ❛♥❞ ●✳ ●♦♦❞✇✐♥✳ ❙❛♠♣❧✐♥❣ ✐♥ ❉✐❣✐t❛❧ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣ ❛♥❞ ❈♦♥tr♦❧✳ ❇✐r❦❤❛✉s❡r ❇♦st♦♥✱ ✶✾✾✻✳ ❬❋❧❡t❝❤❡r✭✶✾✷✷✮❪ ❍✳ ❋❧❡t❝❤❡r✳ ❚❤❡ ♥❛t✉r❡ ♦❢ s♣❡❡❝❤ ❛♥❞ ✐ts ✐♥t❡r♣r❡t❛t✐♦♥s✳ ❇❡❧❧ ❙②st✳ ❚❡❝❤✳ ❏✱ ✶✿✶✷✾✕✶✹✹✱ ✶✾✷✷✳ ❬❋❧❡t❝❤❡r ❛♥❞ ●❛❧t✭✶✾✺✵✮❪ ❍✳ ❋❧❡t❝❤❡r ❛♥❞ ❘✳ ❍✳ ●❛❧t✳ P❡r❝❡♣t✐♦♥ ♦❢ s♣❡❡❝❤ ❛♥❞ ✐ts r❡❧❛t✐♦♥ t♦ t❡❧❡♣❤♦♥②✳ ❏✳ ❆❝♦✉st✐❝✳ ❙♦❝✳ ❆♠❡r✐❝❛✱ ✷✷✿✽✾✕✶✺✶✱ ✶✾✺✵✳ ❬❋❧❡t❝❤❡r ❛♥❞ ▼✉♥s♦♥✭✶✾✸✼✮❪ ❍✳ ❋❧❡t❝❤❡r ❛♥❞ ❲✳ ❆✳ ▼✉♥s♦♥✳ ❘❡❧❛t✐♦♥ ❜❡✲ t✇❡❡♥ ❧♦✉❞♥❡ss ❛♥❞ ♠❛s❦✐♥❣✳ ❏✳ ❆❝♦✉st✐❝✳ ❙♦❝✳ ❆♠❡r✐❝❛✱ ✾✿✶✕✶✵✱ ✶✾✸✼✳ ❬❋❧❡t❝❤❡r✭✶✾✽✼✮❪ ❘✳ ❋❧❡t❝❤❡r✳ ❧❡②✱❙❡❝♦♥❞ ❡❞✐t✐♦♥✱ ✶✾✽✼✳

Pr❛❝t✐❝❛❧ ▼❡t❤♦❞s ♦❢ ❖♣t✐♠✐③❛t✐♦♥✳

✇✐✲

❬❋♦✇❧❡r ❛♥❞ ❍✉❛✭✷✵✵✷✮❪ ❏❛♠❡s✳ ❊ ❋♦✇❧❡r ❛♥❞ ▲✐ ❍✉❛✳ ❲❛✈❡❧❡t tr❛♥s❢♦r♠s ❢♦r ✈❡❝t♦r ✜❡❧❞s ✉s✐♥❣ ♦♠♥✐❞✐r❡❝t✐♦♥❛❧❧② ❜❛❧❛♥❝❡❞ ♠✉❧t✐✇❛✈❡❧❡ts✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✺✵✭✶✷✮✿✸✵✶✽✕✸✵✷✼✱ ✷✵✵✷✳ ❬❋r❛s❡r✳❆✳▼ ❛♥❞ ❙✇✐♥♥❡②✳❍✳▲✭✶✾✽✻✮❪ ❋r❛s❡r✳❆✳▼ ❛♥❞ ❙✇✐♥♥❡②✳❍✳▲✳ ■♥❞❡✲ ♣❡♥❞❡♥t ❝♦♦r❞✐♥❛t❡s ❢♦r str❛♥❣❡ ❛ttr❛❝t♦rs ❢r♦♠ ♠✉t✉❛❧ ✐♥❢♦r♠❛t✐♦♥✳ P❤②s✳ ❘❡✈✳ ❆✱ ✸✸✿✶✶✸✹✕✶✶✹✵✱ ✶✾✽✻✳ ✷✻✵

■♥tr♦❞✉❝t✐♦♥ t♦ P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥ ✲ ❙t❛t✐st✐❝❛❧✱ ❙tr✉❝t✉r❛❧✱ ◆❡✉❛r❛❧ ❛♥❞ ❋✉③③② ❧♦❣✐❝ ❛♣♣r♦❛❝❤✳ ❲❆❚ ❛♥❞ ❚❞ ❙❝✐❡♥t✐✜❝✱ ✶✾✾✾✳

❬❋r✐❡❞♠❛♥ ❛♥❞ ❑❛♥❞❡❧✭✶✾✾✾✮❪ ▼✳ ❋r✐❡❞♠❛♥ ❛♥❞ ❆✳ ❑❛♥❞❡❧✳

❬❋r② ❛♥❞ ❉❡♥❡s✭✶✾✺✾✮❪ ❉✳ ❇✳ ❋r② ❛♥❞ P✳ ❉❡♥❡s✳ ❚❤❡ ❞❡s✐❣♥ ❛♥❞ ♦♣❡r❛t✐♦♥ ♦❢ t❤❡ ♠❡❝❤❛♥✐❝❛❧ s♣❡❡❝❤ r❡❝♦❣♥✐③❡r ❛t ✉♥✐✈❡rs✐t② ❝♦❧❧❡❣❡ ❧♦♥❞♦♥✳

❇r✐t✐s❤ ■♥st✳ ❘❛❞✐♦ ❊♥❣r✱ ✶✾✭✹✮✿✷✶✶✕✷✷✾✱ ✶✾✺✾✳

❬●✐♠s♦♥✳❆✭✶✾✼✷✮❪ ●✐♠s♦♥✳❆✳

❏✳

❆♥ ■♥tr♦s✉❝t✐♦♥ t♦ Pr♦♥✉♥❝✐❛t✐♦♥ ♦❢ ❊♥❣❧✐s❤✳

❊❞✇❛r❞ ❆r♥♦❧❞ ▲t❞✳✒ ✶✾✼✷✳ ❬●♦❧❞ ❛♥❞ ▼♦r❣❛♥✭✷✵✵✵✮❪ ❇ ●♦❧❞ ❛♥❞ ◆ ▼♦r❣❛♥✳

Pr♦❝❡ss✐♥❣✳

❙♣❡❡❝❤ ❛♥❞ ❆✉❞✐♦ ❙✐❣♥❛❧

◆❡✇❨♦r❦✿❏♦❤♥ ❲✐❧❡② ❛♥❞ ❙♦♥s ■♥❝✱ ✷✵✵✵✳

❬●♦♣❛❧❛❦r✐s❤♥❛ ❛♥❞ ❙✳P✳❑✐s❤♦r❡✭✷✵✵✺✮❪ ❙✳ ❏♦s❤✐ ❘✳ ❑✉♠❛r ❙✳ ❙✐♥❣❤ ❘✳ ◆✳ ❱ ❙✐t❛r❛♠ ●♦♣❛❧❛❦r✐s❤♥❛✱ ❆✳❘✳❈❤✐tt✉r✐ ❛♥❞ ❙✳P✳❑✐s❤♦r❡✳ ❉❡✈❡❧♦♣♠❡♥t ♦❢ ✐♥❞✐❛♥ ❧❛♥❣✉❛❣❡ s♣❡❡❝❤ ❞❛t❛❜❛s❡s ❢♦r ❧❛r❣❡ ✈♦❝❛❜✉❧❛r② s♣❡❡❝❤ r❡❝♦❣✲ ■♥ Pr♦❝✳ t❤❡ ■♥t❡r♥❛t✐♦♥❛❧❈♦♥❢❡r❡♥❝❡ ♦♥ ❙♣❡❡❝❤ ❛♥❞ ❈♦♠♣✉t❡r ✭❙P❊❈❖▼✮✱ P❛tr❛s✱ ✷✵✵✺✳

♥✐t✐♦♥ s②st❡♠s✳

❬●r❛ss❜❡r❣❡r ❛♥❞ Pr♦❝❛❝❝✐❛✭✶✾✽✸✮❪ P✳ ●r❛ss❜❡r❣❡r ❛♥❞ ■✳ Pr♦❝❛❝❝✐❛✳ ❊st✐♠❛✲ t✐♦♥ ♦❢ t❤❡ ❦♦❧♠♦❣♦r♦✈ ❡♥tr♦♣② ❢r♦♠ ❛ ❝❤❛♦t✐❝ s✐❣♥❛❧✳

P❤②s✳ ❘❡✈✳ ❆✱ ✷✽

✭✹✮✿✸✹✻✕✸✺✷✱ ✶✾✽✸✳ ❬●r❡❡♥❜❡r❣✭✶✾✾✾✮❪ ❙ ●r❡❡♥❜❡r❣✳ ❙♣❡❛❦✐♥❣ ✐♥ s❤♦t❤❛♥❞✲❛ s②❧❧❛❜❧❡ ❝❡♥tr✐❝ ♣❡r✲ s♣❡❝t✐✈❡ ❢♦r ✉♥❞❡rst❛♥❞✐♥❣ ♣r♦♥♦✉♥❝✐❛t✐♦♥ ✈❛r✐❛t✐♦♥✳

♥✐❝❛t✐♦♥✱ ✷✾✭✷✲✹✮✿✶✺✾✕✶✼✻✱ ✶✾✾✾✳

❙♣❡❡❝❤ ❈♦♠♠✉✲

❬●✉♣t❛ ❛♥❞ ●✐❧❜❡rt✭✷✵✵✶✮❪ ▼✳ ●✉♣t❛ ❛♥❞ ❆✳ ●✐❧❜❡rt✳ ❘♦❜✉st s♣❡❡❝❤ r❡❝♦❣✲

■♥ Pr♦❝✳ ■❊❊❊ ❲♦r❦s❤♦♣ ♦♥ ❆✉t♦♠❛t✐❝ ❙♣❡❡❝❤ ❘❡❝♦❣♥t✐♦♥ ❛♥❞ ❯♥❞❡rst❛♥❞✐♥❣✱ ■t❛❧②✱ ♣❛❣❡s ✹✹✺✕✹✹✽✱

♥✐t✐♦♥ ✉s✐♥❣ ✇❛✈❡❧❡t ❝♦❡✣❝✐❡♥t ❢❡❛t✉r❡s✳ ✷✵✵✶✳

❬❍✳ ❉✉❞❧❡② ❛♥❞ ❲❛t❦✐♥s✭✶✾✸✾✮❪ ❘✳ ❘✳ ❘✐❡s③ ❍✳ ❉✉❞❧❡② ❛♥❞ ❙✳ ❆✳ ❲❛t❦✐♥s✳ ❆ s②♥t❤❡t✐❝ s♣❡❛❦❡r✳ ✶✾✸✾✳ ❬❍❛❛r✭✶✾✶✵✮❪ ❆❧❢r❡❞ ❍❛❛r✳ ❩✉r t❤❡♦r✐❡ ❞❡r ♦rt❤♦❣♦♥❛❧❡♥ ❢✉♥❦t✐♦♥❡♥s②st❡♠❡✳

▼❛t❤❡♠❛t✐s❝❤❡ ❆♥♥❛❧❡♥✱ ✻✾✿✸✸✶✕✸✼✶✱ ✶✾✶✵✳

❬❍❛♥❞✭✶✾✽✶✮❪ ❉ ❏ ❍❛♥❞✳

❉✐s❝r✐♠✐♥❛t✐♦♥ ❛♥❞ ❈❧❛ss✐✜❝❛t✐♦♥✳

◆❡✇②♦r❦✿❲✐❧❡②✱

✶✾✽✶✳ ❬❍❛rr✐♥❣t♦♥ ❛♥❞ ❈❛ss✐❞②✭✶✾✾✾✮❪ ❏✳ ❍❛rr✐♥❣t♦♥ ❛♥❞ ❙✳ ❈❛ss✐❞②✳

❙♣❡❡❝❤ ❆❝♦✉st✐❝s✳

❑❧✉✇❡r ❆❝❛❞❡♠✐❝ P✉❜❧✐s❤❡rs✱ ✶✾✾✾✳

✷✻✶

❚❡❝❤♥✐q✉❡s ✐♥

❬❍❛②❦✐♥✭✷✵✵✹✮❪ ❙✳ ❍❛②❦✐♥✳ ◆❡✉r❛❧ ◆❡t✇♦r❦s✿ ❆ ❈♦♠♣r❡❤❡♥s✐✈❡ ❋♦✉♥❞❛t✐♦♥✳ Pr❡♥t✐❝❡ ❍❛❧❧ ♦❢ ■♥❞✐❛ P✈t✳ ▲t❞✱ ✷✵✵✹✳ ❬❍❡❛rst✭✶✾✾✽✮❪ ▼✳ ❆✳ ❍❡❛rst✳ ❙✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡s✳ ■❊❊❊ ■♥t❡❧❧✐❣❡♥t ❙②st❡♠s✱ ♣❛❣❡s ✶✽✕✷✽✱ ✶✾✾✽✳ ❬❍✐❧❜♦r♥✳❘✭✶✾✾✹✮❪ ❍✐❧❜♦r♥✳❘✳ ❈❤❛♦s ❛♥❞ ◆♦♥❧✐♥❡❛r ❉②♥❛♠✐❝s✳ ❖①❢♦r❞✿ ❖①✲ ❢♦r❞ ❯♥✐✈❡rs✐t② Pr❡ss✱ ✶✾✾✹✳ ❬❍♦❧♠❡s ❛♥❞ ❍♦❧♠❡s✭✷✵✵✶✮❪ ❏ ❍♦❧♠❡s ❛♥❞ ❲ ❍♦❧♠❡s✳ ❙♣❡❡❝❤ ❙②♥t❤❡s✐s ❛♥❞ ❘❡❝♦❣♥✐t✐♦♥✳ ❚❛✐❧♦r ❛♥❞ ❋r❛♥❝✐s✱ ✷✵✵✶✳ ❬❍♦♥❣②✉ ▲✐❛♦ ❛♥❞ ❈♦❝❦❜✉r♥✭✷✵✵✹✮❪ ▼r✐♥❛❧ ❑r ❍♦♥❣②✉ ▲✐❛♦ ❛♥❞ ❇r✉❝❡ ❋ ❈♦❝❦❜✉r♥✳ ❊✣❝✐❡♥t ❛r❝❤✐t❡❝t✉r❡s ❢♦r ✶✲❞ ❛♥❞ ✷✲❞ ❧✐❢t✐♥❣ ❜❛s❡❞ ✇❛✈❡❧❡t tr❛♥s❢♦r♠✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✺✷✭✺✮✿✶✸✶✺✕✶✸✷✻✱ ✷✵✵✹✳ ❬❍✉tt♦♥✳▲✳❱✭✶✾✾✷✮❪ ❍✉tt♦♥✳▲✳❱✳ ❯s✐♥❣ st❛t✐st✐❝s t♦ ❛ss❡ss t❤❡ ♣❡r❢♦r♠❛♥❝❡ ♦❢ ♥❡✉r❛❧ ♥❡t✇♦r❦ ❝❧❛ss✐✜❡rs✳ ❏♦❤♥s✲❍♦♣❦✐♥s✲❆P▲ ❚❡❝❤♥✐❝❛❧ ❉✐❣❡st✱ ✶✸ ✭✷✮✱ ✶✾✾✷✳ ❬■❞❛♥ ❘❛♠ ❛♥❞ ❈♦❤❡♥✭✷✵✶✶✮❪ ▼✐❝❤❛❡❧ ❊❧❛❞ ■❞❛♥ ❘❛♠ ❛♥❞ ■sr❛❡❧ ❈♦❤❡♥✳ ●❡♥❡r❛❧✐③❡❞ tr❡❡✲❜❛s❡❞ ✇❛✈❡❧❡t tr❛♥s❢♦r♠✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❙✐❣♥❛❧ Pr♦✲ ❝❡ss✐♥❣✱ ✺✾✭✾✮✿✹✶✾✾✕✹✷✵✾✱ ✷✵✶✶✳ ❬■♥❣❡ ●❛✈❛t ❛♥❞ ❧❛♥❝✉✭✶✾✾✽✮❪ ●❜❛r✐❛❧ ❈♦st❛❝❤❡ ■♥❣❡ ●❛✈❛t ❛♥❞ ❈❧❛✉❞✐❛ ❧❛♥❝✉✳ ❘♦❜✉st s♣❡❡❝❤ r❡❝♦❣♥✐③❡r ✉s✐♥❣ ♠✉❧t✐❝❧❛ss s✈♠✳ ■♥ ✼t❤ ❙❡♠✐❛♥r ♦♥ ◆❡✉r❛❧ ◆❡t✇♦r❦ ❆♣♣❧✐❝❛t✐♦♥s ✐♥ ❊❧❡❝tr✐❝❛❧ ❊♥❣✐♥❡❡r✐♥❣✱ ✶✾✾✽✳ ❬■♦s✐❢ ▼♣♦r❛s ❛♥❞ ❋❛❦♦t❛❦✐s✭✷✵✵✻✮❪ P❛♥❛❣✐♦t✐s ❩❡r✈❛s ■♦s✐❢ ▼♣♦r❛s✱ ❚♦❞♦r ●❛♥❝❤❡✈ ❛♥❞ ◆✐❦♦s ❋❛❦♦t❛❦✐s✳ ❘❡❝♦❣♥✐t✐♦♥ ♦❢ ❣r❡❡❦ ♣❤♦♥❡♠❡s ✉s✐♥❣ s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡s✳ ❆❞✈❛♥❝❡s ✐♥ ❆rt✐✜❝✐❛❧ ■♥t❡❧❧✐❣❡♥❝❡✱ ✸✾✺✺✴✷✵✵✻✿✷✾✵✕✸✵✵✱ ✷✵✵✻✳ ❬■s❤✐③❛❦❛✳❑ ❛♥❞ ❋❧❛♥❛❣❛♥✳❏✳▲✭✶✾✼✷✮❪ ■s❤✐③❛❦❛✳❑ ❛♥❞ ❋❧❛♥❛❣❛♥✳❏✳▲✳ ❙②♥t❤❡✲ s✐s ♦❢ ✈♦✐❝❡❞ s♦✉♥❞s ❢r♦♠ ❛ t✇♦â⑨➇♠❛ss ♠♦❞❡❧ ♦❢ t❤❡ ✈♦❝❛❧ ❝❤♦r❞s✳ ❇❡❧❧ ❙②st❡♠ ❚❡❝❤♥✐❝❛❧ ❏♦✉r♥❛❧✱ ✺✶✭✶✮✿✶✷✸✸✕✶✷✻✽✱ ✶✾✼✷✳ ❬■t❛❦✉r❛✭✶✾✼✷✮❪ ■t❛❦✉r❛✳ ❆♥ ❛✉❞✐♦ rs♣♦♥s❡ ✉♥✐t ❜❛s❡❞ ♦♥ ♣❛rt✐❛❧ ❛✉t♦❝♦rr❡✲ ❧❛t✐♦♥✳ ✶✾✼✷✳ ❬■t❛❦✉r❛✭✶✾✼✺✮❪ ❋✳ ■t❛❦✉r❛✳ ▼✐♥✐♠✉♠ ♣r❡❞✐❝t✐♦♥ r❡s✐❞✉❛❧ ♣r✐♥❝✐♣❧❡ ❛♣♣❧✐❡❞ t♦ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■❊❊❊ ❚r❛♥s✳ ❆❝♦✉st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦✲ ❝❡ss✐♥❣✱ ❆❙❙P ✷✸✿✺✼✕✼✷✱ ✶✾✼✺✳ ✷✻✷

❬■t❛❦✉r❛ ❛♥❞ ❙❛✐t♦✭✶✾✼✵✮❪ ❋✳ ■t❛❦✉r❛ ❛♥❞ ❙✳ ❙❛✐t♦✳ ❆ st❛t✐st✐❝❛❧ ♠❡t❤♦❞ ❢♦r ❡st✐♠❛t✐♦♥ ♦❢ s♣❡❡❝❤ s♣❡❝tr❛❧ ❞❡♥s✐t② ❛♥❞ ❢♦r♠❛♥t ❢r❡q✉❡♥❝✐❡s✳ ❊❧❡❝✲ tr♦♥✐❝s ❛♥❞ ❈♦♠♠✉♥✐❝❛t✐♦♥s ✐♥ ❏❛♣❛♥✱ ✺✸ ❆✿✸✻✕✹✸✱ ✶✾✼✵✳ ❬❏✳ P❧❛tt ❛♥❞ ❙❜❛✇❡✲❚❛②❧♦r✭✶✾✾✾✮❪ ◆✳ ❈❤r✐st✐❛♥✐♥✐ ❏✳ P❧❛tt ❛♥❞ ❏✳ ❙❜❛✇❡✲ ❚❛②❧♦r✳ ▲❛r❣❡ ♠❛r✐❣✐♥ ❞❛❣s ❢♦r ♠✉❧t✐❝❧❛ss ❝❧❛ss✐✜❝❛t✐♦♥✳ ❆❞✈❛♥❝❡s ✐♥ ◆❡✉r❛❧ ■♥❢♦r♠❛t✐♦♥ Pr♦❝❡ss✐♥❣ ❙②st❡♠✱ ✶✷✿✺✹✼✕✺✺✸✱ ✶✾✾✾✳ ❬❏❛♥ss❡♥✳❘✳❉✳❚ ❛♥❞ ❈♦❧❡✳❘✳❆✭✶✾✾✶✮❪ ❋❛♥t②✳▼ ❏❛♥ss❡♥✳❘✳❉✳❚ ❛♥❞ ❈♦❧❡✳❘✳❆✳ ❙♣❡❛❦❡r✲✐♥❞❡♣❡♥❞❡♥t ♣❤♦♥❡t✐❝ ❝❧❛ss✐✜❝❛t✐♦♥ ✐♥ ❝♦♥t✐♥✉♦✉s ❡♥❣❧✐s❤ ❧❡t✲ t❡rs✳ ■♥ ✐♥ Pr♦❝✳ ■❊❊❊ ■♥t❡r♥❛t✐♦♥❛❧ ❏♦✐♥t ❈♦♥❢❡r❡♥❝❡ ♦♥ ◆❡✉r❛❧ ◆❡t✲ ✇♦r❦s✱ ■❏❈◆◆✲✾✶✱ ✶✾✾✶✳ ❬❏✐♥❣ ❇❛✐✭✷✵✵✻✮❪ ❨✉❡✲❧✐♥❣ ●✉♦ ❏✐♥❣ ❇❛✐✱ ❳✉❡✲②✐♥❣ ❩❤❛♥❣✳ ❙♣❡❡❝❤ r❡❝♦❣♥✐✲ t✐♦♥ ❜❛s❡❞ ♦♥ ❛ ❝♦♠♣♦✉♥❞ ❦❡r♥❡❧ s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡✳ ■♥ Pr♦❝ ♦❢ t❤❡ ✶✶t❤ ■❊❊❊ ✐♥t❡r♥❛t✐♦♥❛❧ ❝♦♥❢❡r❡♥❝❡ ♦♥ ❈♦♠♠✉♥✐❝❛t✐♦♥ ❚❡❝❤♥♦❧♦❣②

✱ ✷✵✵✻✳

Pr♦❝❡❡❞✐♥❣s

❬❏✉❛♥❣✭✶✾✽✺✮❪ ❇✳ ❍✳ ❏✉❛♥❣✳ ▼❛①✐♠✉♠ ❧✐❦❡❧✐❤♦♦❞ ❡st✐♠❛t✐♦♥ ❢♦r ♠✐①t✉r❡ ♠✉❧✲ t✐✈❛r✐❛t❡ st♦❝❤❛st✐❝ ♦❜s❡r✈❛t✐♦♥s ♦❢ ♠❛r❦♦✈ ❝❤❛✐♥s✳ ❆❚ ❛♥❞ ❚ ❚❡❝❤✳ ❏✱ ✻✹✭✻✮✿✶✷✸✺✕✶✷✹✾✱ ✶✾✽✺✳ ❬❏✉r❛❢s❦② ❛♥❞ ▼❛rt✐♥✭✷✵✵✹✮❪ ❉❛♥✐❛❧ ❏✉r❛❢s❦② ❛♥❞ ❏❛♠❡s✳ ❍✳ ▼❛rt✐♥✳

❆♥ ■♥✲

tr♦❞✉❝t✐♦♥ t♦ ◆❛t✉r❛❧ ▲❛♥❣✉❛❣❡ Pr♦❝❡ss✐♥❣✱ ❈♦♠♣t❛t✐♦♥❛❧ ▲✐♥❣✉✐st✐❝s

✳ P❡❛rs♦♥ ❊❞✉❝❛t✐♦♥✱ ✷✵✵✹✳

❛♥❞ ❙♣❡❡❝❤ Pr♦❞✉❝t✐♦♥

❬❏✳❲✳❋♦r❣✐❡ ❛♥❞ ❈✳❉✳❋♦r❣✐❡✭✶✾✺✾✮❪ ❏✳❲✳❋♦r❣✐❡ ❛♥❞ ❈✳❉✳❋♦r❣✐❡✳ ❘❡s✉❧ts ♦❜✲ t❛✐♥❡❞ ❢r♦♠ ❛ ✈♦✇❡❧ r❡❝♦❣♥✐t✐♦♥❝♦♠♣✉t❡r ♣r♦❣r❛♠✳ ❏♦✉r♥❛❧ ♦❢ ❆❝♦✉s✲ t✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✐❝❛✱ ✸✶✿✶✹✽✵✕✶✹✽✾✱ ✶✾✺✾✳ ❬❑ ❇❡♥♥❡t ❛♥❞ ❲✉✭✷✵✵✵✮❪ ❏ ❙❤❛✇❡✲❚❛②❧♦r ❑ ❇❡♥♥❡t✱ ◆ ❈r✐st✐❛♥✐♥✐ ❛♥❞ ❉ ❲✉✳ ❊♥❧❛r❣✐♥❣ t❤❡ ♠❛r❣✐♥ ✐♥ ♣❡r❝❡♣tr♦♥ ❞❡❝✐s✐♦♥ tr❡❡s✳ ▼❛❝❤✐♥❡ ▲❡❛r♥✐♥❣✱ ✹✶✱ ✷✵✵✵✳ ❬❑✳ ❍✳ ❉❛✈✐s ❛♥❞ ❇❛❧❛s❤❡❦✭✶✾✺✷✮❪ ❘✳ ❇✐❞❞✉❧♣❤ ❑✳ ❍✳ ❉❛✈✐s ❛♥❞ ❙✳ ❇❛❧❛s❤❡❦✳ ❆✉t♦♠❛t✐❝ r❡❝♦❣♥✐t✐♦♥ ♦❢ s♣♦❦❡♥ ❞✐❣✐ts✳ ❆❝♦✉st✳ ❙♦❝✳ ❆♠❡r✐❝❛✱ ✷✹✭✻✮✿ ✻✷✼✕✻✹✷✱ ✶✾✺✷✳ ❬❑✳ ◆❛❣❛t❛ ❛♥❞ ❈❤✐❜❛✭✶✾✻✸✮❪ ❨✳ ❑❛t♦ ❑✳ ◆❛❣❛t❛ ❛♥❞ ❙✳ ❈❤✐❜❛✳ ❙♣♦❦❡♥ ❞✐❣✐t r❡❝♦❣♥✐③❡r ❢♦r ❥❛♣❛♥❡s❡ ❧❛♥❣✉❛❣❡✳ ◆❊❈ ❘❡s✳ ❉❡✈❡❧♦♣✱ ✭✻✮✱ ✶✾✻✸✳ ❬❑❛♥t③ ❛♥❞ ❙❝❤r❡✐❜❡r✳❚✭✷✵✵✸✮❪ ❍✳ ❑❛♥t③ ❛♥❞ ❙❝❤r❡✐❜❡r✳❚✳ ❙❡r✐❡s ❆♥❛❧②s✐s✳ ❈❛♠❜r✐❞❣❡ ❯♥✐✈❡rs✐t② Pr❡ss✱ ✷✵✵✸✳

✷✻✸

◆♦♥❧✐♥❡❛r ❚✐♠❡

❬❑❛r✭✷✵✶✶✮❪ ❍❛r❛♥❛t❤ ❑❛r✳ ❆s②♠♣t♦t✐❝ st❛❜✐❧✐t② ♦❢ ✜①❡❞✲♣♦✐♥t st❛t❡✲s♣❛❝❡ ❞✐❣✐t❛❧ ✜❧t❡rs ✇✐t❤ ❝♦♠❜✐♥❛t✐♦♥s ♦❢ q✉❛♥t✐③❛t✐♦♥ ❛♥❞ ♦✈❡r✢♦✇ ♥♦♥❧✐♥✲ ❡❛r✐t✐❡s✳ ❊❧s❡✈✐❡r ❏♦✉r♥❛❧ ♦❢ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✾✶✿✷✻✻✼✕✷✻✼✵✱ ✷✵✶✶✳ ❬❑❡❛r♥❡②✳▼✳❏ ❛♥❞ ❙t❛r❦✳❏✭✶✾✾✷✮❪ ❑❡❛r♥❡②✳▼✳❏ ❛♥❞ ❙t❛r❦✳❏✳ ❆♥ ✐♥tr♦❞✉❝t✐♦♥ t♦ ❝❤❛♦t✐❝ s✐❣♥❛❧ ♣r♦❝❡ss✐♥❣✳ ●❊❈ ❏♦✉r♥❛❧ ♦❢ ❘❡s❡❛r❝❤✱ ✶✵✭✶✮✿✺✷✕✺✽✱ ✶✾✾✷✳ ❬❑❡♥♥❡❧✳▼✳❇ ❛♥❞ ❆❜❛r❜❛♥❡❧✳❍✳❉✳■✭✶✾✾✷✮❪ ❇r♦✇♥✳❘ ❑❡♥♥❡❧✳▼✳❇ ❛♥❞ ❆❜❛r✲ ❜❛♥❡❧✳❍✳❉✳■✳ ❉❡t❡r♠✐♥✐♥❣ ❡♠❜❡❞❞✐♥❣ ❞✐♠❡♥s✐♦♥ ❢♦r ♣❤❛s❡✲s♣❛❝❡ r❡❝♦♥✲ str✉❝t✐♦♥ ✉s✐♥❣ ❛ ❣❡♦♠❡tr✐❝❛❧ ❝♦♥str✉❝t✐♦♥✳ P❤②s✳ ❘❡✈✳ ❆✱ ✹✺✿✸✹✵✸✕✸✹✶✶✱ ✶✾✾✷✳ ❬❑❡✈✐♥ ▼✳ ■♥❞r❡❜♦✭✷✵✵✺✮❪ ▼✐❝❤❛❡❧ ❚✳ ❏♦❤♥s♦♥ ❑❡✈✐♥ ▼✳ ■♥❞r❡❜♦✱ ❘✐❝❤❛r❞ ❏✳ P♦✈✐♥❡❧❧✐✳ ❚❤✐r❞✲♦r❞❡r ♠♦♠❡♥ts ♦❢ ✜❧t❡r❡❞ s♣❡❡❝❤ s✐❣♥❛❧s ❢♦r r♦❜✉st s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■♥ ✐♥ Pr♦❝✳ ■♥t✳ ❈♦♥❢✳ ♦♥ ◆♦♥✲▲✐♥❡❛r ❙♣❡❡❝❤ Pr♦✲ ❝❡ss✐♥❣✱ ♣❛❣❡s ✶✺✶✕✶✺✼✱ ✷✵✵✺✳ ❬❑✐✲❙❡♦❦✲❑✐♠ ❛♥❞ ❍❡❡✲❨❡✉♥❣✲❍✇❛♥❣✭✶✾✾✶✮❪ ❑✐✲❙❡♦❦✲❑✐♠ ❛♥❞ ❍❡❡✲❨❡✉♥❣✲ ❍✇❛♥❣✳ ❆ st✉❞② ♦♥ t❤❡ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ♦❢ ❦♦r❡❛♥ ♣❤♦♥❡♠❡s ✉s✐♥❣ r❡✲ ❝✉rr❡♥t ♥❡✉r❛❧ ♥❡t✇♦r❦ ♠♦❞❡❧s✳ ❚r❛♥s❛❝t✐♦♥s✲♦❢✲t❤❡✲❑♦r❡❛♥✲■♥st✐t✉t❡✲ ♦❢✲❊❧❡❝tr✐❝❛❧✲❊♥❣✐♥❡❡rs✱ ✹✵✭✽✮✱ ✶✾✾✶✳ ❬❑♦❤♦♥❡♥✭✶✾✽✽✮❪ ❚✳ ❑♦❤♦♥❡♥✳ ❆♥ ✐♥tr♦❞✉❝t✐♦♥ t♦ ♥❡✉r❛❧ ❝♦♠♣✉t✐♥❣✳ ◆❡t✇♦r❦s✱ ✶✿✸✕✶✻✱ ✶✾✽✽✳

◆❡✉r❛❧

❬❑♦✐③✉♠✐✳❚ ❛♥❞ ❍✐r♦♠✐ts✉✳❙✭✶✾✽✺✮❪ ❚❛♥✐❣✉❝❤✐✳❙ ❑♦✐③✉♠✐✳❚ ❛♥❞ ❍✐✲ r♦♠✐ts✉✳❙✳ ●❧♦tt❛❧ s♦✉r❝❡✲✈♦❝❛❧ tr❛❝t ✐♥t❡r❛❝t✐♦♥✳ ❏♦✉r♥❛❧ ♦❢ t❤❡ ❆❝♦✉st✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✐❝❛✱ ✼✽✭✶✮✿✶✺✹✶✕✶✺✹✼✱ ✶✾✽✺✳ ❬❑♦✐③✉♠✐✳❚ ❛♥❞ ❍✐r♦♠✐ts✉✳❙✭✶✾✽✼✮❪ ❚❛♥✐❣✉❝❤✐✳❙ ❑♦✐③✉♠✐✳❚ ❛♥❞ ❍✐✲ r♦♠✐ts✉✳❙✳ ❚✇♦â⑨➇♠❛ss ♠♦❞❡❧s ♦❢ t❤❡ ✈♦❝❛❧ ❝♦r❞s ❢♦r ♥❛t✉r❛❧ s♦✉♥❞✐♥❣ ✈♦✐❝❡ s②♥t❤❡s✐s✳ ❏♦✉r♥❛❧ ♦❢ t❤❡ ❆❝♦✉st✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✐❝❛✱ ✽✷✿✶✶✼✾✕✶✶✾✷✱ ✶✾✽✼✳ ❬❑✉❜✐♥✳●✭✶✾✾✺✮❪ ❑✉❜✐♥✳●✳ ✶✾✾✺✳

◆♦♥❧✐♥❡❛r s♣❡❡❝❤ ♣r♦❝❡ss✐♥❣✳

❊❧s❡✈✐❡r ❙❝✐❡♥❝❡✱

❬❑✉♠❛r✳❆ ❛♥❞ ▼✉❧❧✐❝❦✳❙✳❑✭✶✾✾✻✮❪ ❑✉♠❛r✳❆ ❛♥❞ ▼✉❧❧✐❝❦✳❙✳❑✳ ◆♦♥❧✐♥❡❛r ❞②✲ ♥❛♠✐❝❛❧ ❛♥❛❧②s✐s ♦❢ s♣❡❡❝❤✳ ❏♦✉r♥❛❧ ♦❢ t❤❡ ❆❝♦✉st✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✲ ✐❝❛✱ ✶✵✵✿✻✶✺✕✻✷✾✱ ✶✾✾✻✳ ❬▲✳ ❘✳ ❇❛❤❧ ❛♥❞ ▼❡r❝❡r✭✶✾✽✻✮❪ P✳ ❱✳ ❞❡❙♦✉③❛ ▲✳ ❘✳ ❇❛❤❧✱ P✳ ❋✳ ❇r♦✇♥ ❛♥❞ ▲✳ ❘✳ ▼❡r❝❡r✳ ▼❛①✐♠✉♠ ♠✉t✉❛❧ ✐♥❢♦r♠❛t✐♦♥❡st✐♠❛t✐♦♥ ♦❢ ❤✐❞❞❡♥ ✷✻✹

♠❛r❦♦✈ ♠♦❞❡❧ ♣❛r❛♠❡t❡rs ❢♦r s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■♥ ✽✻✱❏❛♣❛♥✱ ✶✾✽✻✳

Pr♦❝✳ ■❈❆❙❙P

❬▲✳ ❘✳ ❘❛❜✐♥❡r ❛♥❞ ❲✐❧♣♦♥✭✶✾✼✾✮❪ ❆ ❊ ❘♦s❡♥❜❡r❣ ▲✳ ❘✳ ❘❛❜✐♥❡r✱ ❙ ❊ ▲❡✈✐♥✲ s♦♥ ❛♥❞ ❏ ● ❲✐❧♣♦♥✳ ❙♣❡❛❦❡r ✐♥❞❡♣❡♥❞❡♥t r❡❝♦❣♥✐t✐♦♥ ♦❢ ✐s♦❧❛t❡❞ ✇♦r❞s ✉s✐♥❣ ❝❧✉st❡r✐♥❣ t❡❝❤♥✐q✉❡s✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❆❝♦✉st✐❝✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣✲ ♥❛❧ Pr♦❝❡ss✐♥❣✱ ✶✾✼✾✳ ❬▲❛❞❡❢♦❣❡❞✭✷✵✵✹✮❪ P❡t❡r ▲❛❞❡❢♦❣❡❞✳

❱♦✇❡❧s ❛♥❞ ❈♦♥s♦♥❛♥ts ✲ ❆♥ ■♥tr♦❞✉❝✲

✳ ❇❧❛❝❦❲❡❧❧ P✉❜❧✐s❤✐♥❣✱ ✷✵✵✹✳

t✐♦♥ t♦ t❤❡ ❙♦✉♥❞s ♦❢ ▲❛♥❣✉❛❣❡

❬▲❛❥✐s❤ ❛♥❞ ◆❛r❛②❛♥❛♥✭✷✵✵✹✮❪ ❘ ❑ ❙✉♥✐❧❦✉♠❛r ❱ ▲ ▲❛❥✐s❤ ❛♥❞ ◆ ❑ ◆❛r❛②❛♥❛♥✳ ❱♦✇❡❧ ♣❤♦♥❡ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ③❡r♦ ❝r♦ss✐♥❣ ❜❛s❡❞ ❞✐s✲ tr✐❜✉t✐♦♥ ♣❛tt❡r♥ ❛♥❞ ❛rt✐✜❝✐❛❧ ♥❡✉r❛❧ ♥❡t✇♦r❦s✳ ■♥ ■♥ Pr♦❝✳ ◆❛t✐♦♥❛❧ ❈♦♥❢❡r❡♥❝❡ ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ■♥t❡❧❧✐❣❡♥t ❙②st❡♠s ❛♥❞ ◆❡t✇♦r❦✐♥❣✱

✱ ♣❛❣❡s ✹✾✕✺✹✱ ✷✵✵✹✳

❙P■◆ ✷✵✵✸

❬▲❛♥❣✐✳❆ ❛♥❞ ❑✐♥s♥❡r✳❲✭✶✾✾✺✮❪ ▲❛♥❣✐✳❆ ❛♥❞ ❑✐♥s♥❡r✳❲✳ ❈♦♥s♦♥❛♥t ❝❤❛r✲ ❛❝t❡r✐③❛t✐♦♥ ✉s✐♥❣ ❝♦rr❡❧❛t✐♦♥ ❢r❛❝t❛❧ ❞✐♠❡♥s✐♦♥ ❢♦r s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■♥ Pr♦❝✳ ♦❢ ■❊❊ ❈♦♥❢✳ ❈♦♠♠✉♥✐❝❛t✐♦♥s✱ P♦✇❡r✱ ❛♥❞ ❈♦♠♣✉t✐♥❣✱ ✶✾✾✺✳ ❬▲❛♥❣♠❛♥♥ ❛♥❞ ❞❡♥ ❖s✭✶✾✾✻✮❪ ❘✳ ❍❛❡❜✲❯♠❜❛❝❤ ▲✳ ❇♦✈❡s ▲❛♥❣♠❛♥♥✱ ❉✳ ❛♥❞ ❊✳ ❞❡♥ ❖s✳ ❋r❡s❝♦✿ ❚❤❡ ❢r❡♥❝❤ t❡❧❡♣❤♦♥❡ s♣❡❡❝❤ ❞❛t❛ ❝♦❧❧❡❝t✐♦♥ ✲ ♣❛rt ♦❢ t❤❡ ❡✉r♦♣❡❛♥ s♣❡❡❝❤❞❛t✭♠✮ ♣r♦❥❡❝t✳ ❢r❡s❝♦✳ ■♥ ❋♦✉rt❤ ■♥t❡r♥❛✲ t✐♦♥❛❧ ❈♦♥❢❡r❡♥❝❡ ♦♥ ❙♣♦❦❡♥ ▲❛♥❣✉❛❣❡ Pr♦❝❡ss✐♥❣✳ P❤✐❧❛❞❡❧♣❤✐❛✱ ♣❛❣❡ ✶✾✶✽✱ ✶✾✾✻✳ ❬▲❡✐ ❛♥❞ ❏✳❩✉♦✭✷✵✵✾✮❪ ❨❛❣✉♦ ▲❡✐ ❛♥❞ ▼✐♥❣ ❏✳❩✉♦✳ ●❡❛r ❝r❛❝❦ ❧❡✈❡❧ ✐❞❡♥t✐✲ ✜❝❛t✐♦♥ ❜❛s❡❞ ♦♥ ✇❡✐❣❤t❡❞ ❦✲♥❡❛r❡st ♥❡✐❣❤❜♦r ❝❧❛ss✐✜❝❛t✐♦♥ ❛❧❣♦r✐t❤♠✳ ▼❡❝❤❛♥✐❝❛❧ ❙②st❡♠s ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✷✸✿✶✺✸✺✕✶✺✹✼✱ ✷✵✵✾✳ ❬▲✐♣♣♠❛♥♥✭✶✾✽✼✮❪ ❘✳ P✳ ▲✐♣♣♠❛♥♥✳ ❆♥ ✐♥tr♦❞✉❝t✐♦♥ t♦ ❝♦♠♣✉t✐♥❣ ✇✐t❤ ♥❡✉✲ r❛❧ ♥❡ts✳ ■❊❊❊ ❆❙❙P ▼❛❣❛③✐♥❡✱ ✺✿✶✶✺✕✶✸✸✱ ✶✾✽✼✳ ❬▲♦❢t❣❛❛r❞❡♥ ❛♥❞ ◗✉❡❡♥s❜r❡rr②✭✶✾✻✺✮❪ ❉✳ ❖ ▲♦❢t❣❛❛r❞❡♥ ❛♥❞ ❈✳ P✳ ◗✉❡❡♥s✲ ❜r❡rr②✳ ❆ ♥♦♥♣❛r❛♠❡tr✐❝ ❡st✐♠❛t❡ ♦❢ ❛ ♠✉❧t✐✈❛r✐❛t❡ ❞❡♥s✐t② ❢✉♥❝t✐♦♥✳ ❆♥♥❛❧s ♦❢ ▼❛t❤❡♠❛t✐❝❛❧ ❙t❛t✐st✐❝s✱ ✸✻✿✶✵✹✾✕✶✵✺✶✱ ✶✾✻✺✳ ❬▼ ▲❛♥❣ ❛♥❞ ❘✳❖✳❲❡❧❧s✭✶✾✾✻✮❪ ❏✳❊✳❖❞❡❣❛r❞ ❈✳❙✳❇✉rr✉s ▼ ▲❛♥❣✱ ❍✳ ●✉♦ ❛♥❞ ❘✳❖✳❲❡❧❧s✳ ◆♦✐s❡ r❡❞✉❝t✐♦♥ ✉s✐♥❣ ❛♥ ✉♥❞❡❝✐♠❛t❡❞ ❞✐s❝r❡t❡ ✇❛✈❡❧❡t tr❛♥s❢♦r♠✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✸✭✶✮✿✶✵✕✶✷✱ ✶✾✾✻✳ ❬▼❛❤✳❘✳❙✳❍ ❛♥❞ ❈❤❛❦r❛✈❛rt❤②✳❱✭✶✾✾✷✮❪ ▼❛❤✳❘✳❙✳❍ ❛♥❞ ❈❤❛❦r❛✈❛rt❤②✳❱✳ P❛tt❡r♥ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ❛rt✐✜❝✐❛❧ ♥❡✉r❛❧ ♥❡t✇♦r❦✳ ❈♦♠♣✉t❡rs ❛♥❞ ❈❤❡♠✐❝❛❧ ❊♥❣✐♥❡❡r✐♥❣✱ ✶✻✭✹✮✱ ✶✾✾✷✳ ✷✻✺

❬▼❛❦❤♦✉❧✭✶✾✼✺✮❪ ❏❤♦♥ ▼❛❦❤♦✉❧✳ ❧✐♥❡❛r ♣r❡❞✐❝t✐♦♥✿ ❆ t✉t♦r✐❛❧ r❡✈✐❡✇✳ ❝❡❡❞✐♥❣s ♦❢ t❤❡ ■❊❊❊✱ ✻✸✭✹✮✿✺✻✶✕✺✽✵✱ ✶✾✼✺✳

Pr♦✲

❬▼✳❆❦s❡❧❛ ❛♥❞ ❏✳▲❛❛❦s♦♥❡♥✭✷✵✵✻✮❪ ▼✳❆❦s❡❧❛ ❛♥❞ ❏✳▲❛❛❦s♦♥❡♥✳ ◆❡✐❣❤❜♦r✲ ❤♦♦❞ s✐③❡ s❡❧❡❝t✐♦♥ ✐♥ t❤❡ ❦✲♥❡❛r❡st ♥❡✐❣❤❜♦r r✉❧❡ ✉s✐♥❣ st❛t✐st✐❝❛❧ ❝♦♥✲ ✜❞❡♥❝❡✳ P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥✱ ✸✾✿✹✶✼✕✹✷✸✱ ✷✵✵✻✳ ❬▼❛❧❧❛t✭✶✾✽✾✮❪ ❙ ▼❛❧❧❛t✳ ❆ t❤❡♦r② ❢♦r ♠✉❧t✐ r❡s♦❧✉t✐♦♥ s✐❣♥❛❧ ❞❡❝♦♠♣♦s✐t✐♦♥ ✿ ❚❤❡ ✇❛✈❡❧❡t r❡♣r❡s❡♥t❛t✐♦♥✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ P❛tt❡r♥ ❆♥❛❧②s✐s ❛♥❞ ♠❛❝❤✐♥❡ ■♥t❡❧❧✐❣❡♥❝❡✱ ✶✶✭✶✮✿✻✼✹✕✻✾✸✱ ✶✾✽✾✳ ❬▼❛❧❧❛t✭✷✵✵✾✮❪ ❙✳ ▼❛❧❧❛t✳ ❆ ✇❛✈❡❧❡t ❚♦✉r ❲❛②✳ ◆❡✇❨♦r❦ ✿ ❆❝❛❞❡♠✐❝✱ ✷✵✵✾✳

♦❢ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ❚❤❡ ❙♣❛rs❡

❬▼❛r❦❡❧ ❛♥❞ ❆✳❍✳●r❛②✭✶✾✼✹✮❪ ❉✳ ▼❛r❦❡❧ ❛♥❞ ❆✳❍✳●r❛②✳ ❆ ❧✐♥❡❛r ♣r❡❞✐❝t✐♦♥ ✈♦❝♦❞❡r s✐♠✉❧❛t✐♦♥ ❜❛s❡❞ ✉♣♦♥ t❤❡ ❛✉t♦❝♦rr❡❧❛t✐♦♥ ♠❡t❤♦❞✳ ■❊❊❊ ❚r❛♥s✳ ❛❝♦✉st✳ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ❆❙❙P✕✷✷✭✶✮✱ ✶✾✼✹✳ ❬▼❝❈✉❧❧♦✉❣❤ ❛♥❞ P✐tts✭✶✾✹✸✮❪ ❲ ❈ ▼❝❈✉❧❧♦✉❣❤ ❛♥❞ ❲ ❍ P✐tts✳ ❆ ❧♦❣✐❝❛❧ ❝❛❧❝✉❧✉s ♦❢ ✐❞❡❛s ✐♠♠❛♥❡♥t ✐♥ ♥❡r✈♦✉s ❛❝t✐✈✐t②✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❆❝♦✉st✐❝ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣ ▼❛❣❛③✐♥❡✱ ✻✶✿✹✕✷✷✱ ✶✾✹✸✳ ❬▼✐❝❤❛❡❧ ❚ ❏❤♦♥s♦♥ ❛♥❞ ■♥❞r❡❜♦✭✷✵✵✺✮❪ ❆♥❞r❡✇ ❈ ▲✐♥❞❣r❡♥ ❏✐♥❥✐♥ ❨❡ ❳✐✲ ❛♦❧✐♥ ▲✐✉ ▼✐❝❤❛❡❧ ❚ ❏❤♦♥s♦♥✱ ❘❝❤❛r❞ ❏ P♦✈✐♥❛❧❧✐ ❛♥❞ ❑❡✈✐♥ ■♥❞r❡❜♦✳ ❚✐♠❡ ❞♦♠❛✐♥ ✐s♦❧❛t❡❞ ♣❤♦♥❡♠❡ ❝❧❛ss✐✜❝❛t✐♦♥ ✉s✐♥❣ r❡❝♦♥str✉❝t❡❞ ♣❤❛s❡ s♣❛❝❡✳ ■❊❊❊ ❚r❛♥s✳ ❖♥ ❙♣❡❡❝❤ ❛♥❞ ❆✉❞✐♦ Pr♦❝❡ss✐♥❣✱ ✶✸✭✹✮✿✹✺✽✕✹✻✻✱ ✷✵✵✺✳ ❬▼✐♥s❦② ❛♥❞ P❛♣❡rt✭✶✾✻✾✮❪ ▼✳ ▼✐♥s❦② ❛♥❞ ❙✳ P❛♣❡rt✳ P❡r❝❡♣tr♦♥s✿ tr♦❞✉❝t✐♦♥ t♦ ❈♦♠♣✉t❛t✐♦♥❛❧ ●❡♦♠❡tr②✳ ▼■❚ Pr❡ss✱ ✶✾✻✾✳

❆♥ ■♥✲

❬▼✉t❤✉s❛♠② ❛♥❞ ●♦❞❢r❡②✭✶✾✾✺✮❪ ❊✳ ❍♦❧❧✐♠❛♥ ❇✳ ❲❤❡❛t❧❡② ❏✳ P✐❝♦♥❡ ▼✉t❤✉s❛♠②✱ ❨✳ ❛♥❞ ❏✳ ●♦❞❢r❡②✳ ❱♦✐❝❡ ❛❝r♦ss ❤✐s♣❛♥✐❝ ❛♠❡r✐❝❛✿ ❆ t❡❧❡♣❤♦♥❡ s♣❡❡❝❤ ❝♦r♣✉s ♦❢ ❛♠❡r✐❝❛♥ s♣❛♥✐s❤✳ ❛❝♦✉st✐❝s✳ ■♥ Pr♦❝✳ t❤❡ ■❈❙❙P✱ ♣❛❣❡ ✽✺✱ ✶✾✾✺✳ ❬◆ ❑ ◆❛r❛②❛♥❛♥ ❛♥❞ ❙❛s✐♥❞r❛♥✭✷✵✵✵✮❪ P Pr❛❥✐t❤ ◆ ❑ ◆❛r❛②❛♥❛♥ ❛♥❞ ❊ ❙❛s✐♥❞r❛♥✳ ❆♣♣❧✐❝❛t✐♦♥ ♦❢ ♣❤❛s❡ s♣❛❝❡ ♠❛♣ ❛♥❞ ♣❤❛s❡ s♣❛❝❡ ♣♦✐♥t ❞✐str✐❜✉t✐♦♥ ✐♥ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■♥ ■♥ Pr♦❝✳✱ ■❊❊❊✱ ■♥t✳ ❈♦♥❢✳ ❖♥ ❈♦♠♠✉♥✐❝❛t✐♦♥✱ ❝♦♥tr♦❧ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ■■❙❝✱ ❇❛♥❣❛❧♦r❡✱ ♣❛❣❡s ✾✽✕✶✵✶✱ ✷✵✵✵✳ ❬◆❛r❛②❛♥❛♥✭✶✾✾✾✮❪ ◆ ❑ ◆❛r❛②❛♥❛♥✳ ❱♦✐❝❡❞ ✴ ✉♥✈♦✐❝❡❞ ❝❧❛ss✐✜❝❛t✐♦♥ ✉s✐♥❣ s❡❝♦♥❞ ♦r❞❡r ❛ttr❛❝t♦r ❞✐♠❡♥s✐♦♥ ❛♥❞ s❡❝♦♥❞ ♦r❞❡r ❦♦❧♠♦❣r♦✈ ❡♥tr♦♣② ♦❢ s♣❡❡❝❤ s✐❣♥❛❧s✳ ❏✳❆❝♦✉s✳❙♦❝✳■♥❞✱ ✷✼✿✶✽✶✕✶✽✺✱ ✶✾✾✾✳ ✷✻✻

❬◆❛r❛②❛♥❛♥✳◆✳❑ ❛♥❞ ❙r✐❞❤❛r✳❈✳❙✭✶✾✽✽✮❪ ◆❛r❛②❛♥❛♥✳◆✳❑ ❛♥❞ ❙r✐❞❤❛r✳❈✳❙✳ P❛r❛♠❡tr✐❝ r❡♣r❡s❡♥t❛t✐♦♥ ♦❢ t❤❡ ❞②♥❛♠✐❝❛❧ ✐♥st❛❜✐❧✐t✐❡s ❛♥❞ ❞❡t❡r♠✐♥✲ ✐st✐❝ ❝❤❛♦s ✐♥ s♣❡❡❝❤ s✐❣♥❛❧s✳ ■♥ Pr♦❝✳ ❙②♠♣♦s✐✉♠ ♦♥ ❙✐❣♥❛❧s✱ ❙②st❡♠s ❛♥❞ ❙♦♥❛rs✱ ◆P❖▲✱ ❈♦❝❤✐♥✱ ✶✾✽✽✳ ❬◆❛r❛②❛♥❛♥✳❙✳❙ ❛♥❞ ❆❧✇❛♥✳❆✳❆✭✶✾✾✺✮❪ ◆❛r❛②❛♥❛♥✳❙✳❙ ❛♥❞ ❆❧✇❛♥✳❆✳❆✳ ❆ ♥♦♥❧✐♥❡❛r ❞②♥❛♠✐❝❛❧ s②st❡♠s ❛♥❛❧②s✐s ♦❢ ❢r✐❝❛t✐✈❡ ❝♦♥s♦♥❛♥ts✳ ❏♦✉r♥❛❧ ♦❢ t❤❡ ❆❝♦✉st✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✐❝❛✱ ✾✼✿✷✺✶✶✕✷✺✷✹✱ ✶✾✾✺✳ ❬❖ ❋❛r♦♦q ❛♥❞ ❙❤r♦tr✐②❛✭✷✵✶✵✮❪ ❙ ❉✉tt❛ ❖ ❋❛r♦♦q ❛♥❞ ▼ ❈ ❙❤r♦tr✐②❛✳ ❲❛✈❡❧❡t s✉❜ ❜❛♥❞ ❜❛s❡❞ t❡♠♣♦r❛❧ ❢❡❛t✉r❡s ❢♦r r♦❜✉st ❤✐♥❞✐ ♣❤♦♥❡♠❡ r❡❝♦❣♥✐t✐♦♥✳ ■♥t✳ ❏♦✉r♥❛❧ ♦❢ ❲❛✈❡❧❡ts✱ ▼✉❧t✐ r❡s♦❧✉t✐♦♥ ❛♥❞ ■♥❢♦r♠❛t✐♦♥ Pr♦❝❡ss✐♥❣✱ ✽✭✻✮✿✽✹✼✕✽✺✾✱ ✷✵✶✵✳ ❬❖✳❍❡❜❜✭✶✾✹✾✮❪ ❉✳ ❖✳❍❡❜❜✳ ❚❤❡ ♦r❣❛♥✐③❛t✐♦♥ ❧♦❣✐❝❛❧ ❚❤❡♦r②✳ ❲✐❧❡② ✲ ◆❡✇❨♦r❦✱ ✶✾✹✾✳ ❬❖tt✭✶✾✾✸✮❪ ❊✳ ❖tt✳ Pr❡ss✱ ✶✾✾✸✳

♦❢ ❇❡❤❛✈✐♦✉r✱ ❆ ◆❡✉r♦♣s②❝❤♦✲

✳ ❈❛♠❜r✐❞❣❡ ❯♥✐✈❡rs✐t②

❈❤❛♦s ✐♥ ❉②♥❛♠✐❝❛❧ ❙②st❡♠s

❬❖tt✳❊ ❛♥❞ ❙❛✉❡r✳❚✭✶✾✾✹✮❪ ❖tt✳❊ ❛♥❞ ❙❛✉❡r✳❚✳ ◆❡✇ ❨♦r❦✱ ✶✾✾✹✳

✳ ❲✐❧❡②✱

❈♦♣✐♥❣ ✇✐t❤ ❈❤❛♦s

❬P Pr❛❥✐t❤ ❛♥❞ ◆❛r❛②❛♥❛♥✭✷✵✵✹✮❪ ◆ ❙ ❙r❡❡❦❛♥t❤ P Pr❛❥✐t❤ ❛♥❞ ◆ ❑ ◆❛r❛②❛♥❛♥✳ ♣❤❛s❡ s♣❛❝❡ ♣❛r❛♠❡t❡r ❢♦r ♥❡✉r❛❧ ♥❡t✇♦r❦ ❜❛s❡❞ ✈♦✇❡❧ r❡❝♦❣♥✐t✐♦♥✳ ■♥ Pr♦❝✳ ■♥tr♥❛t✐♦♥❛❧ ❈♦♥❢❡r❡♥❝❡ ♦♥ ◆❡✉r❛❧ ■♥❢♦r♠❛t✐♦♥ Pr♦❝❡ss✐♥❣✱■◆❉■❆✱ ♣❛❣❡s ✶✷✵✹✕✶✷✵✾✱ ✷✵✵✹✳ ❬P❛❝❦❛r❞✳◆✳❍ ❛♥❞ ❙❤❛✇✳❘✳❙✭✶✾✽✵✮❪ ❋❛r♠❡r✳❏✳❉ P❛❝❦❛r❞✳◆✳❍✱ ✜❡❧❞✳❏✳P ❛♥❞ ❙❤❛✇✳❘✳❙✳ ●❡♦♠❡tr② ❢r♦♠ ❛ t✐♠❡ s❡r✐❡s✳ P❤②s✳ ✹✺✿✼✶✷✕✼✶✼✱ ✶✾✽✵✳

❈r✉t❝❤✲ ❘❡✈✳ ▲❡tt✱

❬P❛❧ ❛♥❞ ▼✐tr❛✭✶✾✽✽✮❪ ❙❛♥❦❛r✳ ❑✳ P❛❧ ❛♥❞ ❙✉s❤♠✐t❛ ▼✐tr❛✳ ▼✉❧t✐❧❛②❡r ♣❡r✲ ❝❡♣tr♦♥✱ ❢✉③③② s❡ts✱ ❛♥❞ ❝❧❛ss✐✜❝❛t✐♦♥✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ◆❡✉r❛❧ ◆❡t✇♦r❦s✱ ✸✭✺✮✿✻✽✸✕✻✾✼✱ ✶✾✽✽✳ ❬P❛t✐❧ ❛♥❞ ❇❛s✉✭✷✵✵✽✮❪ ❍❡♠❛♥t❤ ❆ P❛t✐❧ ❛♥❞ ❚ ❑ ❇❛s✉✳ ▲♣ s♣❡❝tr❛ ✈s✳ ♠❡❧ s♣❡❝tr❛ ❢♦r ✐❞❡♥t✐✜❝❛t✐♦♥ ♦❢ ♣r♦❢❡ss✐♦♥❛❧ ♠✐♠✐❝s ✐♥ ✐♥❞✐❛♥ ❧❛♥❣✉❛❣❡s✳ ■♥t❡r♥❛t✐♦♥❛❧ ❏♦✉r♥❛❧ ♦❢ ❙♣❡❡❝❤ ❚❡❝❤♥♦❧♦❣②✱ ✶✶✿✶✕✶✻✱ ✷✵✵✽✳ ❬P❡r♥❦♦♣❢✭✷✵✵✺✮❪ ❋✳ P❡r♥❦♦♣❢✳ ❇❛②❡s✐❛♥ ♥❡t✇♦r❦ ❝❧❛ss✐✜❡rs ✈❡rs✉s s❡❧❡❝t✐✈❡ ❦ â⑨➇♥♥ ❝❧❛ss✐✜❡r✳ P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥✱ ✸✽✿✶✕✶✵✱ ✷✵✵✺✳

✷✻✼

❬P❡rs♦♥♥❛③✳▲ ❛♥❞ ❉r❡②❢✉s✳●✭✶✾✾✵✮❪ P❡rs♦♥♥❛③✳▲ ❛♥❞ ❉r❡②❢✉s✳●✳ ◆❡✉r❛❧ ♥❡t✲ ✇♦r❦s✿ st❛t❡ ♦❢ t❤❡ ❛rt ❛♥❞ ❢✉t✉r❡ ♣r♦s♣❡❝ts✳ ❘❡✈✉❡✲●❡♥❡r❛❧❡✲❞❡✲ ❧✬❊❧❡❝tr✐❝✐t❡✱ ✺✱ ✶✾✾✵✳ ❬P✐ts✐❦❛❧✐s✳❱ ❛♥❞ ▼❛r❛❣♦s✳P✭✷✵✵✸✮❪ P✐ts✐❦❛❧✐s✳❱ ❛♥❞ ▼❛r❛❣♦s✳P✳ ❙♦♠❡ ❛❞✲ ✈❛♥❝❡s ♦♥ s♣❡❡❝❤ ❛♥❛❧②s✐s ✉s✐♥❣ ❝❤❛♦t✐❝ ♠♦❞❡❧s✳ ■♥ Pr♦❝✳ ■❙❈❆ ❚✉t♦r✐❛❧ ❛♥❞ ❘❡s❡❛r❝❤ ❲♦r❦s❤♦♣ ♦♥ ◆♦♥❧✐♥❡❛r ❙♣❡❡❝❤ Pr♦❝❡ss✐♥❣ ✭◆❖▲■❙P✮✱ ✷✵✵✸✳ ❬P♦✈✐♥❡❧❧✐✳❘✳❏✭✷✵✵✻✮❪ ❆♥❞r❡✇ ❈✳ ▲✐♥❞❣r❡♥ ❋❡❧✐❝❡ ▼✳ ❘♦❜❡rts ❏✐♥❥✐♥ ❨❡ P♦✈✐♥❡❧❧✐✳❘✳❏✱ ▼✐❝❤❛❡❧ ❚✳ ❏♦❤♥s♦♥✳ ❙t❛t✐st✐❝❛❧ ♠♦❞❡❧s ♦❢ r❡❝♦♥str✉❝t❡❞ ♣❤❛s❡ s♣❛❝❡s ❢♦r s✐❣♥❛❧ ❝❧❛ss✐✜❝❛t✐♦♥✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✺✹✭✻✮✿✷✶✼✽✕✷✶✽✻✱ ✷✵✵✻✳ ❬Pr❛❥✐t❤✭✷✵✵✽✮❪ P✳ Pr❛❥✐t❤✳ ■♥✈❡st✐❣❛t✐♦♥s ♦♥ t❤❡ ❆♣♣❧✐❝❛t✐♦♥s ♦❢ ❉②♥❛♠✐✲ ❝❛❧ ■♥st❛❜✐❧✐t✐❡s ❛♥❞ ❉❡t❡r♠✐♥✐st✐❝ ❈❤❛♦s ❢♦r ❙♣❡❡❝❤ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✳ P❤❉ t❤❡s✐s✱ ❯♥✐✈❡rs✐t② ♦❢ ❈❛❧✐❝✉t✱ ✷✵✵✽✳ ❬Pr❛❥✐t❤ ❛♥❞ ◆❛r❛②❛♥❛♥✭✷✵✵✻✮❪ P Pr❛❥✐t❤ ❛♥❞ ◆ ❑ ◆❛r❛②❛♥❛♥✳ ◆♦♥❧✐♥❡❛r ♣❤❛s❡ s♣❛❝❡ ❢❡❛t✉r❡s ❢♦r ♣✐t❝❤ ❞❡t❡❝t✐♦♥✳ ■♥ ■♥ Pr♦❝✳ ◆❛t✐♦♥❛❧ ❙②♠♣♦✲ s✐✉♠ ♦♥ ❆❝♦✉st✐❝s ✭◆❙❆✮✱ ◆P▲✱ ◆❡✇ ❉❡❧❤✐❡✱ ✷✵✵✻✳ ❬❘ ❑r♦♥❧❛♥❞✲▼❛rt✐♥❡t ❛♥❞ ●r♦ss♠❛♥✭✶✾✽✼✮❪ ❏ ▼♦r❧❡t ❘ ❑r♦♥❧❛♥❞✲▼❛rt✐♥❡t ❛♥❞ ❆ ●r♦ss♠❛♥✳ ❆♥❛❧②s✐s ♦❢ s♦✉♥❞ ♣❛tt❡r♥s t❤r♦✉❣❤ ✇❛✈❡❧❡t tr❛♥s✲ ❢♦r♠s✳ ■♥t✳ ❏♦✉r♥❛❧ ♦❢ P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥✱ ❆rt✐✜❝✐❛❧ ■♥t❡❧❧✐❣❡♥❝❡✱ ✶✭✷✮✿ ✷✼✸✕✸✵✷✱ ✶✾✽✼✳ ❬❘ ❙❛r✐❦❛②❛ ❛♥❞ ❍❛♥s❡♥✭✶✾✾✽✮❪ ❇ ▲ P❡❧❧♦♠ ❘ ❙❛r✐❦❛②❛ ❛♥❞ ❏ ❍ ▲ ❍❛♥s❡♥✳ ❲❛✈❡❧❡t ♣❛❝❦❡t tr❛♥s❢♦r♠ ❢❡❛t✉r❡s ✇✐t❤ ❛♣♣❧✐❝❛t✐♦♥ t♦ s♣❡❛❦❡r ✐❞❡♥t✐✲ ✜❝❛t✐♦♥✳ ■♥ ■♥ Pr♦❝ ♦❢ t❤❡ ✸r❞ ■❊❊❊ ◆♦r❞✐❝ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣ ❙②♠♣♦✲ s✐✉♠✱ ♣❛❣❡s ✽✶✕✽✹✱ ✶✾✾✽✳ ❬❘❛❜✐♥❡r ❛♥❞ ❏✉❛♥❣✭✶✾✾✸✮❪ ▲ ❘❛❜✐♥❡r ❛♥❞ ❇ ❍ ❏✉❛♥❣✳ ❋✉♥❞❛♠❡♥t❛❧s ♦❢ ❙♣❡❡❝❤ ❘❡❝♦❣♥✐t✐♦♥✳ P❡❛rs♦♥ ❊❞✉❝❛t✐♦♥✱ ✶✾✾✸✳ ❬❘❛❜✐♥❡r ❛♥❞ ❏✉❛♥❣✭✷✵✵✸✮❪ ▲ ❘❛❜✐♥❡r ❛♥❞ ❇ ❍ ❏✉❛♥❣✳ ❋✉♥❞❛♠❡♥t❛❧s ♦❢ ❙♣❡❡❝❤ ❘❡❝♦❣♥✐t✐♦♥✳ P❡❛rs♦♥ ❊❞✉❝❛t✐♦♥✱ ✷✵✵✸✳ ❬❘❛❜✐♥❡r ❛♥❞ ❙❝❤❛❢❡r✭✶✾✼✽✮❪ ▲✳ ❘✳ ❘❛❜✐♥❡r ❛♥❞ ❘✳ ❲✳ ❙❝❤❛❢❡r✳ ❉✐❣✐t❛❧ Pr♦✲ ❝❡ss✐♥❣ ♦❢ ❙♣❡❡❝❤ ❙✐❣♥❛❧s✳ Pr❡♥t✐❝❡✲❍❛❧❧✱ ❊❛❣❧❡✇♦♦❞ ❈❧✐✛s✱ ✶✾✼✽✳ ❬❘❛❥❛ ❛♥❞ ❉❛♥❞❛♣t✭✷✵✶✵✮❪ ●✳ ❙❡♥t❤✐❧ ❘❛❥❛ ❛♥❞ ❙✳ ❉❛♥❞❛♣t✳ ❙♣❡❛❦❡r r❡❝♦❣✲ ♥✐t✐♦♥ ✉♥❞❡r str❡ss❡❞ ❝♦♥❞✐t✐♦♥✳ ■♥t❡r♥❛t✐♦♥❛❧ ❏♦✉r♥❛❧ ♦❢ ❙♣❡❡❝❤ ❚❡❝❤✲ ♥♦❧♦❣②✱ ✶✸✿✶✹✶✕✶✻✶✱ ✷✵✶✵✳ ✷✻✽

❬❘❛② ❛♥❞ ❇✭✶✾✽✹✮❪ ❆✳ ❑✳ ❘❛② ❛♥❞ ❈❤❛tt❡r❥❡❡ ❇✳ ❉❡s✐❣♥ ♦❢ ❛ ♥❡❛r❡st ♥❡✐❣❤❜♦r ❝❧❛ss✐✜❡r s②st❡♠ ❢♦r ❜❡♥❣❛❧✐ ❝❤❛r❛❝t❡r r❡❝♦❣♥✐t✐♦♥✳ ❏♦✉r♥❛❧ ♦❢ ■♥st✳ ❊❧❡❝✳ ❚❡❧❡❝♦♠✳ ❊♥❣✱ ✸✵✿✷✷✻✕✷✷✾✱ ✶✾✽✹✳ ❬❘❡❞❞②✭✶✾✻✻✮❪ ❉✳ ❘✳ ❘❡❞❞②✳ ❆♥ ❛♣♣r♦❛❝❤ t♦ ❝♦♠♣✉t❡r s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ❜② ❞✐r❡❝t ❛♥❛❧②s✐s ♦❢ t❤❡ s♣❡❡❝❤ ✇❛✈❡✳ ❚❡❝❤♥✐❝❛❧ r❡♣♦rt✱ ❈♦♠♣✉t❡r ❙❝✐❡♥❝❡ ❉❡♣t✳✱❙t❛♥❢♦r❞ ❯♥✐✈❡r✐st②✱ ✶✾✻✻✳ ❬❘✐❝❤❛r❞ ❖✳ ❉✉❞❛ ❛♥❞ ●✳❙t♦r❦✭✷✵✵✺✮❪ P❡t❡r ❊✳❍❛rt ❘✐❝❤❛r❞ ❖✳ ❉✉❞❛ ❛♥❞ ❉❛✈✐❞ ●✳❙t♦r❦✳ P❛tt❡r♥ ❈❧❛ss✐✜❝❛t✐♦♥✳ ❏♦❤♥ ❲✐❧❡② ❛♥❞ ❙♦♥s✱ ✷✵✵✺✳ ❬❘✐♣❧❡②✭✶✾✾✻✮❪ ❇✳ ❉✳ ❘✐♣❧❡②✳ P❛tt❡r♥ ❘❡❝♦❣♥✐t✐♦♥ ❛♥❞ ◆❡✉r❛❧ ◆❡t✇♦r❦s✳ ❈❛♠✲ ❜r✐❞❣❡ ❯♥✐✈❡rs✐t② Pr❡ss✱ ✶✾✾✻✳ ❬❘♦❜✐♥s♦♥ ❛♥❞ ❘❡♥❛❧s✭✶✾✾✺✮❪ ❏✳ ❋r❛♥s❡♥ ❉✳ P②❡ ❏✳ ❋♦♦t❡ ❘♦❜✐♥s♦♥✱ ❚✳ ❛♥❞ ❙✳ ❘❡♥❛❧s✳ ❲s❥❝❛♠♦✿ ❆ ❜r✐t✐s❤ ❡♥❣❧✐s❤ s♣❡❡❝❤ ❝♦r♣✉s ❢♦r ❧❛r❣❡ ✈♦❝❛❜✉✲ ❧❛r② ❝♦♥t✐♥✉♦✉s s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■♥ Pr♦❝✳ ■❈❆❙❙P ✾✺✱❏❛♣❛♥✱ ♣❛❣❡s ✽✶✕✽✹✱ ✶✾✾✺✳ ❬❘♦♥❛❧❞ ❲ ▲✐♥❞s❛② ❛♥❞ ❘♦t❤r♦❝❦✭✶✾✾✻✮❪ ❉♦♥❛❧❞ ❇ P❡r❝✐✈❛❧ ❘♦♥❛❧❞ ❲ ▲✐♥❞s❛② ❛♥❞ ❉ ❆♥❞r❡✇ ❘♦t❤r♦❝❦✳ ❚❤❡ ❞✐s❝r❡t❡ ✇❛✈❡❧❡t tr❛♥s❢♦r♠ ❛♥❞ t❤❡ s❝❛❧❡ ❛♥❛❧②s✐s ♦❢ t❤❡ s✉r❢❛❝❡ ♣r♦♣❡rt✐❡s ♦❢ s❡❛ ✐❝❡✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ●❡♦ ❙❝✐❡♥❝❡ ❛♥❞ ❘❡♠♦t❡ ❙❡♥s✐♥❣✱ ✸✹✭✸✮✿✼✼✶✕✼✽✼✱ ✶✾✾✻✳ ❬❘♦s❡♥❜❧❛tt✭✶✾✺✾✮❪ ❋✳ ❘♦s❡♥❜❧❛tt✳ ❚✇♦ t❤❡♦r❡♠s ♦❢ st❛t✐st✐❝❛❧ s❡♣❛r❛❜✐❧✐t② ✐♥ t❤❡ ♣❡r❝❡♣t✐♦♥✳ ■♥ ❙②♠♣♦s✐✉♠ ♦♥ t❤❡ ▼❡❝❤❛♥✐③❛t✐♦♥ ♦❢ ❚❤♦✉❣❤t Pr♦❝❡ss✱ ♣❛❣❡s ✹✷✶✕✹✺✻✱ ✶✾✺✾✳ ❬❘✉❜❡♥ ❙♦❧❡r❛✲✉r❡♥❛ ❛♥❞ ❞❡ ▼❛r✐❆✭✷✵✶✷✮❪ ❈❆r♠❡♥ P❡❧❛❡③✲▼♦r❡♥♦ ❘✉❜❡♥ ❙♦❧❡r❛✲✉r❡♥❛✱ ❆♥❛ ■❝❛❜❡❧ ●r❛❝✐❛✲▼♦r❡❧ ❛♥❞ ❋❡t♥❛♥❞♦ ❉✐❛③ ❞❡ ▼❛r✐❆✳ ❘❡❛❧✲t✐♠❡ r♦❜✉st ❛✉t♦♠❛t✐❝ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ❝♦♠♣❛❝t s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡s✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❆✉❞✐♦✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✷✵✭✹✮✿✶✸✹✼✕✶✸✻✶✱ ✷✵✶✷✳ ❬❘✉♠❡❧❍❛rt✳ ❉✳ ❊✳ ❛♥❞ ❏✳✭✶✾✽✻✮❪ ❍✐♥t♦♥✳ ●✳ ❊✳ ❘✉♠❡❧❍❛rt✳ ❉✳ ❊✳ ❛♥❞ ❲✐❧❧✐❛♠s✳ ❘✳ ❏✳ ▲❡❛r♥✐♥❣ ■♥t❡r♥❛❧ ❘❡♣r❡s❡♥t❛t✐♦♥s ❜② ❊rr♦r Pr♦♣❛✲ ❣❛t✐♦♥✳ ▼■❚ Pr❡ss✱ ✶✾✽✻✳ ❬❙ ❛♥❞ ❋✭✶✾✾✷✮❪ ❑❛❞❛♠❜❡ ❙ ❛♥❞ ❇♦✉❞r❡❛✉① ❇❛rt❡❧s ● ❋✳ ❆♣♣❧✐❝❛t✐♦♥ ♦❢ t❤❡ ✇❛✈❡❧❡t tr❛♥s❢♦r♠ ❢♦r ♣✐t❝❤ ❞❡t❡❝t✐♦♥ ♦❢ s♣❡❡❝❤ s✐❣♥❛❧✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ■♥❢♦r♠❛t✐♦♥ ❚❤❡♦r②✱ ✸✽✿✾✶✼✕✾✷✹✱ ✶✾✾✷✳

❙✐♥✲ ❣❧❡ ▲❛②❡r ▲❡❛r♥✐♥❣ ❘❡✈✐s✐t❡❞✿ ❆ ❙t❡♣✇✐s❡ Pr♦❝❡❞✉r❡ ❢♦r ❇✉✐❧❞✐♥❣ ❛♥❞

❬❙ ❑♥❡rr ❛♥❞ ❉r❡②❢✉s✭✶✾✾✵✮❪ ▲ P❡rs♦♥♥❛③ ❙ ❑♥❡rr ❛♥❞ ● ❉r❡②❢✉s✳

✷✻✾

❚r❛✐♥✐♥❣ ❛ ◆❡✉r❛❧ ◆❡t✇♦r❦✱◆❡✉r♦❝♦♠♣✉t✐♥❣✿

❆❧❣♦r✐t❤♠s✱❆r❝❤✐t❡❝t✉❡s

✳ ❙♣r✐♥❣❡r✱ ✶✾✾✵✳

❛♥❞ ❆♣♣❧✐❝❛t✐♦♥s

❬❙❛❦❛✐ ❛♥❞ ❉♦s❤✐t❛✭✶✾✻✷✮❪ ❏✳ ❙❛❦❛✐ ❛♥❞ ❙✳ ❉♦s❤✐t❛✳ ❚❤❡ ♣❤♦♥❡t✐❝ t②♣❡✇r✐t❡r✳ Pr♦❝✳ ■❋■P ❈♦♥❣r❡ss✱ ♣❛❣❡s ✶✾✸✕✷✶✷✱ ✶✾✻✷✳ ❬❙❛❦❛✐ ❛♥❞ ❉♦s❤✐t❛✭✶✾✻✸✮❪ ❚✳ ❙❛❦❛✐ ❛♥❞ ❙✳ ❉♦s❤✐t❛✳ ❚❤❡ ❛✉t♦♠❛t✐❝ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ s②st❡♠ ❢♦r ❝♦♥✈❡rs❛t✐♦♥❛❧ s♦✉♥❞s✳ ■❊❊❊ ❚r❛♥s✳ ❊❧❡❝tr♦♥✳ ❈❛♠♣✉t✱ ❊❈✲✶✷✿✽✸✺✕✽✹✻✱ ✶✾✻✸✳ ❬❙❛❦♦❡ ❛♥❞ ❈❤✐❜❛✭✶✾✼✽✮❪ ❍✳ ❙❛❦♦❡ ❛♥❞ ❙✳ ❈❤✐❜❛✳ ❉②♥❛♠✐❝ ♣r♦❣r❛♠♠✐♥❣ ❛❧✲ ❣♦r✐t❤♠ q✉❛♥t✐③❛t✐♦♥ ❢♦r s♣♦❦❡♥ ✇♦r❞ r❡❝♦❣♥✐t✐♦♥✳ ■❊❊❊ ❚r❛♥s✳ ❆❝♦✉s✲ t✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝✳✱ ❆❙❙P✲✷✻✭✶✮✿✹✸✕✹✾✱ ✶✾✼✽✳ ❬❙❛♠❜✉r✭✶✾✼✺❛✮❪ ▼✳ ❘✳ ❙❛♠❜✉r✳ ❆♥ ❡✣❝✐❡♥t ❧✐♥❡❛r ♣r❡❞✐❝t✐♦♥ ✈♦❝♦❞❡r✳ ❇❡❧❧ ❙②st❡♠ ❚❡❝❤✳ ❏♦✉r♥❛❧✱ ✺✹✭✶✵✮✱ ✶✾✼✺❛✳

❚❤❡

❬❙❛♠❜✉r✭✶✾✼✺❜✮❪ ▼✳ ❘✳ ❙❛♠❜✉r✳ ❙❡❧❡❝t✐♦♥ ♦❢ ❛❝♦✉st✐❝ ❢❡❛t✉r❡s ❢♦r s♣❡❛❦❡r ✐❞❡♥t✐✜❝❛t✐♦♥✳ ■❊❊❊ ❚r❛♥s✳ ❆❈♦✉st✳ ❙♣❡❡❝❤ ❛♥❞ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ❆❙❙P✲✷✸✿✶✼✻✕✶✽✷✱ ✶✾✼✺❜✳ ❬❙❛♠❜✉r ❛♥❞ ❏❛②❛♥t✭✶✾✼✻✮❪ ▼✳ ❘✳ ❙❛♠❜✉r ❛♥❞ ◆ ❙ ❏❛②❛♥t✳ ❙♣❡❡❝❤ ❡♥❝r✐♣✲ t✐♦♥ ❜② ♠❛♥✐♣✉❧❛t✐♦♥s ♦❢ ❧♣❝ ♣❛r❛♠❡t❡rs✳ ❚❤❡ ❇❡❧❧ ❙②st❡♠ ❚❡❝❤✳ ❏♦✉r✲ ♥❛❧✱ ✺✺✭✾✮✱ ✶✾✼✻✳ ❬❙❛♥❞❤②❛ ❆r♦r❛ ❛♥❞ ❇❛s✉✭✷✵✶✵✮❪ ▼✐t❛ ◆❛s✐♣✉r✐✲▲✳ ▼❛❧✐❦ ▼✳ ❑✉♥❞✉ ❙❛♥❞✲ ❤②❛ ❆r♦r❛✱ ❉❡❜♦t♦s❤ ❇❤❛tt❛❝❤❛r❥❡❡ ❛♥❞ ❉✳ ❑✳ ❇❛s✉✳ P❡r❢♦r♠❛♥❝❡ ❝♦♠♣❛r✐s♦♥ ♦❢ s✈♠ ❛♥❞ ❛♥♥ ❢♦r ❤❛♥❞✇r✐tt❡♥ ❞❡✈♥❛❣❛r✐ ❝❤❛r❛❝t❡r r❡❝♦❣✲ ♥✐t✐♦♥✳ ■♥t❡r♥❛t✐♦♥❛❧ ❏♦✉r♥❛❧ ♦❢ ❈♦♠♣✉t❡r ❙❝✐❡♥❝❡ ■ss✉❡s✱ ✼✭✸✮✿✶✽✕✷✻✱ ✷✵✶✵✳ ❬❙❛✉❡r✳❚ ❛♥❞ ❈❛s❞❛❣❧✐✳▼✭✶✾✾✶✮❪ ❨♦r❦❡✳❏✳❆ ❙❛✉❡r✳❚ ❛♥❞ ❈❛s❞❛❣❧✐✳▼✳ ❊♠✲ ❜❡❞♦❧♦❣②✳ ❏♦✉r♥❛❧ ♦❢ ❙t❛t✐st✐❝❛❧ P❤②s✐❝s✱ ✻✺✭✸✮✿✺✼✻✕✻✶✻✱ ✶✾✾✶✳ ❬❙❝❤❛✛❡r ❛♥❞ ❘❛❜✐♥❡r✭✶✾✼✵✮❪ ❘✳ ❲✳ ❙❝❤❛✛❡r ❛♥❞ ▲✳ ❘✳ ❘❛❜✐♥❡r✳ ❙②st❡♠s ❢♦r ❛✉t♦♠❛t✐❝ ❢♦r♠❛♥t ❛♥❛❧②s✐s ♦❢ ✈♦✐❝❡❞ s♣❡❡❝❤✳ ❏✳ ❆❝♦✉st✳ ❙♦❝✳ ❆♠❡r✐❝❛✱ ✹✼✿✻✸✹✕✻✹✽✱ ✶✾✼✵✳ ❬❙❝❤♦❡♥t❣❡♥✳❏✭✶✾✾✵✮❪ ❙❝❤♦❡♥t❣❡♥✳❏✳ ◆♦♥✲❧✐♥❡❛r s✐❣♥❛❧ r❡♣r❡s❡♥t❛t✐♦♥ ❛♥❞ ✐ts ❛♣♣❧✐❝❛t✐♦♥ t♦ t❤❡ ♠♦❞❡❧❧✐♥❣ ♦❢ t❤❡ ❣❧♦tt❛❧ ✇❛✈❡❢♦r♠✳ ❙♣❡❡❝❤ ❈♦♠♠✉✲ ♥✐❝❛t✐♦♥s✱ ✾✿✶✽✾✕✷✵✶✱ ✶✾✾✵✳ ❬❙❝❤✉❧t③✭✷✵✵✷✮❪ ❚✳ ●❧♦❜❛❧♣❤♦♥❡ ❙❝❤✉❧t③✳ ❆ ♠✉❧t✐❧✐♥❣✉❛❧ s♣❡❡❝❤ ❛♥❞ t❡①t ❞❛t❛❜❛s❡✳ ■♥ Pr♦❝✳ ■❈❙▲P ✾✺✱❯❙❆✱ ✷✵✵✷✳ ✷✼✵

❬❙❤❡♥❣✭✷✵✵✵✮❪ ❨✉♥❧♦♥❣ ❙❤❡♥❣✳ ❲❛✈❡❧❡t ❚r❛♥s❢♦r♠✲❚❤❡ ❚r❛♥s❢♦r♠s ❛♥❞ ❆♣✲ ♣❧✐❝❛t✐♦♥ ❍❛♥❞❜♦♦❦✳ ❈❘❈ Pr❡ss ▲▲❈✱ ✷✵✵✵✳ ❬❙❤❡♥s❛✭✶✾✾✷✮❪ ▼ ❏ ❙❤❡♥s❛✳ ❆✣♥❡ ✇❛✈❡❧❡ts✿ ❲❡❞❞✐♥❣ t❤❡ ❛tr♦✉s ❛♥❞ ♠❛❧❧❛t ❛❧❣♦r✐t❤♠s✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✹✵✿✷✹✻✹✕✷✹✽✷✱ ✶✾✾✷✳ ❬❙✐♠♣s♦♥✭✶✾✾✵✮❪ P✳ ❑✳ ❙✐♠♣s♦♥✳ ❆rt✐✜❝✐❛❧ ◆❡✉r❛❧ ❙②st❡♠s✳ P❡r❣❛♠♦♥ Pr❡ss✱ ✶✾✾✵✳ ❬❙♦♠❛♥ ❛♥❞ ❘❛♠❛❝❤❛♥❞r❛♥✭✷✵✵✺✮❪ ❑ P ❙♦♠❛♥ ❛♥❞ ❑ ■ ❘❛♠❛❝❤❛♥❞r❛♥✳ ■♥✲ s✐❣❤t ✐♥t♦ ❲❛✈❡❧❡ts✱ ❢r♦♠ ❚❤❡♦r② t♦ Pr❛❝t✐❝❡✳ Pr❡♥t✐❝❡ ❍❛❧❧✱ ■♥❞✐❛✱ ✷✵✵✺✳ ❬❙t❛♥❡❧② ❙♠✐t❤ ❙t❡✈❡♥s ❛♥❞ ◆❡✇♠❛♥✭✶✾✸✼✮❪ ❏♦❤♥ ❱♦❧❦♠❛♥ ❙t❛♥❡❧② ❙♠✐t❤ ❙t❡✈❡♥s ❛♥❞ ❊❞✇✐♥ ◆❡✇♠❛♥✳ ❆ s❝❛❧❡ ❢♦r t❤❡ ♠❡❛s✉r❡✲ ♠❡♥t ♦❢ t❤❡ ♣s②❝❤♦❧♦❣✐❝❛❧ ♠❛❣♥✐t✉❞❡ ♣✐t❝❤✳ ❏♦✉r♥❛❧ ♦❢ ❆❝♦✉st✐❝❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✐❝❛✱ ✽✭✸✮✿✶✽✺✼✕✶✾✵✱ ✶✾✸✼✳ ❬❙t❡✐♥❡❝❦❡✳■ ❛♥❞ ❍❡r③❡❧✳❍✭✶✾✾✺✮❪ ❙t❡✐♥❡❝❦❡✳■ ❛♥❞ ❍❡r③❡❧✳❍✳ ❇✐❢✉r❝❛t✐♦♥s ✐♥ ❛♥ ❛s②♠♠❡tr✐❝ ✈♦❝❛❧✲❢♦❧❞ ♠♦❞❡❧✳ ❏♦✉r♥❛❧ ♦❢ ❆❝♦✉st❝✐❛❧ ❙♦❝✐❡t② ♦❢ ❆♠❡r✲ ✐❝❛✱ ✾✼✿✶✽✼✹✕✶✽✽✹✱ ✶✾✾✺✳ ❬❙✉♥✐❧❦✉♠❛r✭✷✵✵✷✮❪ ❘✳ ❑✳ ❙✉♥✐❧❦✉♠❛r✳ ❱♦✇❡❧ P❤♦♥❡♠❡ ❘❡❝♦❣♥✐t✐♦♥ ❢r♦♠ ❩❡r♦ ❈r♦ss✐♥❣ ❜❛s❡❞ P❛r❛♠❡t❡rs ✉s✐♥❣ ❆rt✐✜❝✐❛❧ ◆❡✉r❛❧ ◆❡t✇♦r❦s✳ P❤❉ t❤❡s✐s✱ ❯♥✐✈❡rs✐t② ♦❢ ❈❛❧✐❝✉t✱ ✷✵✵✷✳ ❬❙✉♥✐❧❦✉♠❛r ❛♥❞ ▲❛❥✐s❤✭✷✵✶✷✮❪ ❘ ❑ ❙✉♥✐❧❦✉♠❛r ❛♥❞ ❱ ▲ ▲❛❥✐s❤✳ P❤♦♥❡♠❡ r❡❝♦❣♥✐t✐♦♥ ✉s✐♥❣ ③❡r♦❝r♦ss✐♥❣ ✐♥t❡r✈❛❧ ❞✐str✐❜✉t✐♦♥ ♦❢ s♣❡❡❝❤ ♣❛tt❡r♥s ❛♥❞ ❛♥♥✳ ■♥t❡r♥❛t✐♦♥❛❧ ❏♦✉r♥❛❧ ♦❢ ❙♣❡❡❝❤ ❚❡❝❤♥♦❧♦❣②✱ ❖♥❧✐♥❡ ❋✐rst✱ ✷✵✶✷✳ ❬❙✉③✉❦✐ ❛♥❞ ◆❛❦❛t❛✭✶✾✻✶✮❪ ❏✳ ❙✉③✉❦✐ ❛♥❞ ❑✳ ◆❛❦❛t❛✳ ❘❡❝♦❣♥✐t✐♦♥ ♦❢ ❥❛♣❛♥❡s❡ ✈♦✇❡❧sâ⑨➈♣r❡❧✐♠✐♥❛r② t♦ t❤❡ r❡❝♦❣♥✐t✐♦♥ ♦❢ s♣❡❡❝❤✳ ❏✳ ❘❛❞✐♦ ❘❡s✳ ▲❛❜✱ ✸✼✭✽✮✿✶✾✸✕✷✶✷✱ ✶✾✻✶✳ ❬❚✳ ❇✳ ▼❛rt✐♥ ❛♥❞ ❩❛❞❡❧❧✭✶✾✻✹✮❪ ❆✳ ▲✳ ◆❡❧s♦♥ ❚✳ ❇✳ ▼❛rt✐♥ ❛♥❞ ❍✳ ❏✳ ❩❛❞❡❧❧✳ ❙♣❡❡❝❤ ❘❡❝♦❣♥✐t✐♦♥ ❜② ❋❡❛t✉r❡ ❆❜str❛❝t✐♦♥ ❚❡❝❤♥✐q✉❡s✳ ❚❡❝❤✳ ❘❡♣♦rt ❆▲✲❚❉❘✲✻✹✲✶✼✻✱ ❆✐r ❋♦r❝❡ ❆✈✐♦♥✐❝s ▲❛❜✱ ✶✾✻✹✳ ❬❚❛❦❡♥s✭✶✾✽✵✮❪ ❋✳ ❚❛❦❡♥s✳ ❉❡t❡❝t✐♥❣ str❛♥❣❡ ❛ttr❛❝t♦rs ✐♥ t✉r❜✉❧❡♥❝❡✳ ■♥ Pr♦❝✳ ❉②♥❛♠✐❝❛❧ ❙②❛t❡♠s ❛♥❞ ❚✉r❜✉❧❡♥❝❡✱ ❲❛r✇✐❝❦✱ ♣❛❣❡s ✸✻✻✕✸✽✶✱ ✶✾✽✵✳

✷✼✶

❬❚❛♥ ❛♥❞ ❲❛♥❣✭✷✵✵✹✮❪ ❨✐♥❣ ❚❛♥ ❛♥❞ ❏✉♥ ❲❛♥❣✳ ❆ s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡ ✇✐t❤ ❛ ❤②❜r✐❞ ❦❡r♥❡❧ ❛♥❞ ♠✐♥✐♠❛❧ ✈❛♣♥✐❦ ✲ ❝❤❡r✈♦♥❡♥❦✐♥s ❞✐♠❡♥s✐♦♥✳ ■❊❊❊ tr❛♥s✳ ♦♥ ❑♥♦✇❧❡❞❣❡ ❛♥❞ ❉❛t❛ ❊♥❣✐♥❡❡r✐♥❣✱ ✶✵✭✹✮✿✸✽✺✕✸✾✺✱ ✷✵✵✹✳ ❬❚❡❛❣❡r✳❙✳▼✭✶✾✽✾✮❪ ❚❡❛❣❡r✳❍✳▼✳❛♥❞ ❚❡❛❣❡r✳❙✳▼✳ ❊✈✐❞❡♥❝❡ ❢♦r ♥♦♥❧✐♥❡❛r s♦✉♥❞ ♣r♦❞✉❝t✐♦♥ ♠❡❝❤❛♥✐s♠s ✐♥ t❤❡ ✈♦❝❛❧ tr❛❝t✳ ■♥ ❋r❛♥❝❡ ◆❆❚❖ ❆❞✲ ✈❛♥❝❡❞ ❙t✉❞② ■♥st✐t✉t❡ ❙❡r✐❡s ❉✱ ❇♦♥❛s✱ ❡❞✐t♦r✱ ❙♣❡❡❝❤ Pr♦❞✉❝t✐♦♥ ❛♥❞ ❙♣❡❡❝❤ ▼♦❞❡❧❧✐♥❣✱ ❲✳❏✳ ❍❛r❞❝❛st❧❡ ❛♥❞ ❆✳ ▼❛r❝❤❛❧✱ ❊❞s✱ ✈♦❧✉♠❡ ✺✺✱ ✶✾✽✾✳ ❬❚✐s❤❜②✳◆✭✶✾✾✵✮❪ ❚✐s❤❜②✳◆✳ ❆ ❞②♥❛♠✐❝❛❧ s②st❡♠s ❛♣♣r♦❛❝❤ t♦ s♣❡❡❝❤ ♣r♦✲ ❝❡ss✐♥❣✳ ■♥ Pr♦❝✳ ■♥t✳ ❈♦♥❢✳ ❆❝♦✉st✐❝s✱ ❙♣❡❡❝❤ ❛♥❞ s✐❣♥❛❧ ♣r♦❝❡ss✐♥❣✱ ❆❧❜✉q✉❡rq✉❡✱ ✶✾✾✵✳ ❬t♦rr❡♥❝❡ ❛♥❞ ❈♦♠♣♦✭✶✾✾✽✮❪ ❈ t♦rr❡♥❝❡ ❛♥❞ ● P ❈♦♠♣♦✳ ❆ ♣r❛❝t✐❝❛❧ ❣✉✐❞❡ t♦ ✇❛✈❡❧❡t ❛♥❛❧②s✐s✳ ❇✉❧❧✳ ❆♠❡r✳ ▼❡t❡♦r♦❧♦❣✐❝❛❧✳ ❙♦❝✐❡t②✱ ✼✾✭✶✮✿✻✶✕✼✽✱ ✶✾✾✽✳ ❬❚♦✉ ❛♥❞ ●♦♥③❛❧❡③✭✶✾✼✺✮❪ ❏✳ ❚✳ ❚♦✉ ❛♥❞ ❘✳ ❈✳ ●♦♥③❛❧❡③✳ P❛tt❡r♥ ❘❡❝♦❣♥✐✲ t✐♦♥ Pr✐♥❝✐♣❧❡✳ ❆❞❞✐s♦♥ ❲❡s❧❡②✱ ✶✾✼✺✳ ❬❚s❛♥❣✲▲♦♥❣ P❛♦ ❛♥❞ ▲✐✭✷✵✵✻✮❪ ❏✉♥❍❡♥❣ ❨❡❤ ❚s❛♥❣✲▲♦♥❣ P❛♦✱ ❨✉✲❚❡ ❈❤❡♥ ❛♥❞ P❡✐✲❏✐❛ ▲✐✳ ▼❛♥❞❛r✐♥ ❡♠♦t✐♦♥❛❧ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥ ❜❛s❡❞ ♦♥ s✈♠ ❛♥❞ ♥♥✳ ■♥ Pr♦❝ ♦❢ t❤❡ ✶✽t❤ ✐♥t❡r♥❛t✐♦♥❛❧ ❝♦♥❢❡r❡♥❝❡ ♦♥ P❛tt❡r♥ ❘❡❝♦❣✲ ♥✐t✐♦♥✱ ✷✵✵✻✳ ❬❚s❡♥❣ ❛♥❞ ❍✉❛♥❣✭✷✵✸✮❪ ❨✳ ❈❤❡♥❣✲❲✳ ▲❡❡ ❚s❡♥❣✱ ❈✳ ❛♥❞ ❋✳ ❍✉❛♥❣✳ ❈♦❧✲ ❧❡❝t✐♥❣ ♠❛♥❞❛r✐♥ s♣❡❡❝❤ ❞❛t❛❜❛s❡s ❢♦r ♣r♦s♦❞② ✐♥✈❡st✐❣❛t✐♦♥s✳ ■♥ Pr♦❝✳ ❚❤❡ ❖r✐❡♥t❛❧ ❈❖❈❖❙❉❆✱ ✷✵✸✳ ❬❯✳❑r❡❜❡❧✭✶✾✾✾✮❪ ❯✳❑r❡❜❡❧✳ P❛✐r✇✐s❡ ❈❧❛ss✐✜❝❛t✐♦♥ ❛♥❞ ❙✉♣♣♦rt ❱❡❝t♦r ▼❛✲ ❝❤✐♥❡s✳ ▼■❚ Pr❡ss✿❈❛♠❜r✐❞❣❡✱ ✶✾✾✾✳ ❬❱❛♣♥✐❦✭✶✾✾✺✮❪ ❱✳ ◆✳ ❱❛♣♥✐❦✳ ❚❤❛ ◆❛t✉r❡ ♦❢ ❙t❛t✐st✐❝❛❧ ▲❡❛r♥✐♥❣ ❚❤❡♦r②✳ ◆❡✇②♦r❦ ✿ ❙♣r✐♥❣❡r✲❱❡r❧❛❣✱ ✶✾✾✺✳ ❬❱❛♣♥✐❦✭✶✾✾✽✮❪ ❱✳ ◆✳ ❱❛♣♥✐❦✳ ❙t❛t✐st✐❝❛❧ ▲❡❛r♥✐♥❣ ❚❤❡♦r②✳ ◆❡✇②♦r❦ ✿ ✇✐❧❡②✱ ✶✾✾✽✳ ❬❱❡tt❡r❧✐✭✶✾✾✷✮❪ ◆❛rt✐♥ ❱❡tt❡r❧✐✳ ❲❛✈❡❧❡ts ❛♥❞ ✜❧t❡r ❜❛♥❦s✿ ❚❤❡♦r② ❛♥❞ ❞❡✲ s✐❣♥✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✹✵✭✾✮✿✷✷✵✼✕✷✷✸✷✱ ✶✾✾✷✳ ❬❱❡tt❡r❧② ❛♥❞ ❍❡r❧❡②✭✶✾✾✷✮❪ ▼ ❱❡tt❡r❧② ❛♥❞ ❈ ❍❡r❧❡②✳ ❲❛✈❡❧❡ts ❛♥❞ ✜❧t❡r ❜❛♥❦s ✿ ❚❤❡♦r② ❛♥❞ ❞❡s✐❣♥✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✹✵✭✾✮✿ ✷✷✵✼✕✷✷✸✷✱ ✶✾✾✷✳ ✷✼✷

❬❱✐♥ts②✉❦✭✶✾✻✽✮❪ ❚✳ ❑✳ ❱✐♥ts②✉❦✳ ❙♣❡❡❝❤ ❞✐s❝r✐♠✐♥❛t✐♦♥ ❜② ❞②♥❛♠✐❝ ♣r♦✲ ❣r❛♠♠✐♥❣✳ ❑✐❜❡r♥❡t✐❦❛✱ ✹✭✷✮✿✽✶✕✽✽✱ ✶✾✻✽✳ ❬❱✐t❡r❜✐✭✶✾✻✼✮❪ ❆✳ ❏✳ ❱✐t❡r❜✐✳ ❊rr♦r ❜♦✉♥❞s ❢♦r ❝♦♥✈♦❧✉t✐♦♥❛❧ ❝♦❞❡s ❛♥❞ ❛♥ ❛s②♠♣t♦t✐❝❛❧❧② ♦♣t✐♠❛❧ ❞❡❝♦❞✐♥❣ ❛❧❣♦r✐t❤♠✳ ■❊❊❊ ❚r❛♥s✳ ■♥❢♦r♠❛✐t♦♥ ❚❤❡♦r②✱ ■❚✲✶✸✭✶✮✿✷✻✵✕✷✻✾✱ ✶✾✻✼✳ ❬❲❡✐❜❡❧✳❆ ❛♥❞ ▲❛♥❣✳❑✭✶✾✽✽✮❪ ❍✐♥t♦♥✳● ❙❤✐❦❛♥♦✳❑ ❲❡✐❜❡❧✳❆✱ ❍❛♥❛③❛✈❛✳❚ ❛♥❞ ▲❛♥❣✳❑✳ P❤♦♥❡♠❡ r❡❝♦❣♥✐t✐♦♥ ✿ ◆❡✉r❛❧ ♥❡t✇♦r❦s ✈s✳ ❤✐❞❞❡♥ ♠❛r❦♦✈ ♠♦❞❡❧s✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ◆❡✉r❛❧ ◆❡t✇♦r❦s✱ ✽✭✷✮✱ ✶✾✽✽✳ ❬❲❤✐t♥❡②✭✶✾✸✻✮❪ ❍✳ ❲❤✐t♥❡②✳ ❉✐✛❡r❡♥t✐❛❜❧❡ ♠❛♥✐❢♦❧❞s✳ ❆♥♥✳ ▼❛t❤✳✱ s❡r✳ ✷♥❞✱ ✸✼✿✻✹✺✕✻✽✵✱ ✶✾✸✻✳ ❬❲✐❧❧✐❛♠ ❍✉❛♥❣ ❛♥❞ ●♦❧❞✭✶✾✽✽✮❪ ❘✐❝❤❛r❞ ▲✐♣♣♠❛♥♥ ❲✐❧❧✐❛♠ ❍✉❛♥❣ ❛♥❞ ❇❡♥ ●♦❧❞✳ ❆ ♥❡✉r❛❧ ♥❡t ❛♣♣r♦❛❝❤ t♦ s♣❡❡❝❤ r❡❝♦❣♥✐t✐♦♥✳ ■❊❊❊ ❚r❛♥s✲ ❛❝t✐♦♥s ♦♥ ◆❡✉r❛❧ ◆❡t✇♦r❦s✱ ✽✭✷✮✱ ✶✾✽✽✳ ❬❳ ❍✉❛♥❣ ❛♥❞ ❍♦♥✭✷✵✵✶✮❪ ❆ ❆❝❡r♦ ❳ ❍✉❛♥❣ ❛♥❞ ❍ ❍♦♥✳ ❙♣♦❦❡♥ ▲❛♥❣✉❛❣❡ Pr♦❝❡ss✐♥❣✲❆ ●✉✐❞❡ t♦ ❚❤❡♦r②✱ ❆❧❣♦r✐t❤♠✱ ❛♥❞ ❙②st❡♠ ❉❡✈❡❧♦♣♠❡♥t✳ Pr❡♥t✐❝❡ ❍❛❧❧✱ ✷✵✵✶✳ ❬❳❛✈✐❡r ❉♦♠♦♥t ❛♥❞ ●♦❡r✐❝❦✭✷✵✵✼✮❪ ❍❡✐❦♦ ❲❡rs✐♥❣ ❋r❛♥❦ ❏♦✉❜❧✐♥ ❙t❡❢❛♥ ▼❡♥③❡❧ ❇❡r♥❤❛r❞ ❙❡♥❞❤♦✛ ❳❛✈✐❡r ❉♦♠♦♥t✱ ▼❛rt✐♥ ❍❡❝❦♠❛♥♥ ❛♥❞ ❈❤r✐st✐❛♥ ●♦❡r✐❝❦✳ ❲♦r❞ r❡❝♦❣♥✐t✐♦♥ ✇✐t❤ ❛ ❤✐❡r❛r❝❤✐❝❛❧ ♥❡✉r❛❧ ♥❡t✲ ✇♦r❦✳ ■♥ ✧✐♥ Pr♦❝✳ ■♥t✳ ❈♦♥❢✳ ♦♥ ◆♦♥✲▲✐♥❡❛r ❙♣❡❡❝❤ Pr♦❝❡ss✐♥❣✱ ◆❖✲ ▲■❙P✧✱ ✷✵✵✼✳ ❬❳✐❛ ❛♥❞ ❩❤❛♥❣✭✶✾✾✸✮❪ ❳✐❛♥❣✲●❡♥ ❳✐❛ ❛♥❞ ❩❤❡♥ ❩❤❛♥❣✳ ❖♥ s❛♠♣❧✐♥❣ t❤❡✲ ♦r❡♠✱ ✇❛✈❡❧❡ts ❛♥❞ ✇❛✈❡❧❡t tr❛♥s❢♦r♠s✳ ■❊❊❊ ❚r❛♥s❛❝t✐♦♥s ♦♥ ❙✐❣♥❛❧ Pr♦❝❡ss✐♥❣✱ ✹✶✭✶✷✮✿✸✺✷✹✕✸✺✸✺✱ ✶✾✾✸✳ ❬❨♦♥❣ ❛♥❞ ❚✐♥❣✭✷✵✶✶✮❪ ❇ ❋ ❨♦♥❣ ❛♥❞ ❍ ◆ ❚✐♥❣✳ ❙♣❡❛❦❡r✲✐♥❞❡♣❡♥❞❡♥t ✈♦✇❡❧ r❡❝♦❣♥✐t✐♦♥ ❢♦r ♠❛❧❛② ❝❤✐❧❞r❡♥ ✉s✐♥❣ t✐♠❡✲❞❡❧❛② ♥❡✉r❛❧ ♥❡t✇♦r❦✳ ■♥ ■♥ Pr♦❝✳ ✐♥t✳ ❈♦♥❢✳ ♦♥ ❇✐♦♠❡❞✐❝❛❧ ❊♥❣✐♥❡❡r✐♥❣✱ ♣❛❣❡s ✺✻✺✕✺✻✽✱ ✷✵✶✶✳ ❬❨✉✭✷✵✶✶✮❪ ▼✐♥✲❈❤✉♥ ❨✉✳ ▼✉❧t✐ ✲ ❝r✐t❡r✐❛ ❛❜❝ ❛♥❛❧②s✐s ✉s✐♥❣ ❛rt✐✜❝✐❛❧ ✐♥t❡❧✲ ❧✐❣❡♥❝❡ ❜❛ss❡❞ ❝❧❛ss✐✜❝❛t✐♦♥ t❡❝❤♥✐q✉❡s✳ ❊❧s❡✈✐❡r✲❊①♣❡rt ❙②st❡♠s ✇✐t❤ ❆♣♣❧✐❝❛t✐♦♥s✱ ✸✽✿✸✹✶✻✕✸✹✷✶✱ ✷✵✶✶✳ ❬❨✉❝❡❧ ❖③❜❡❦ ❛♥❞ ❉❡♠✐r❡❦❧❡r✭✷✵✶✷✮❪ ▼❛r❦ ❍❛s❡❣❛✇❛ ❏♦❤♥s♦♥ ❨✉❝❡❧ ❖③❜❡❦ ❛♥❞ ▼✉❜❡❝❝❡❧ ❉❡♠✐r❡❦❧❡r✳ ❖♥ ✐♠♣r♦✈✐♥❣ ❞②♥❛♠✐❝ st❛t❡ s♣❛❝❡ ❛♣✲ ♣r♦❛❝❤❡s t♦ ❛rt✐❝✉❧❛t♦r② ✐♥✈❡rs✐♦♥ ✇✐t❤ ♠❛♣✲❜❛s❡❞ ♣❛r❛♠❡t❡r ❡st✐♠❛✲ t✐♦♥✳ ■❊❊❊ ❚r❛♥s✳ ♦♥ ❆✉❞✐♦✱ ❙♣❡❡❝❤ ❛♥❞ ▲❛♥❣✉❛❣❡ Pr♦❝❡ss✐♥❣✱ ✷✵✭✶✮✿ ✻✼✕✽✶✱ ✷✵✶✷✳ ✷✼✸

❬❩❛♥✉②✭✷✵✵✼✮❪ ▼❛r❝♦s ❋❛✉♥❞❡③ ❩❛♥✉②✳ ❖♥ t❤❡ ✉s❡❢✉❧♥❡ss ♦❢ ❧✐♥❡❛r ❛♥❞ ♥♦♥✲ ❧✐♥❡❛r ♣r❡❞✐❝t✐♦♥ r❡s✐❞✉❛❧ s✐❣♥❛❧s ❢♦r s♣❡❛❦❡r r❡❝♦❣♥✐t✐♦♥✳ ■♥ Pr♦❝✳ ■♥t✳ ❈♦♥❢✳ ♦♥ ◆♦♥✲▲✐♥❡❛r ❙♣❡❡❝❤ Pr♦❝❡ss✐♥❣✱ ◆❖▲■❙P✱ ✷✵✵✼✳ ❬❩❤❡♥❣ ❛♥❞ ❲✉✭✷✵✵✷✮❪ P✳ ❨❛♥ ❍✳ ❙✉♥ ▼✳ ❳✉ ❩❤❡♥❣✱ ❚✳❋✳ ❛♥❞ ❲✳ ❲✉✳ ❈♦❧✲ ❧❡❝t✐♦♥ ♦❢ ❛ ❝❤✐♥❡s❡ s♣♦♥t❛♥❡♦✉s t❡❧❡♣❤♦♥❡ s♣❡❡❝❤ ❝♦r♣✉s ❛♥❞ ♣r♦♣♦s❛❧ ♦❢ r♦❜✉st r✉❧❡s ❢♦r r♦❜✉st ♥❛t✉r❛❧ ❧❛♥❣✉❛❣❡ ♣❛rs✐♥❣✳ ■♥ Pr♦❝✳ ❏♦✐♥t ■♥t❡r♥❛t✐♦♥❛❧ ❈♦♥❢❡r❡♥❝❡ ♦❢ ❙◆▲P✲❖❈❖❈❖❙❉❆✱ ♣❛❣❡s ✻✵✕✻✼✱ ✷✵✵✷✳ ❬❩❤✉♦✲♠✐♥❣ ❈❤❡♥ ❛♥❞ t❛♦ ❨❛♦✭✷✵✶✶✮❪ ❏✐❛♥✲❤✉✐ ❩❤❛♦ ❩❤✉♦✲♠✐♥❣ ❈❤❡♥✱ ❲❡✐✲ ①✐♥ ▲✐♥❣ ❛♥❞ ❚❛♦ t❛♦ ❨❛♦✳ ❈♦♥s♦♥❛♥t r❡❝♦❣♥✐t✐♦♥ ♦❢ ❞②s❛rt❤r✐❛ ❜❛s❡❞ ♦♥ ✇❛✈❡❧❡t tr❛♥s❢♦r♠ ❛♥❞ ❢✉③③② s✉♣♣♦rt ✈❡❝t♦r ♠❛❝❤✐♥❡s✳ ❏♦✉r♥❛❧ ♦❢ ❙♦❢t✇❛r❡✱ ✻✭✺✮✿✽✽✼✕✽✾✸✱ ✷✵✶✶✳

✷✼✹

Internati onal Journal of Innovations & Advancement in Computer Science IJ IACS ISSN 2347 – 8616 Volume 3, Issue 6 August 2014

LDA Classifier for Animal Identification Mamatha G M.Tech Co mputer Science and Engineering Nitte Meenakshi Institute of Technology Bangalore-560064

Dr Jharna Maju mdar DEAN, R & D, Prof. & Head, Dept. of CSE (PG) Nitte Meenakshi Institute of Technology Bangalore-560064

Abstract—Nowadays animal identificat ion is an emerging area, due to extinction and endangering of animals. In order to identify an animal in an image, the image has to be described or represented by certain features. Shape is an important visual feature of an image. Searching for images using shape features has attracted much attention. There are many shape representation and description techniques in the literature. An approach for recognizing animal using Agglomerative Hierarchical Clustering and LDA (Linear Discriminate Analysis) is brought forward. Different steps are used to identify the images. First step is preprocessing, Second step is feature extraction, and the next step is agglomerative hierarchical clustering. Then, Fisher’s Linear Discriminant Analysis (FLDA) is used to classify the testing images to their appropriate categories. Key words: AHC (Agglomerative Hierarchical Clustering), Feature extraction, and LDA (Linear Discriminate Analysis), Preprocessing. I. INTRO DUC TIO N Due to the rapid concretization of the lands and urbanization, the shelter places of the animals, birds etc. are getting dimin ished and these creatures are finding it very difficult to survive and some species are becoming extinct. To survive life of animal intrusion in residential area fro m residential people to get inform to prevent an accidental hurt to wild life and save both civilian and wild life. This system prevents the intruders from entering into the restricted area. Moreover the detection must be done on animals. This is so because in many cases the animals hurt themselves by the residential people. To p revent this, one detects animal if in the case of to prevent the animals fro m harm. There are so many human detectors wh ich are capable of detecting only humans. But those detectors cannot be used for animal detection [7]. Since the humans and animals shapes and texture varies the human detectors cannot be simply app licable to animals. This urged to find the detector for the animal. The process is init iated once the object captured by the camera in the surveillance area. Object detection is an important problem in the field of Co mputer Vision. Also, the detection of animals in wildlife videos is a topic of great interest among biologists and

66

Mamatha G, Dr Jharna Majumdar, Bhuvaneswari

Mrs. Bhuvaneswari S. Patil Assistant Professor, Co mputer Science and Engineering Nitte Meenakshi Institute of Technology Bangalore-560064.

wild life enthusiasts. While recognition of humans in images has been actively explored by researchers in the past, but little work has been done on the problem o f animal detection. Work done for images in the case of animal detection is classification into the appropriate species based on the model that we will train using samples. Clustering is one solution to the case of unsupervised learning, where class labeling information of the data is not available. Clustering is a method where data is divided into groups (clusters) which „seem‟ to make sense. Clustering algorithms are usually fast and quite simple. They need no beforehand knowledge of the used data and form a solution by comparing the given samples to each other and to the clustering criterion. Image classification is one of the primary tasks in geocomputation, that being used to categorize for fu rther analysis such as land management, potential mapping, forecast analysis and soil assessment etc. Image classification is method by which labels or class identifiers are attached to individual pixels on basis of their characteristics. Traditionally, classification tasks are based Linear Discrimination Analysis (LDA). Th is classifier is generally characterized by having an exp licit underlying probability model, wh ich provides a probability of being in each class rather than simply a classification. In this paper, an attempt has been made to develop a LDA classification algorithm specifically for the classification . The proposed algorithm is coded in Visual studio 6.0 language. For the classification of the image, the training sets are chosen for different classes.

II. R ELATED W O R K The main steps in digital image processing: Image acquisition: In the image acquisition step using the suitable camera, the image of the co mponent is acquired and then subjected to digitizat ion. Preprocessing: The pre-processing is a series of operations performed on the scanned input image. It essentially enhances the image rendering it suitable for feature extraction. Preprocessing methods is use a small neighborhood of a pixel in an input image to get a new brightness value in the output

Internati onal Journal of Innovations & Advancement in Computer Science IJ IACS ISSN 2347 – 8616 Volume 3, Issue 6 August 2014

image. Such pre-processing operations are also called filtration.

whether shape features are extracted from the whole shape region [2]. The contour and region based methods are classified into two methods: Structural approaches Global approaches. Gl obal method: It computes features on the whole shape. It takes less computation when compare to the structural method [3]. The global methods are Area, perimeter, Bounding box, half area, rectangularity, black occupancy, slimness, roundness, centroid, circularity, eccentricity, major axis orientation etc. Structural method: It computes features on subsets of the shape. It has more computation when compare to global method. It has more features to extract from an image. The structural features are convex hull, principal axes method, chain code. Clustering using hierarchical agglomerati ve clustering

Figure 1: set of animal database The pre-processing stage takes a raw image then following operations are applied on it: Threshol ding: Raw image either color or grey is converted into binary image. Binary Image: Binary image is the simp lest type of images and can take on two values, typically black and white, or „0‟ and „1‟. A binary image is referred to as a 1 bit/pixel image because it takes only 1 binary digit to represent each pixel. In this project Automatic thresholding is used. Noise reduction: Various techniques like mo rphological operations are used to connect unconnected pixels, to remove isolated pixels, to smooth pixels boun dary. Normalization: The segmented image is normalized to 32* 32, 64*64, 128* 128 matrixes. Features extraction First of all, Several Features were incorporated in the classification system will be described. The Features can be categorized into three features such as Shape features, Color features, Texture features. The Shape Features are Area, Perimeter, Bounding box, half area, Black occupancy, Slimness, Rectangularity, Eccentricity, Roundness, Dispersion, Solidity, Centroid, and Convexity etc. Shape representation and description techniques can be generally classified into two classes of methods: Contour-based methods Region-based methods. The classification is based on whether shape features are ext racted from the contour only. The classification is based on

67

Mamatha G, Dr Jharna Majumdar, Bhuvaneswari

Learning: “learning denotes the changes in a system that enable a system to do the same task more efficient ly the next time”. Machine learning: Machine learning is a scientific discipline concerned with the design and development of algorithms that allow computers to evolve behaviors based on data, such as fro m data or database. Learning can take an advantage of data to capture characteristics of interest. Types of machine learning: 1. Supervised learning 2. Unsupervised learning Unsupervised learning: If data has to be processed by the machine learning method, where the where the desired output is not given, then the learning task is called unsupervised. The task of grouping related data points together without labeling them. Clustering: Cluster is a group of objects that belong to the same class. In other words the similar object are grouped in one cluster and dissimilar are grouped in other cluster. Hierarchical Clustering: Hierarchical clustering algorith ms are either top-down or bottom-up. Bottom-up algorithms t reat each document as a singleton cluster at the outset and then successively merge (or agglomerate) pairs of clusters until all clusters have been merged into a single cluster that contains all documents. Bottom-up hierarch ical clustering is therefore called hierarch ical agglomerat ive clustering or HAC [4]. Topdown clustering requires a method for splitting a cluster. It can be broadly classified into two categories: agglomerative and divisive.

Internati onal Journal of Innovations & Advancement in Computer Science IJ IACS ISSN 2347 – 8616 Volume 3, Issue 6 August 2014

method maximizes the ratio of between -class variance to the within-class variance in any particular data set thereby guaranteeing maximal separability. The use of Linear Discriminant Analysis for data classification is applied to classification problem. We decided to implement an algorithm for LDA in hopes of providing better classification compared to Principal Co mponents Analysis.

Extracted feature file

Agglomerati ve hierarchi -cal cluster

Class 2

Class N Figure 3: cluster module. The agglomerative algorith m for h ierarchical clustering starts by placing each of the objects in the data set in an individual cluster and then gradually merges those individual clusters. The divisive algorithm however, starts with the whole data set as a single cluster and then breaks it down into fewer clusters. Single Link and Complete Link are two hierarch ical agglomerative clustering procedures. In the Single Lin k clustering algorithm, clusters are merged at each stage by the single shortest link between them. During each iteration, after the clusters p and q are merged, the distance between the new cluster, say n, and some other cluster, say r, is given by dr = min(dpr, dqr), where dnr denotes the distance between the two closest members of clusters n and r. If the clusters n and z were to be merged, then for any object in the resulting cluster, the distance to its nearest neighbor would be at most dnr shown in figure 3. Classification using LDA Supervised learning : Tasks where the desired output for each object is given and called as supervised and the desired outputs called targets. It generates a function that maps inputs to desired outputs. The task is to assigning instances to predefined classes . Classification: Classification is a supervised learning. Classification consists of assigning a class label to a set of unclassified cases [3]. Supervised Classification : The set of possible classes is known in advance. There are many possible techniques for classification of data. Principal Co mponent Analysis (PCA) and Linear Discriminant Analysis (LDA) are two common ly used techniques for data classification and dimensionality reduction [8]. Linear Discriminant Analysis easily handles the case where the within-class frequencies are unequal and their performances have been examined on randomly generated test data. This

68

Mamatha G, Dr Jharna Majumdar, Bhuvaneswari

The prime difference between LDA and PCA is that PCA does mo re of feature classification and LDA does data classification. In PCA, the shape and location of the original data sets changes when transformed to a different space whereas LDA doesn‟t change the location but only tries to provide more class separability and draw a decision region between the given classes. This method also helps to better understand the distribution of the feature data. LDA that Perform dimensionality reduction while preserving as much of the class discriminatory informat ion as possible [6]. There are two methods of LDA Fisher‟s Lin ear Discriminant method. Mult iple Discriminant method.

The main Objective Fisher‟s Linear Discriminant is to find a projection which separates data clusters. Maximize the between-class scatter while min imizing the within -class scatter. The algorithm steps for FLDA are: Input: Training images, classes, Test image of size m x n. Output: Recognition of image. m: nu mber of rows, n: nu mber of colu mns. Step1: calculate the mean ( ) for all the train ing set of images and the class mean ( i ). Step2: The between-class scatter matrix is the scatter of the expected vectors around the mixtu re mean: k

Sb

Pk (

)T

k

k 1

)(

k

k= nu mber of classes.

Pk 

Nk k

Nl l 1

Step3: The within-class scatter matrix is the scatter of the expected vectors around the mixtu re mean:

Sw

Sw

i

Step4: Using between class and within class value extraction of the eigen vector value (v). Step4: Calculate the Fisher factor value using this eigen vector (FFi). Step5: Calculate the value of the projected image (Px) for FFi * i . training set using Px

Internati onal Journal of Innovations & Advancement in Computer Science IJ IACS ISSN 2347 – 8616 Volume 3, Issue 6 August 2014

Step6: Calculate the value of the projected image (Py) for test image using Py FFi * Y . Step7: Calcu late the distance between the Px and Py using Euclidian d istance. Step8: Recognition of the image.

is used in animal identificat ion each of the new d imensions is a linear combination of pixel values, which form a template. The linear co mbinations obtained using Fisher's linear Discriminant are called Fisher faces. R EF E R EN C E S

III. Experimental results and analysis The organization of the project, as seen in figure 1 clearly separates different parts of the program. The Core contains the main processing like image acquisition, pre-processing, and feature extraction. The dataset contains 12 kinds of animals further imp rove the quality of a clustering; we used them to refine the solutions produced by the hierarch ical clustering algorith m described in Section II that is based on the generalized group average model. The feature values are used as input to the clustering algorithm so that the output is the classes. Recall from Sections 2 that the hierarch ical clustering algorith m that was used to obtained the clustering solutions as well as the clustering objective functions that are used by classification algorith ms. So the next step is clustering. Output of the agglomerative hierarch ical clustering method is the classes. This trained data is used for the classification and

CONCLUS ION In this paper, the feature extraction can be done for gray-level images from predefined set of database in figure 1. Image clustering is done using the agglomerative hierarchical clustering. Next using the clusters as the input to the Linear Discriminant Analysis, classification method is done to identify the animals. The application of LDA is in Face recognition. In computerized face recognition, each face is represented by a large nu mber of pixel values. Linear Discriminant analysis is primarily used here to reduce the number of features to a more manageable number before classification. So that this method

69

Mamatha G, Dr Jharna Majumdar, Bhuvaneswari

[1] Mukesh B Rangdal, and Dinesh B. Hanchate, “Animal Detection Using Histogram Oriented Gradient,” International Journal on Recent and Innovation Trends in Co mputing and Co mmunication , vol. 2, feb 2014. [2] Abdul Kadir, Lukito Ed i Nug roho, Adhi Susanto and Leaf Identification System Using Principal Co mponent Analysis” Gadjah Mada University, Indonesia, vol. 44, July 2012. [3] Tuyen Ly, John Hammer, “Classification Of Cats And Dogs,” Dept. Of Mathemat ics, 2012. [4] Dr. N. Rajalingam and K. Ranjini, “Hierarch ical Clustering Algorith m –A Co mparative Study,” International Journal of Co mputer Applications, Vo l. 19– No.3, April 2011. [5] Jukka Kainulainen, “Clustering Algorithms: Basics and Visualizat ion,” Helsinki Un iversity o f Information Science. October 2002. [6] Sebastian, Mikat, Gunnar fitscht, Jason Weston, Bernhard Scholkopft, and Klaus-Robert Mullert. “Fisher Discriminant Analysis with Kernels,” University of London, Egham, Surrey, UK, IEEE 1999. [7] Ian T. Young , Jan J. Gerbrands and Lucas J. van Vliet, “Fundamentals of Image Processing” Delft University of Technology, Version 2.3, 1995-2007. [8] S. Balakrishnama and A. Ganapathiraju, “Linear Discriminant Analysis - A Brief Tutorial” Institute for Signal and Information Processing Department of Electrical and Co mputer Engineering.

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

368

Article

Tilted Plane Symmetric Magnetized Cosmological Models D. D. Pawar#*, V. J. Dagwal @ & Y. S. Solanke & #

School of Mathematical Sciences, Swami Ramanand Teerth Marathwada University, Vishnupuri, Nanded-431606, (India) @ Dept. of Mathematics, Govt. College of Engineering, Amravati 444 604, India & Mungsaji Maharaj Mahavidyalay, Darwha ,Yawatmal

Abstract In this paper we have investigated tilted plane symmetric cosmological models in presence and absence of magnetic field. To get the deterministic model, we have assumed the supplementary conditions p = 0 and B = An where A and B are metric potentials and n is constant. Some geometric aspects of the model are also discussed. Keywords: tilted models, plane symmetric, dust fluid.

1. Introduction An anisotropic cosmological model plays an important role in the large scale behavior of the universe. The many researchers working on cosmology by using relativistic cosmological models have not given proper reasons to believe in a regular expansion for the description of the early stages of the universe. There are some experimental data of CMR and theoretical arguments which supports the existence of anisotropic universe (Verma et.al [1], Chimento [2], Misner [3], Land et .al [4], M. Demianski [5], Chawala et al [6]). The considerable interest has been focused in investigating spatially homogeneous and anisotropic universes in which the matter does not move orthogonally to the hyper surface of homogeneity in recent years. These are called tilted universes. King and Ellis [7]; Ellis and King [8]; Collins and Ellis [9] have studied the general dynamic of tilted cosmological models. Dunn and Tupper [10,11] have been studied tilted Bianchi type I cosmological model for perfect fluid and shown that a Bianchi tilting universe is possible when electromagnetic field is present. Many other researchers like Matravens et al. [12], Bali and Sharma [13], Horwood et al. [14], Hewit et al. [15], Aposotolopoulos [16] have studied different aspects of tilted cosmological models. Bianchi type V tilted cosmological models in the Scale-covariant theory derived by *

Correspondence Author: D. D. Pawar, School of Mathematical Sciences, Swami Ramanand Teerth Marathwada University, Vishnupuri, Nanded-431606, (India). E-mail: [email protected]

ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

369

Beesham [17]. Bali and Meena [18] have investigated tilted cosmological models filled with disordered radiation of perfect fluid, heat flow. Pradhan and Rai [19] have obtained tilted Bianchi type V cosmological models filled with disordered radiation in the presence of bulk viscous fluid and heat flow. Bhaware et al. [20] studied tilted cosmological models with varying

 . Pawar et al. [21] have investigated tilted plane symmetric cosmological models with heat conduction and disordered radiation. Pawar and Dagwal [22,23] have studied confirmally flat tilted cosmological models and recently two fluids tilted cosmological models in general relativity. Banerjee et al. [24] have studied an axially symmetric Bianchi type I string dust cosmological model in presence and absence of magnetic field. LRS Bianchi type string dust-magnetized cosmological models have been investigated by Bali and Upadhyay [25]. Stationary distribution of dust and electromagnetic field in general relativity has been investigated by Banerjee and Banerjee [26]. Pawar et al. [27, 28] have studied bulk viscous fluid with plane symmetric string dust magnetized cosmological model in general relativity and Lyra manifold. Patil et al. [29, 30] obtained on thick domain walls with viscous field coupled with electromagnetic field in general relativity and Lyra geometry. Bayaskar et al. [31] derived cosmological models of perfect fluid and massless scalar field with electromagnetic field. Bali et al. [32] have studied magnetized tilted universe for perfect fluid distribution in general relativity. Bagora et al. [33] have investigated tilted Bianchi type I dust fluid magnetized cosmological model in general relativity.

2. Field Equation We consider metric in the form –





ds 2   dt 2  A 2 dx 2  dy 2  B 2 dz 2 ,

(1)

where A and B are the functions of t alone. The Einstein’s field equation is – 1 Ri j  Rg ij  8 Ti j e (2) 2 The energy-momentum tensor for perfect fluid distribution with heat conduction is given by – Ti j    p  vi v j  p g ij  qi v j  vi q j  Ei j .

(3)

With g ij v i v j  1 , ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

(4) www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

370

qi q j  0 ,

(5)

qi v i  0 .

(6)

Here Ei j is the energy momentum tensor of electromagnetic field given by   2 1  Ei j    h  vi v j  g ij   hi h j  , 2    

(7)

where  is magnetic permeability and the magnetic flux vector hi is given by hi 

g  i j k F k v j , 2

(8)

Fk  is the electromagnetic field tensor,  i j k  is the Levi-Civita tensor density, P is pressure,  is

density and q i is heat conduction vector orthogonal to the fluid flow vector v i . The fluid flow sin h   vector v i has the components  0,0, , cos h  , satisfying condition (4); and  is the tilt B   angle. The incident magnetic field is taken on z-axis so that

h1  0 , h2  0 , h3  0 , h4  0 .

(9)

The first set of Maxwell’s equation is Fi j ; k  F j k ; i  Fk i ; j  0 .

(10)

F12  constant  M(say) .

(11)

Here Here F14  F24  F34  0 , due to assumption of infinite electrical conductivity. The only non-vanishing component of Fi j is F12 . Hence h3 

BM cos h ,  A2

h4 

M



(12)

sin h

(13)

and 2

h  hi h i 

M2 .  2 A4

(14)

From (7) we have, E11  E22 

M2   E33   E44 . 2 A4

(15)

The field equation (2) for metric (1) reduces to ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

371

 A AB B M2      8  p  , A AB B 2  A4  

(16)

 A2 2 A M2  sin h 2 p h p q    8 sin 2         3  , A2 A B 2  A4  

(17)

 sin h A2 2 AB M2  2    8      p  cos h  p  2q3  , 2  A4  A2 AB B 

(18)

h 0. cos h

  p B sin h cos h  q3 cos h  q3 sin

2

(19)

Here the dot (.) over a field variable denotes the differentiation with respect to time t.

3. Solution of the Field Equations The set (16) – (19) being highly non-linear containing six unknowns (A, B, , , p and q3). So to obtain a determinate solution we have to use two additional constraints. Let us first assume that the model is filled with dust which leads to P=0 (20) and secondly we consider that the scalar expansion θ is proportional to the shear scalar σ leads to B = An , (21) where n is constant. Equation (17), (18) lead to 2 A2 2 A 2 AB 8 M 2 8      .  A4 A2 A AB

(22)

Equating (16) and (20) we have A AB B 8 M 2    . A AB B 2  A4

(23)

Equating (21) and (23) we have 2A 

2 n 2 A2 N ,   n  1 A  n  1 A3

(24)

where N

 8 M 2



.

(25)

The equation can be rewritten as

ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

372

df 2 a 2 N  f  , n  1 A3 dA A

(26)

2n 2 , where a  n  1

(27)

A  f  A ,

(28)

and

A  ff '

f '

when

df . dA

(29)

Equation (26) leads to 2

N C  dA   a ,    2 2a1 A A  dt 

where a1  n 2  n  1 ,

a

(30)

2n 2 & C is constant of integration . n 1

The metric (1) reduces to form 2

 dt  2 2 2 2 2n 2 ds 2     dA  A dx  dy  A dz  dA  Equating (31) we get





1

(31)

 N C  a  dT 2  T 2 dX 2  dY 2  T 2 n dZ 2 ds    2 T   2a1T



2



(32)

Where A  T , dx  dX , dy  dY , dz  dZ .

4. Some Physical and Geometrical Properties The density for the model (31) is given by a N aC 8  2 4  3a4 , n ≠ -1 T T where a2 





 n 2  2n  3 , n2  n  1

a3 



(33)



2 2n 2  2n  1 n  1

&

a4 





2 n2  n  1 . n  1

The tilt angle  is given by 1 2

 a NT  a7  cos h   5  , a6  a8 NT  a9  a6

where a5   n 3  2n 2  5n  2  , a6 

ISSN: 2153-8301



(34)



2 n2  n 1 , n  1

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models







a7  n 4  2n 3  3n 2  4n  1 C ,

and





373



a8  2 n 2  1 ,

a9  2  2 n 4  4 n 3  2 n C . 1

 a NT a6  a11  2 sin h   10 a6  ,  a8 NT  a9 

where a10   n 3  5n  4 ,

(35)





a11   3n 4  6n 3  3n 2  8n  1 C .

The expansion () calculated for the flow vector v i is given by



n  2  T

1 2

N C   a NT  a7   ,  a   5  a6 2 a T T 2 a NT a  9   8  1 a6

n ≠ -2

(36)

The flow vector v i and heat conduction vector q i for the metric (31) are given by 1

1  a NT a6  a11  2 v 3  n  10 a6  , T  a8 NT  a9 

(37)

1

 a NT a6  a7  2 v 4   5 a6  ,  a8 NT  a9 

(38)

1

 a10 NT a6  a11  2  a5 NT a6  a7      a8 NT a6  a9   a8 NT a6  a9  3  , q   1  n N 2nn  1C  n T 8    4 T a4  a1T 

(39)

1

 a10 NT a6  a11   a5 NT a6  a7  2    a8 NT a6  a9   a8 NT a6  a9  4  . q   1  n N 2nn  1C  8    4 T a4  a1T 

(40)

The non-vanishing components of shear tensor  i j  and rotation tensor wi j  are given by

2 

21  n   N C  a  2 3T  2a1T T 

1 2

n≠1

,

(41)

   a 2 N a 3C     a 4  T N C   a10 NT  a11   T 4   n 1  . 34  nT   a    a6 2 T   a8 NT  a9   1  n  N a  n  1    2a1T    4 nT a4     a 1T 1 2

a6

1 2

(42)

The rate of expansion H i in the direction of x, y and z-axis respectively are given by ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

374

1

2 N C 2 H1  H 2     , T  2a1T 2 T a 

(43)

1

C 2 2n  N  H3    . T  2a1T 2 T a 

(44)

5. Discussion At the initial moment energy density is infinite where as it vanishes when T is infinite. i.e The model (32) starts with big bang at T = 0 and stops at T   .Thus model has point -type singularity at T = 0.The tilt angles and the flow vectors are a7 , sin h  a9

cos h 

Whereas cos h 

a7 , when T = 0 provided n  1 , a9

a11 , v3   , v 4  a9

a5 , sin h  a8

a10 , v3  0 , v 4  a8

a5 , when T   provided n  1 . Thus a8

tilt angles are constant for both T = 0 and T   . From (36), initially the rate of expansion  is infinite, it decreases when time increases and the expansion stop at T   . At T = 0 shear scalar  is infinite provided n  1 where as it vanishes when T is infinite. Initially directional Hubble parameters are infinite and it vanishes for large value of T. Heat conduction vector q 3 is infinite at T = 0 and vanishes at infinite time provided n > 4 and n > a4 and q 4 vanishes when T = 0 but it is infinite for large value of T. Since lim  T   

   0 the model not approach isotropy for large value of T. 

In this case, the models are expanding, shearing, rotating for tilted universe.

When the magnetic field is absent ( N  0 ): ds 2  dt 2   kT 

2

k

 dx

2



 dy 2   kT 

2n

k

dz 2

(45)

n2  n  1 n 1 The density for the model (45) is given by

where T   ct  d  & k 

ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models



375

2c 2  n  k  2  8  kT 

(46)

2

Tilt angles are given by 2k  3 2 1  n  k 

cos h 

(47)

4k  2n  5 2 1  n  k 

sin h 

(48)

The flow vector v i and heat conduction vector q i for the metric (45) are given by 1

v3 

1

 kT 

n

k

 4k  2n  5  2    2 1  n  k  

(49)

1

 2k  3  2 v4     2 1  n  k   q 

c 2  2k  3

3

8  kT 

q  4

4k  2n  5 2 1  n  k 

(2 k  n ) k

c 2  4k  2n  5  8  kT 

(50)

2

(51)

2k  3 2 1  n  k 

(52)

Scalar expansion (), the non-vanishing component of shear tensor  i j  and rotation tensor

  respectively are given by ij



 n  2 kT

2k  3 2 1  n  k 

2  c 1  n       3  kT 

(53)

2

2

34  nc  kT 

nk k

(54)

 k  n  2  4k  2n  5  n  k  1  2 1  n  k   

(55)

The rate of expansion H i in the direction of x, y and z-axis respectively are given by H1  H 2 

ISSN: 2153-8301

2c kT

(56)

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

H3 

2nc kT

376

(57)

At the initial moment energy density and scalar expansion are infinite where as energy density and scalar expansion vanish when T is infinite. The tilt angles are (47) and (48) which are the same at any instant. The flow vectors v3 is infinite at T= 0 and vanishes when time is infinite. The flow vectors v4 is constant for both T = 0 and T   . Initially heat conduction vector q 3 & q 4 are infinite and it vanishes for large value of T. At T = 0 shear scalar  is infinite

where as it vanishes when T is infinite. Initially directional Hubble parameters are infinite and it vanishes for large value of T. lim   Since T     0 the model not approach isotropy for large value of T.  

7. Conclusion In the present paper we have constructed tilted plane symmetric dust fluid cosmological model in presence and absence of the magnetic field. We have obtained a determinate solution by assuming the conditions that model is filled with dust (which leads to the zero pressure) and the expansion scalar ө is proportional to the shear scalar σ. We have discussed the physical behavior of the models in presence and absence of the magnetic field. Energy density, expansion scalar and shear scalar have the similar behavior in presence and absence of magnetic field. At the initial epoch all these parameters are infinite and decreases with increase in cosmic time. In the presence and absence of the magnetic field the models are expanding, shearing, rotating and tilted. Only variations in tilt angles are found. In the presence of the magnetic field tilt angles are the functions of time and it tends to finite number when time is infinite and they are infinite when time zero. Whereas tilted angles are constant (very small) throughout the evolution of the universe in absence of magnetic field. The model has real singularity at T=0 and it start with big bang and stops when cosmic time is infinite, in other word it becomes asymptotically empty. The present model does not approach isotropy in presence and absence of magnetic field lim

since

T 

  

   0. 

References 1) M. K. Verma et.al (2011). Rom-Journ. Phys., 56,3-4,616-626 2) L.P. Chimento (2004).phy. Rev. D .69.123517. 3) C.W. Misner (1968) .Astrophys.J.151, 431. ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

Prespacetime Journal| April 2014 | Volume 5 | Issue 5 | pp. 368-377 Pawar, D. D., Dagwal V. J. & Solanke, Y. S., Tilted Plane Symmetric Magnetized Cosmological Models

377

4) K. Land, J. Maqueijo (2005).Phys. Revlelt.95, 071301. 5) M. Demianski (1992) Phy. Rev. D. 46, 4, 1391-1398. 6) Chanchal Chawla, R. K. Mishra, Anirudh Pradhan (2012): arXiv 1203.4014v3 [physics.gen-ph]. 7) A. R. King, G. G. R. Ellis (1973). Comm. Math. Phys., 31. 209. 8) G. G. R. Ellis, A. R. King (1974). Comm. Math. Phys., 38, 119. 9) C. B. Collins, G. G. R. Ellis (1979). Phys. Rep., 56, 65. 10) K. A. Dunn, B. O. J. Tapper (1978). Astrophys. J. 222, 405. 11) K. A. Dunn, B. O. J. Tupper (1980). Astrophys. J., 235, 307. 12) D. R. Matravers, M. s. Madsen, D. L. Vogel (1958). Astrophys. Space Sci., 112, 193. 13) R. Bali, K. Sharma (2002). Pramana J. Phys., 58, 457. 14) J. T. Horwood, M. J. Hancock, D. The Wainwright (2002). Preprint gr-ac/0210031. 15) C. G. Hewitt, R. Bridson, J. Wainstright (2001). Gen. Relativ. Gravitation, 36, 65. 16) P. S. Apostolopoulos (2003). Preprint gr-ac/0310033. 17) A. Beesham (1986). Astrophys. J., 125, 99. 18) R. Bali, B. L. Meena (2002). Astrophys. J., 281, 565. 19) A. Pradhan, A. Rai (2004). Astrophys. Space Sci., 291, 149. 20) S. W. Bhaware, D. D. Pawar, A. G. Deshmukh (2010), JVR 5, 3, 42-53. 21) D. D. Pawar, S. W. Bhaware, A. G. Deshmukh (2009). Room. J. Phys., 54, 187-194. 22) D. D. Pawar, V. J. Dagwal (2010). Bulg. J. Phys., 37, 165-175. 23) D. D. Pawar, V. J. Dagwal (2014) Int. J. Theor. Phys DOI:10.1007/s10773-014-2043-7. 24) A. Banerjee, A. K. Sanayal and S. Chakraborti (1990). Pramana J. Phys., 34, 1. 25) R. Bali, R. D. Upadhyaya (2003). Astrophys. Space Sci., 283, 97. 26) A. Banerjee, S. Banerjee (1968). J. Phys. A. Proc. Phys. Soc. (Gen) (GB), 2: 188. 27) D. D. Pawar, S. W. Bhaware, A. G. Deshmukh (2008). Int. J. Theor. Phys. 47 , 599–605 28) D. D. Pawar, V. R. Patil (2013) Prespacetime Journal, Vol. 4-3, pp.312-320. 29) V. R. Patil, D. D. Pawar, A. G. Deshmukh (2010) Rom. Rep. Phys. 62(4), 722–730. 30) V. R. Patil, D. D. Pawar, G. U. Khapekar (2012) Int J Theor Phys 51:2101–2108. 31) S. N. Bayaskar, D. D. Pawar, A. G. Deshmukh (2010) Rom. J. Phys. Vol. 54, Nos. 7-8, 763–740. 32) R. Bali, K. Sharma (2003). Astrophys. J., 283, 11. 33) A. Bagora, G. S. Rathore, P. Bagora (2009). Turk. J. Phys., 33, 167-177.

ISSN: 2153-8301

Prespacetime Journal Published by QuantumDream, Inc.

www.prespacetime.com

TELKOMNIKA, Vol. 13, No. 4, December 2015, pp. 1422~1436 ISSN: 1693-6930, accredited A by DIKTI, Decree No: 58/DIKTI/Kep/2013 DOI: 10.12928/TELKOMNIKA.v13i4.xxxx

 1422

Computing Game and Learning State in Serious Game for Learning Ririn Dwi Agustin, Ayu Purwarianti, Kridanto Surendro, Iping S Suwardi STEI ITB, Jalan Ganesa No. 10 Bandung, West Java, Indonesia *Corresponding author, email: [email protected], [email protected], [email protected], [email protected]

Abstract In order to support the adaptive SGfL, teaching materials must be representedin game component that becomes the target of adaptivity. If adaptive architecture of the game only use ‘game state’ (GS) to recognize player's state, SGfL require another indicator –‘learning state’ (LS)– to identify the learning progress. It is a necessary to formulate computational framework for both states in SGfL. The computational framework was divided into two moduls, macro-strategy and micro-strategy. Macro-strategy control the learning path based on learning map in AND-OR Graph data stucture. This paper focus on the Macro-strategy modul, that using online, direct, and centralized adaptivity method. The adaptivity in game has five components as its target. Based on those targets, eight development models of SGfL concept was enumerated. With similarity and difference analysis toward possibility of united LS and GS in computational framework to implement the nine SGfL concept into design and application, there are three groups of the development models i.e. (1) better united GS and LS, (2) must manage LS and GS as different entity, and (3) can choose whether to be united or not. In the model which is united LS with GS, computing model at the macro-strategy modul use and-or graph and forward chaining. However, in the opposite case, macro-strategy requires two intelligent computing solutions, those are and-or graph with forward chaining to manage LS collaborated with Finite State Automata to manage GS. The proposed computational framework of SGfL was resulted from the similarity and difference analysis toward all possible representations of teaching materials into the adaptive components of the game. It was not dependent of type of learning domain and also of the game genre. Keywords: Serious Gamefor Learming, Learning State, Game State, And-Or Graph, FSA Copyright © 2015 Universitas Ahmad Dahlan. All rights reserved.

1. Introduction Entertainment game requires adaptivity for the game more fun and unpredictable [1]. While in the serious game for learning (SGfL), adaptivity is a necessity because of SGfL should be able to adjust to the progress of player skill proficiency and achievement of learning targets. This paper outlines a study of adaptivity in SGfL by utilizing theories, constructs, methods, techniques, tools, or other artifacts of adaptivity in the game, instructional design framework, adaptivity in a serious game itself. The study results manifested in the form of a flexible computing model to the variety of adaptive game component which represented the teaching material, so versatile also for the learning domain and game genres. On this study, ITS (intelligent tutoring system) will be used to evaluate the completeness of computational models SGfL features an intelligent learning system. This paper is a subset of the research on the development of concept models and design models of SGfL with this approach: the transformation of non-game instructional design into the game. There are three research are closely associated with this paper. The main one is the result of a survey by Lopez about the progress and the movement of research at adaptivity in the area of game [1]. Second is the basic theory from Reygeluth about instructional design framework, which has a micro-strategy and macro-strategy terminology in organizational strategy, which certainly have an impact on the delivery strategy and management strategy[3]. Third, the paper MDKickmeier-Rust, which proposes about adaptivity in a serious game, namely a non-invasive method of micro-adaptivity, within the meaning adaptivity towards learning does not interfere with the flow of game. The Rust's method applied in the case study on narative game based learning. The paper also touched on the need for macro-adaptivity that one of its

Received June 26, 2015; Revised August 13, 2015; Accepted September 2, 2015

1423 

ISSN: 1693-6930

functions is to manage the sequence of the curriculum. Macro-adaptivity is expected to be noninvasive as well [3]. Paper from MD KNickmeier-Rust will be used as the primary means of comparison with SGfL models proposed in this paper. Lopez classify diversity of research results about adaptivity in the game based on the purpose, method and the targeted game component was adapted. This article discuss about SGfL flexible computing models for the 5 different game components as adaptivity target. Studied adaptation method in this paper was limited to the online method with direct adaptivity and centralized mechanism, and variable input only from player skill proficiency aspect. Online adaptivity means that adaptation carried out during running game, controlled by the data about learning progress of player. Direct adaptivity means that the rules for decision-making and choice of actions that can be selected in the decision, has been prepared by the game designer before running the game. Centralized mechanism means that all decision and action for adaptation are done and controlled by one module, not distributed to some independent agents. The scope of adaptation the learning task in macro-strategy is control the learning path based on map of competencies. Among competency have a prerequisite relationship. Macrostrategy ensure that a competency can only be studied if its prerequisite has been mastered. Variety adaptivity that must be provided in macro-strategy of SGfL are (1) to intervene when the player in the stuck, where no game state can be explored (3) lower/raise the minimum the threshold criteria for mastery of competencies based on the trends of the player's ability to learn (4) encourage the player to repeat the game for achieve a higher level of mastery of the competencies. If the adaptivity model that be proposed by MDKickmeier-Rust is non-invasive [3] [9], this research contrary want to examine how to integrate learning with components, flow and logic of the game. The reason is "because SGfL must be adaptive based on the learning progress of the player, and the primary object to be adapted are teaching materials and delivery technique, so the teaching material should be represented in a game component that becomes the target of adaptivity". Research question of the studies reviewed in this paper is how the invasive patterns of the representation of the learning material into 5 different components of the game, and then to found flexible computing model for adaptivity in serious game for learning (SGfL). Reuse component for varian of implementation context is key component of flexible computing model. Domain analysis is a method to find the reuse component, using similarity and difference analysis in the domain problem. Research in this paper used FODA (Feature Oriented Domain Analysis) consist of context analysis: In order to establish scope, domain modeling: in order to define the problem space, architectural modeling, in order to characterize the solution space [4] [5]. Detail of these stage was described in research method. The solution space will be manifested in functional model and architecural model. 2. Research Method Picture 1 described detail step of research in this paper. Contex analysis give source material to further research. Based on literatur review, in this chapter will be explained about the material. The space of problem in the form of an enumeration of the variants development model of SGfL concept. The variants was developed based on representations of teaching materials on a variety of game component. The next step is analysis of similarity and difference of the functional features that required for computing the game engine. That is not on aspects of multimedia interaction. Including experiments to gain a firmer clarity about how to manage learning state and Game State the game will be described also in the research method. Characteristics of the solution is found in the form of features and computational models be written in result and analysis.

TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

TELKOMNIKA

ISSN: 1693-6930

 1424

Figure 1. Research Methodology 2.1. Context Analysis 2.1.1 Adaptivity in Game and the Adoption into SGfL Adaptivity architecture in game could be seen in Figure 2. Adaptation mechanism is listed below. 1. Monitoring player action 2. Interpret player action into variables in player modelling 3. Assign values into player mode 4. Predict Next State Experience using game state and information from player model 5. Construct game elements based on the Next State Experience Generally, adaptivity in serious game could be done by online or offline. Offline mechanism was done by survey approach to user when user login and before the game loaded, so the engine was called as “content generation”. Meanwhile online mechanism is done along the game based on data obtained from model player (driven approach data) so the engine fit with the name “content adaptation”. In online mechanism, rules and technique for decision making and types of decisions could be taken may use two approaches, which are direct adaptation (all things are prepared by designer) or indirect adaptation (using machine’s learning to find customized combination of action). To use indirect adaptation, need long enough learning process towards the system and a lot of data for automatic learning. Refer to the figure 2, there are 5 different components of the game that could be the target of adaptivity. If adaptivity is limited only by the player competence proficiency, not motivation or other mental conditions, then the 5 kinds of components it is an opportunity to represent the teaching materials. That rule was induced from fact that teaching materials and delivery mechanisms that will be the target of the adaptivity in SGfL. Map of learning embodied in the organization of competence in AND-OR GRAPH. Control over the students' learning pathways are the same as controlling the position of students in the map of learning Set of different status in the game where the player must have been in one of the declared status is the game states space. When this architecture is adopted into SGfL, position of the players can be viewed from two angles, the play map (game state), and the angle of

Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

1425 

ISSN: 1693-6930

learning (learning state). It can be presumed that in SGfL should be managed game state (GS) space and learning state (LS) space. How space management for the GS and LS, if it is associated with the representation of teaching materials in five option components of the game that be target of the adaptivity is a research question that becomes the main subject of this paper.

Figure 2. Architecture of Adaptivity in Game – summarized from [1]-[6-10] 2.1.2 Scope of Macrostrategy in Instructional Design Reigeluth defines framework of instructional design consisted of three elements, which are Condition, Method, and Outcome [2]. The purpose of design is to set the right method that suite the condition existed and the outcome expected. There are four aspect of Condition, i.e.: learning content, learner context, learning context, and specific requirement. Method consists of organizational strategy, delivery strategy, and management strategy. Organizational strategy divided into two levels, which are macro strategy and micro strategy. Macro strategy manages “what do I want the Student to learn” and “what did I know about the Student”. Practically, it is organizing learning contents, what the students have to learn to achieve the learning outcome, how to sort, conclude, or synthesize them. Delivery strategy in macro level introduces learning activity and controls Micro strategy module. Management Strategy make decisions towards which contents of learning delivered in what context at every T (time) in learning process. Management strategy needs knowledge about 1) condition and progress of student’s learning, 2) map of material’s organization, context, and interactions, and 3) strategy to match student’s condition with material. Subjects learned are managed as a set of learning state (LS). LS is a cross of competency levels (Bloom/Anderson taxonomy) in one domain and knowledge object (see table 1). LS, one with another, have a relation between, which is usually made in prerequisite form. One LS could contain one or more knowledge object. To measure student’s ability towards an LS, will be needed a standard definition of player’s mastery over an LS. Learning method used for an LS also need to be defined. Details of subjects in an LS will be managed in Micro strategy module. Referred to [11], in the microstrategy was proposed two kind of activity. Those activities are learning activity and assessment activity. In the learning activity, SGfL will provide varian of support for learning based on player capability to learn the new competencies. Judgment about player’s skill proficiency was taken from assessment activity only. So the player is not judged by the length of the learning process. TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

 1426

ISSN: 1693-6930

TELKOMNIKA

Table 1. Learning State was build from cognitive level x Knowledge Object Cognitif Level

Knowledge Object Q R Q1 R1 Q2 R2 Q3 R3

P P1 P2 P3

Remembering Understanding Apply

S S1 S2 S3

Example of learning map in Figure 3, show that learning target consisted of three competences. Competence K1 is a set of competences in node P1, node P2, node S1, node S2, node P3, and node S3. Competence K2 is a set of competences in node P1, node Q1, node S1, node Q2, node T1, node T2, and node T3. Meanwhile competence K3 is a set of competences in node P1, node Q1, node Q2, node R1, node R2, node Q3, and node R3. LS Q1 needs mastery of LS P1 to be played. Management strategy take responsibility to drive the learning path based on the learning map. Example in figure 3, the path only be started from P1, because there is no prerequisite of P1. Learning state P2, Q1, and S1 can be learned after player have mastered P1. Learning state S2 was opened for player if and only if they have mastered P2 and S1, that is the meaning of the two arrows into the S2 is united with curved lines. Criteria of the mastering of competencies is determined by the assessment result of the player skill proficiency in a learning state.

P3

S2

P2

P1

S3 S1 Q2

Q1

T2

T1

R1

T3

Q3 R2 R3

Figure 3. Relationship Prerequisite between LS Learning Policy Adaptation in Macro Level Strategy Along learning process, learners’ ability is different one to another. For provide adaptivity, the threshold value could be differ for each LS, depends on its difficulty level. If it sets an absolute score for every learner in every LS, then some learner did not experience learning process in a few LS, because they can't fulfill threshold value of prerequisite LS. To solve this, learning designer can apply three types of policy or a combination of them. Below are the three policies.give the learners chance to try again in failed LS, with maximum limit of chances after he get game over state. (1) lower treshold for learners who have signs of having less ability, so those learners could experience the next LS but with degraded quality of challenges. SGfL can provide 3 kinds grade, low, medium, and high for the treshold. (2) to learners who have succes through the whole LS required, but not with their optimal result (high treshold), will be given chance to repeat again By repeating, learners are expected to master the LS better. Order of LS opened can be changed to prevent learners from getting bored. It is better for nonlinierity aspect if SGfL have many alternative material resource. 2.1.3. Rule Analysis For Combine 5 Game Compoment as Target of Adaptivity Based on Figure 3, there are five types of components that have possibilities to be adaptive. The order and combinations cannot be done freely, however. It has to refer to components’ relations in game design aspects. Combination proposed in Figure 4 is based on details below. a) Definition of game components according to Richard Rouse [12] and Dave Morris [13] Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

1427 

ISSN: 1693-6930

b) Three types of order in arranging game concept according to [12], which are, i) Gameplay  technology  Story; ii) Technology  story  gameplay; and iii) Story  gameplay  technology c) Framework of gameplay developments and game mechanic from Carlo Fabricatore [14] A few proposition obtained are listed below. a) Quest /Challenge/Puzzle is core of interaction between game with player b) Gameplay, Game Mechanic, and NPC can not define separately  (Gameplay X Game Mechanic X PC/NPC) c) Quest is weak entity toward story or toward (Gameplay X Game Mechanic X NPC) d) Story can be followed by Gameplay or Gameplay be followed by Story e) Gameworld must be relevant with plot. Plot is even in that all of game component collaborate to make story.  changes in the gameworld make impact on the change in the majority of components. f) At player story or emergent story, story was created by Gameplay X Game Mechanic X PC/NPC.  (Gameplay X Game Mechanic X PC/NPC) X Story

Figure 4. Diagram of Combination Rule for Adaptif Component in Game 2.2. Domain Modelling 2.2.1. Enumeration Space of Problem with Rule In this section will be displayed eight kinds of development models of the concept of serious games (see figure 5), based on rule at figure 4 and using the terminology of Carlo Fabricatore and game development framework gameplay mechanic that is concerned with learnability. For gameplay, starting from the core gameplay is facilitated by a core mechanic or more Game play can be enriched with core metagameplay without changing the core mechanic. Core mechanic can be enriched with satelite mechanic, in the form of enhancement or powerup, or alternate mechanic. Peripheral gameplay may be used if the story forced to introduce a new gameplay to the player. Quest is the essence of the interaction of the game to create a challenge. In the quest to be represented SGfL teaching materials. Forms will vary based component quest game where teaching materials are represented. The simplest is a puzzle. The most complicated is if the representation of the material in the form gameplay x mechanic x item/NPC.

TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

 1428

ISSN: 1693-6930

TELKOMNIKA

1. Quest as Adaptive Component, Story  GamePlay  Quest

Knowledge Object  Inspired by Story‐ Support the Quest was presented in  Story  mechanic (follow  item/NPC (follow  Learning  quest (inspired by  State (embbeded) gameplay  story) gameplay) gameplay) Story ‐1 gameplay (core+meta) Core Mechanic [1,2] (quest as item/NPC) P1 q1 Story ‐1+ gameplay (core+meta) Core Mechanic [1,2] (quest as item/NPC) P2 q2 q3 Story ‐1++ gameplay (core+meta) Core Mechanic [1,2,3,4] (quest as item/NPC) P3

Gameworld (context of Story  gameplay, mechanic,item/NPC,  quest) gameworld A gameworld B gameworld C

2. Quest as Adaptive Component, GamePlay  Story  Quest Can be Fixed Component mechanic (follow  gameplay  gameplay) gameplay (core+meta) Core Mechanic [1,2] gameplay (core+meta) Core Mechanic [1,2] gameplay (core+meta) Core Mechanic [1,2,3,4]

item/NPC (follow  gameplay) (quest as item/NPC) (quest as item/NPC) (quest as item/NPC)

Learning  State P1 P2 P3

Knowledge Object  Gameworld  was presented in  Story (context  (present the  quest of quest) story)  q1 Story‐1 gameworld A q2 Story‐2 gameworld B q3 Story‐3 gameworld C

3. Story as adaptive component, Story Gameplay  Quest Knowledge Object  Learning was presented in  Follow the Story State Gameplay Game Mechanic Item/NPC Quest  Story P1 Story‐1 Coregameplay A +meta Core Mechanic [1,2] enemy E1 q1 P2 Story‐2 Coregameplay  A +meta Core Mechanic [1,2] enemy E2, Allied  A1 q2

Gameworld follow  story , gameplay,  mechanic, item Gameworld A Gameworld B

4. Gameplay x Mechanic x Item/Npc as adaptive component , Gameplay Story  Quest Learning  State Gameplay P1 Coregameplay [1] P2 Coregameplay [1,2] P3 Coregameplay [1,2]

Knowledge Object was presented in Game Mechanic

Item/NPC

Core Mechanic [1,2]

enemy E1, Allied A1, A2, A3

Core Mechanic [1,2]

enemy E2, Allied  A1

Core Mechanic [1,2]+PowerUp enemy E1, Allied A1, A2, A3

Story, Inspired by  Ques t follow the  Ga mepl ay,Mecha nic,  Story, ga mepla y,  Item/NPC mechanic,item/NPC

Story‐1 Story‐2 Story‐3

q1 q2 q3

Gameworld  foll ow s tory  and others   element

Gameworld A Gameworld A Gameworld B

5. Quest as adaptive component, Gameplay  Quest ( Player Story) Knowledge  Object was  presented in  quest

Can be Fixed Component gameplay  mechanic (follow gameplay) gameplay (core+meta)Core Mechanic [1,2] gameplay (core+meta)Core Mechanic [1,2] gameplay (core+meta)Core Mechanic [1,2,3,4]

Learning  item/NPC (follow gameplay) State

(quest as subset of  item/NPCP1 (quest as subset of  item/NPCP2 (quest as subset of  item/NPCP3

q1 q2 q3

Gameworld (context  of quest, gameplay,  mechanic,item/NPC)

gameworld A gameworld B gameworld C

6. Gameplay x Mechanic x Item/Npc as adaptive component, gameplay Quest (Player Story) Learning State P1 P2 P3

Gameworld  Quest follow  follow the  gameplay,  gameplay,  mechanic,item/N Item/NPC enemy E1, Allied A1, A2, A q1 Gameworld A enemy E2, Allied  A1 q2 Gameworld A enemy E1, Allied A1, A2, A q3 Gameworld A

Knowledge Object was presented in (create story) Gameplay Game Mechanic Coregameplay [1,2] Core Mechanic [1,2] Coregameplay [1,2] Core Mechanic [1,2] Core meta gameplay [1,2]+Core Mechanic [1,2,3,4]

7. Gameworld as adaptive component Gameworld as LS, Gameworld  Gameplay  Story  Quest Inspired by Game World Knowledge  Learning Object /  GameMechanic Item/NPC State Competence  Gameplay P1 Gameworld X CoreGamePlay Core Mechanic Enemy X, Allied Y, Item Z P2 Gameworld A Peripheral GamePlay A Core Mechanic A Enemy A, Allied A, Item A

Story  Story‐X Story‐A

Quest Follow  Story & gameplay q1 q2

8. Gameworld as adaptive component, Gameworld + Gameplay as LS , Gameworld  Gameplay  Story  Quest Learning  State P1 P2 P3 P4

Knowledge  Object /  Gameworld X Gameworld X Gameworld A Gameworld A

Gameplay CoreGamePlay CoreGamePlay CoreGamePlay +meta Peripheral GamePlay A

Inspired by Game World GameMechanic Item/NPC Core Mechanic [1,2] Enemy X, Allied Y, Item Z Core Mechanic [1,2,3,4] Enemy X, Allied Y, Item Z Core Mechanic [1,2,3,4] Enemy X, Allied Y, Item Z core Mechanic A [1,2] Enemy A, Allied A, Item A

Story  Story‐X Story‐X Story‐X Story‐A

Quest Follow  Story & gameplay q1 q2 q3 q4

Figure 5. Enumerasi of Problem Space: 10 kinds of Development Models SGfL Concept 2.2.2. Similarity and Difference Analysis Of each enumeration can be affirmed, they have the same problem, ie managing learning state as an entity that represents a competency. Characteristics of learning state in macrostrategy level of organizational and management aspects of the strategy outlined in the previous section. The impact of differences in the components used to represent the course material in the game, led to a difference in managing game state (GS) space, whether it can be combined with learning the state or not in the scope of Macrostrategy. Possibility analysis of an integrated managementtoward GS and LS for each enumeration can be seen in tabel 2.

Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

1429 

ISSN: 1693-6930

Table 2. Learning State and Game State Analysis in Development Model of the Game Concept Nu

1 1.a

quest based

2 1.b

quest based

How to find  game concept learning state in macrostrategy story‐gameplay‐ quest class of quest gameplay‐story‐ quest class of quest

3 2.b

story based

story‐gameplay‐ story id (tree) or node in branch of  quest tree of story

4 3.b

gameplay  based

gameplay‐story‐ C(gameplay,mechanic,item/npc) =  quest usually treated as level of game

5 4.a

quest based  +story

gameplay‐quest  create story class of quest

6 4.b

gameplay  based+story

gameplay‐quest  C(gameplay,mechanic,item/npc) =  create story usually treated as level of game

7 5.a

gameworld  based

non hirarchi

gameworld = usually treated as  level of game

8 5.b

gameworld  based

hirarchical  learning state

C(gameworld,gameplay,mechanic,  item/npc) =usually treated as level 

Id

Model

Possibility gamestate           =   game state in macrosrategy learning state fchoose (mechanic, quest as item))   or fchoose (class of quest) possible fchoose (mechanic, quest as item))   or fchoose (class of quest) possible fchoose(mechanics  to choose  branch of story tree) or  possible, for choice of story  fChooce(story id/tree) id more suitable to integrate   GS = LS, as choice  or decided  the levels of game by system impossible , because  fchoose(mechanics, quest as  mechanic must have  item)) for create story functional story impossible , because  fchoose(mechanics, level))  for  mechanic must have  create story functional story more suitable to integrate   GS = LS, as choice  or decided  the levels of game by system more suitable to integrate   GS = LS, as choice  or decided  the levels of game by system

2.2.3. Problem Solving Modelling of Control Learning Path on Learning State Space The structure of the representation of the learning state space in Figure 3, in artificial intelligence terminology known as AND-OR Graph. Its construction consists of a node, directed arc (in / out), and the relation between the arcs-in on a node. Some arc in which adjoined by a curved line on a node declared the relationship AND. Some arc-in on a node that no curved lines express the relation OR [15]. Nodes that do not have the arc-in, referred to as a fact, which does not have the arc-out called goal, and who have both called subgoal. At the practical level computing, AND-OR Graph represented in sentence calculus propositions in the form of a special clause horn. Examples AND-OR graph representation in figure 3 into the calculus of propositions can be seen in the Figure 6 At LS space, a node represents a competence that should be mastered by player. Arc stated prerequisite relationships. Node with a bow in need of competence from its pair node. Control strategy of the learning path starts from the node that does not have the prerequisite (Fact), discover nodes that all prerequisites met (sub Goal), to the node that indicates a complete learning outcomes studied (goal). The target from study is the overall goal node or a part of it. There are two kinds of agoritma to build inference engine for AND-OR graph, ie forward chaining and backward chaining. The algorithm in accordance with the management strategy of macro-modul is a forward chaining. Figure 6 describing the forward chaining algorithm applied to the game "Save The KOD Kingdom". 2.2.4. The Experiment: Build Prototype SGfL about Learning SQL Experiments conducted by building two prototype game with different genres, to support learning on the same topic, ie SQL. Adaptivity target component is a quest. Game version 1 implements enumeration 1 (seet figure 5) which design of the game mechanic at macro-strategy make LS united with GS. The second version implement enumeration number 2, which game mechanic have narrative function, so GS did not unite with LS. Game version 1, where LS = GS, is “SAVE THE KOD KINGDOM”. It was adventure game, embedded story, with only one game mechanic in macro level, that is player choose area where the next quest is waiting to be done. Goal of this SGfL is to collect points from solving problems in a kingdom. Character of the player here is a princess, Ruruna, who gets a sudden duty from his father to rule the kingdom. The story of this game could be seen at figure 8.

TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

TELKOMNIKA

ISSN: 1693-6930

 1430

1. If P1 then P2 Infered: contentedall proposition symbol with related data TabClausa : contentedd norm relational form if IF-THEN clauses 2. If P2 then S1 3. If P2 thenS2 TabCount : contented clausa id and number of premises 4. If S2thenP3 Agenda : stack for computation in Forward chaining 5. If P3 thenT2 Result : contented proposition symbol was known as true 6. If Q1thenQ2 TabTracking : contented proposition have been processed and 7. If Q1thenR1 become true 8. If Q2 AND R1 then R2 Begin 9. If Q2 thenQ3 Input (Tabclauses) 10. If S1 thenQ3 Input (tabInfered) 11. If R2 thenR3 Create(TabCount) 12. If Q3thenT3 Push( all Fakta, Agenda) 13. If R3AND Q3 thenT1 GameOver=False While agenda.notempty()= false and gameover==false do P = Agenda.pop() If Infered[P].value = false Then { the proposition symbol still false} {display property atribut of P} {ask for user, input assesment result of P} IFassresult>= P.treshold Then Push(TabTracking,P); Infered[P].value = True ; ClausesMatch= Select * From Tabclauses Where Left=P.symbol, For i=1 to end ClausesMatch Do Decrement (TabCount[clausesMatch.Norule]. Count) If TabCount[clausesMatch.Norule].Currentcount == 0 then If TabCount[clausesMatch.Norule].Symbol is goal Then {ask for user, input assesment result for the Goal symbol} IF assResult>= Goal.treshold THEN {save the symbol to Result} If Symbol tersebut GoldenGoal then Gameover=true { winner, finish game withexcellcene} endIF ENDIF Else Push (TabCount[clausesMatch.Norule].Symbol) Urutkan agenda berdasarkan nilai heuristic Endif TabCount[clausesMatch.Norule].statusExecute=True Endif endFor {update the count matrix} Endif {display “ try another option”} Endif {proposition symbol have been processed-Skip} Endwhile

Figure 6. Proposition Calculus from Figure 3 and Forward Chaining Algorithm It could be seen in Figure 7, there were two “!”signs. They were provided for the player to choose them. Engine of game that unit learning state and game state controls computation behind these sign choices that appeared on the game interface. Capture of computation process could be seen in Figure 8. There, P1 is already done. P2 and Q1 appeared on the interface as the choices of quest which the player would do next. The other learning states are not yet to be opened. Game version 2 is ALTERCITY. Its genre is career simulation; with game mechanic are actions that relevant with carrier management and daily life. Quest was put as an item, which is done by player in player’s job and training while player was building his carrier. The goal is to collect wealth, to get the highest position in the most prestigious company in a town called Alter city, and to get a prospective couples. Character of player is an informatics technology graduate named Mada who starts his career in Altercity. There are two kinds of learning state, working state and training state. Training state represented “know” and “understand” competence level. Working state represent “apply” competence.

Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

1431 

ISSN: 1693-6930

Figure 7. Illustration of Game Interface version 1- player choose sign!

Figure 8. Capture of Engine Implementation for Control Learning State Figure 9 shows the interface of Altercity. From top-left to bottom-right: MADA boarding to altercity, accompanied by allied Pak Eza, which gives clues about what to do in Altercity. The next is global gameworld of Altercity, first residential for MADA, job announcements, Mada goes to Altermart, and Mada met with HRD of ALTERMART to apply for jobs.The contents of jobs drawn from the results of the execution of the FSA toward the And-Or Graph of LS

Figure 9. Game Interface at ALTERCITY TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

 1432

ISSN: 1693-6930

TELKOMNIKA

Learning state, game state and game mechanic for Altercity can be seen in the FSA in Figure 10 and tabel 3. Game mechanic with parameter wState and tState represent learning activity.

Figure 10. Game State and FSA of Altercity Table 3. Learning State in Altercity Concept Training  Level Basic Junior Intermed Expert

Star‐1 T1‐1 T1‐2 T1‐3 T1‐4

Star‐2 T2‐1 T2‐2 T2‐3 T2‐4

Star‐3 T3‐1 T3‐2 T3‐3 T3‐4

Star‐4 T4‐1 T4‐2 T4‐3 T4‐4

Cognitif Level Know, understand Know, understand Know, understand Know, understand

Concept Star‐1 Mini  Working market Trainee P11 Staff P12 Superviso P13 Analyst P14

Star‐1 Star‐1 Sport  Klinik center P21 P31 P22 P32 P23 P33 P24 P34

Star‐2 Star‐2 Star‐3 District  Super  Bank Office market Q11 Q21 R11 Q12 Q22 R12 Q13 Q23 R13 Q14 Q24 R14

Star‐3Star‐4 Hos pital Bank R21 S11 R22 S12 R23 S13 R24 S14

Star‐4 Town  Office S21 S22 S23 S24

Cognitif  Level Apply Apply Apply Apply

3. Results and Analysis 3.1. The Commonality of Computing Model Similarity and difference analysis and the experiment support the identification of functional features required in the computing engine of SGfL, not on multimedia computing to gaming interface. These features are classified based on management pattern GS and LS state, bounded on Macro strategy module only. The result can be seen in the figure 11 and table 4. There are three groups of issues, (1) The group recommended for unite the GS with LS space, that is, if course material represented in the gameworld or (gameplayxmechanic x item / NPC with designer story). (2) The group that can choose to use a model of integration or separation between GS and LS (3) The groups that should management of GS and LS, as different entity, that is, if the representation of course material in (a) quest or in (b) (gameplay x mechanic x item /NPC) as the level, and the player story. Game designers have to design gameplay as story builder

Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

1433 

ISSN: 1693-6930 1. Manage Learning State (LS) and AND-OR Graph of LS 2. Manage Game State –GS (gameplay x mechanics x item/Npc) in FSA model 3. a Manage Item /NPC, Decoratif, Dialog component to support story and game balance 3.b Manage relation LS and others game component 3.c Manage relation GS and others game component 4.a Manage player model (student model) 4.b Manage player model (gaming variabel) 4.c Manage player model (GS history) 5. Setting and execute Learning Policy as adaptivity in macrolevel 6. Control Learning Path based ( and-or Graph of LS wih Forward Chaining) 7. Control Game State with FSA enginer 8. Execute Game balance 9. Inerfacing with game user interface

Figure 11. Result of Similarity and Difference analysis Table 4. Result of Similarity and Difference analysis ID Fea-01.0.1 Fea-01.0.2 Fea-02.0.1 Fea-03.a.1 Fea-03.b.1 Fea-03.c.1 Fea-04.a.1 Fea-04.b.1 Fea-04.c.1 Fea-08.0.1 Fea-05.0.1 Fea-06.0.1 Fea-07.b.1 Fea-09.0.1 Fea-10.0.1 Fea-10.0.2

Type Data Organizing (CRUDE)

Intellligence Computing Multimedia Programming Information Processing & dashboard

Object Learning State (LS) Prerequisite Relation among LS Game State and rules of its production NPC, Item with behaviour, interactivity, value,dialog,and decorative component to support story, gameplay and gamebalance Game State’srelation with other game elements Learning State’srelation with other game elements Manage Player Model ( Student Model) Manage Player Model (gaming Variabel) Manage Player model (game state history) Execution Game Balance for Gaming Variabel Execute Adaptivity in Learning Policy based on student model Control Learning State Control Game State- FSA Interfacing with UI of game aplication for sensory immersive implementation View Player Model View LS Performance in supporting learning efectivity

Attribute Mandatory Mandatory Mandatory Mandatory Substitutive Fea-3.c.1 Substitutive Fea-3.b.1 Mandatory Mandatory Mandatory for GS#LS Mandatory Mandatory Mandatory Mandatory for GS#LS mandatory Addtional Additional

The experiment has been developed instance two game concepts, core game design, game interface prototype development, and implementation of the core function of computational models needed to macrostrategy. Especially for prototype version 1, "Save the KOD Kingdom", also has made a prototype for Microstrategy. Lesson learned is obtained 1. Confirming the hypothesis that the SGfL, need to be managed LS space in addition to the GS space. However, in several of the concept of the game, both can be united, so that computing becomes more simple. 2. Proving the model representations And-Or Graph, calculus propositions in the form of horn clauses appropriate to implement organizational strategy at the macro level. Forward chaining algorithm modification according to implement a management strategy at the macro level. The solution can re-use for both kinds of LS and GS space management 3. Exemplifying the mechanism of how to interact with Gameplay Mechanic LS using Finite State Automata as a computational model In the experiments conducted there are shortcomings, ie the components supporting the game balance game, story, items, and smart NPC, as well as the gameworld is still modest. For prototype version 1, gameworld represented in the form of comic and implemented as a series of images that display before or after quests of an LS. At Altercity, 3D elements of the room, item, and the allied NPC with their dialogue with the player character has not been put into computing. These need further research and experiment. Using multiple smart NPC Allied / Enemy very interesting for further study, from computing aspect and from learning method aspect.

TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

TELKOMNIKA

ISSN: 1693-6930

 1434

3.2. Functional Model The functional model will describe the implementation from functional features into the process and data as well as the interaction between processes. Figure 12 describe the functional model for two kinds of computational models as a solution of the two kinds of cases, ie for the case (a) LS integrated with GS and (b) a separate LS and interact with GS. Model B requires the addition of a process number 6 and number 7 in the model b compared models a, additional data store for the player models that save game state history. There is a change in behavior on the number 1, in the model (a) related to the game componet. In the model (a), the main control in the second game, while on the model (b) is in the process 7 3.4. Architectural Model Based on architecture of adapativity in game that proposed by Lopez [1], this paper proposed architecure adaptivity in SGfL at figure 13. In the Architecture have been put maping functional feature of ITS (italic font) and aspect in framework instructional desain (bold font). All of the feature and the aspect from ITS and Insructional Design framework was included in the map. Because the ITS feature is used as one of the evaluators about the completeness of features an intelligent system to support learning, then from the SGfL architecture can be concluded that SGfL feature has been intelligent enough to adapt to the needs of the player.

Figure 12. Functional Model of SGfL’s computation where LS=GS (left) and LS # GS (right)

Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

1435 

ISSN: 1693-6930

Figure 13. Architecturall Analysis a) Left: GS = LS b) right: GS separated with LS The proposed architecture model of game as in figure 12, add one entity, that was learning state where the SGfL design, manage the game state as different entity with learning state. Learning state was used to represent learning content at macro level strategy meanwhile game state was a current state of a player as subset of eligible states provided in game. Generally game state is managed by Finite State automata. And-Or Graph was suitable data structure for represent learning state. The structure can be solved with reasoning approach using rule based system, which knowledge base and inference engine are are mandatory component. Propositional calculus in clause horn is simple and powerful for implemented the and-or graph in knowledge base. Forward chaining was suitable for implement management strategy at macrostrategy. 4. Conclusion Some of the conclusions that want to be affirmed are as follows: 1. SGfL must manage learning state (LS) space beside game state (GS) space. The issue would be simpler if the GS can be integrated with the LS space, thus managing only one state in the game. But apparently not in all cases both the state space can be integrated. 2. How to manage LS and GS in SGfL on macro-strategy module, depend on game component where game designer represent the learning material and learning task beside depend on gameplay design. 3. To support good adaptivity in SGfL, LS must be represented in adaptive game component. Based on 5 different game component as target of adaptivity, have been developed 10 kinds of development model of SGfL Concept. 4. Based on possibility analysis to integrate Gs and LS, from 10 kinds development model of SGfL Concept, there were 3 groups. Those are recommended (4), optional (4), and impossible (2). If learning material was presented in quest and use player story approach then LS must separated from GS. the same is true if the learning material presented in gameplay mechanic x item /NPC and uses the player story concept. 5. Experiment in this research affirmed that a. where GSwas integrated with LS, control on game can be handled by rule based system, clausa horn as knowledge representation format and forward chaining b. Where GS was separated with LS, control feature in macrostrategy modul need collaboration of Finite State Automata and rule based system with forward chaining.

TELKOMNIKA Vol. 13, No. 4, December 2015 : 1422 – 1436

TELKOMNIKA

ISSN: 1693-6930

 1436

6. Research on this SGfL need to be continued on the domain analysis of the micro-strategy. Does solution model in macro-strategycan be implemented in micro-strategy? What customization was need for the problem space in micro-strategy. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15]

Lopez Ricardo, Bidarra Rafael. Adaptivity Challenges in Games and Simulation : A Survey. IEEE Transaction on Computational Intelligence and AI in Games. 2011; 3(2). Reigeluth C. Instructional Design: Theory and Model Second Edition. Lawrence Erlbaum Associates, 1983 Kickmeier-Rust MD, Albert. Micro-adaptivity: protecting immersion in didactically adaptive digital educational games. Journal of Computer Assisted Learning. 2010; 26: 95–105. Jatain Aman. Comparison of Domain Analysis Methods and their supporting Quality Attributes. Thesis. Patiala, Thapar University. 2009. Hirave Kanifnath S, Hanchate Dinesh Bhagwan. Survey on Ex Feature: A Feature Modelling and Recommending Technique For Domain Oriented Product Listing. International Journal of Engineering Research and General Science. 2014; 2(6). Westra Josst et.al. Guiding User Adaptation in Serious Games. Agent for Games and SimulationII, LNAI 6525. 2011; pp. 117-131, @ Springer-Verlag Berlin Heidelberg. Westra Joss et.al. Keep the trainee on track. IEEE conference Computing Intelligent Game, Denmark. 2010. Lee Seong Jae, Liu Yun End, Popovic Zoran. Learning Individual Behaviour in Educational Game: a Data Driven Approach. The 7th International Conference on Educational Data Mining (EDM ). 2014. Koidl Kevin, Mehm Kevin et.al, Dynamically adjusting Digital Educational Games Toward Learning Objectives, The 3rd European Conference on Game Based Learning, 2010; 177-184. B Magerko, C Heeter, et.al, Intelligent Adaptation of Digital gamae-based Learning, ACM Conference Future Play, Canada, 2008. Minovic Miroslav, et.al. Visualisation of Student Learning Model in Serious Games. Computers in Human behaviour. 2014. Richard Rouse. Game Design: Theory & Practice Second Edition. Wordware Publishing, Inc, 2005. Rollings A, Morris D. Game Architecture and Desisgn: A New Edition. Indianapolis, New Riders Publishing. 2004. Fabricatore Carlo, Gameplay and Game Mechanics Design: A Key to Quality in Videogames. OECDCERI Expert Meeting on Videogames and Education. 2007. Russel Stuart, Norvig, Peter. Artificial Intelligence a Modern Approach. Prentice Hall, 2003.

Computing Game and Learning State in Serious Game for Learning (Ririn Dwi Agustin)

Chapter 2 REVIEW OF PREVIOUS WORK

2.1 Introduction Automatic Speech Recognition (ASR) is a critical core technology in the field of intelligence communication between human and machine. Despite the long history of research on the acoustic characteristics of Vowel / Consonant-Vowel (V/CV) unit waveforms [Sakai and Doshita(1963)], [Schaffer and Rabiner(1970)], [D. Dutta Majumder and Pal(1976)], [Broad(1972)], current state-of-the-art ASR systems are still incapable of performing accurate recognition for these class of sounds . Beginning 1910, Campell and Crandall from AT & T and Western Electric Engineering initiated a series of experiments to explore the nature of human speech perception. After this in 1918, these experiments were continued by Fletcher and his colleagues at The Bell Telephone Laboratories (Western Electric Engineering until 1925). These studies lead to a speech recognition measure called articulation index, which accurately characterizes speech intelligibly under condition of filtering and noise. All these experiments began with normal conversational speech over a modified telephone channel [Fletcher(1922)] [Fletcher and Munson(1937)] [Fletcher and Galt(1950)]. In the 1930’s Homer

14

Dudley, influenced greatly by Fletcher’s research, developed a speech synthesizer called the VODER (Voice Operating Demonstrator), which was an electrical equivalent (with mechanical control) of Wheatstone’s mechanical speaking machine [H. Dudley and Watkins(1939)]. These two speech poineers thoroughly established the importance of the signal spectrum for reliable identification of the phonetic nature of a speech sound. In 1940’s Dudley had developed mathematical models for speech based on linguistic research that viewed spoken language with the impulses from larynx and the vocal folds as the input to the system, the shape of the vocal tract representing the filter parameter and the speech waveform as the system output [Dudley(1940)].

In 1952,Davis et al., of Bell Laboratories built a system for isolated digit recognition for a single speaker , using the formant frequencies measured (or estimated) from vowel regions of each digit. These trajectories served as the reference pattern for determining the identity of an unknown digit utterance as the best matching digit [K. H. Davis and Balashek(1952)]. In the 1960’s, several Japanese laboratories established their capability of building special purpose hardware to achieve a speech recognition task. Most important were the vowel recognizer of Suzuki and Nakata at the Radio Research Lab in Tokyo [Suzuki and Nakata(1961)], the phoneme recognizer of Sakai and Doshita at Kyoto [Sakai and Doshita(1962)], and the digit recognizer of NEC Laboratories [K. Nagata and Chiba(1963)]. The work of Sakai and Doshita involved the first use of a speech segmenter for analysis and recognition of speech in different portions of the input utterance.

15

In another early recognition system Fry and Denes, at University College in England, built a phoneme recognizer to recognize 4 vowels and 9 consonants [Fry and Denes(1959)]. Including statistical information about allowable phoneme sequences in English, they increased the overall phoneme recognition accuracy for words consisting of two or more phonemes. This work marked the first use of statistical syntax (at the phoneme level) in automatic speech recognition. An alternative to the use of a speech segmenter was the concept of adopting a nonuniform time scale for aligning speech patterns. This concept started to gain acceptance in the 1960’s through the work of Tom Martin at RCA Laboratories [T. B. Martin and Zadell(1964)] and Vintsyuk in the Soviet Union [Vintsyuk(1968)]. Martin recognized the need to deal with the temporal non-uniformity in repeated speech events and suggested a range of solutions, including detection of utterance endpoints, which greatly enhanced the reliability of the recognizer performance. Vintsyuk proposed the use of dynamic programming for time alignment between two utterances in order to derive a meaningful assessment of their similarity [Vintsyuk(1968)]. His work, though largely unknown in the West, appears to have preceded that of Sakoe and Chiba as well as others who proposed more formal methods [Sakoe and Chiba(1978)], generally known as dynamic time warping, in speech pattern matching. Since the late 1970’s,mainly due to the publication by Sakoe and Chiba, dynamic programming, in numerous variant forms (including the Viterbi algorithm which came from the communication theory community), has become an indispensable technique in automatic speech recognition [Viterbi(1967)].

16

In the late 1960’s, Atal and Itakura independently formulated the fundamental concepts of Linear Predictive Coding (LPC) , which greatly simplified the estimation of the vocal tract response from speech waveforms [Atal and Hanauer(1971)] [Itakura and Saito(1970)]. By the mid 1970’s, the basic ideas of applying fundamental pattern recognition technology to speech recognition, based on LPC methods, were proposed by Itakura [Itakura(1975)], Rabiner and Levinson and others [L. R. Rabiner and Wilpon(1979)]. In the early 1980’s at Bell Laboratories, the theory of HMM was extended to mixture densities which have since proven vitally important in ensuring satisfactory recognition accuracy, particularly for speaker independent, large vocabulary speech recognition tasks [Juang(1985)] [B. H. Juang and Sondhi(1986)] . Another technology that was (re)introduced in the late 1980’s was the idea of artificial neural networks (ANN). Neural networks were first introduced in the 1950’s, but failed to produce notable results initially [McCullough and Pitts(1943)]. The advent, in the 1980’s, of a parallel distributed processing (PDP) model, which was a dense interconnection of simple computational elements, and a corresponding training method, called error backpropagation, revived interest around the old idea of mimicking the human neural processing mechanism [Lippmann(1987)] [Kohonen(1988)] [Pal and Mitra(1988)].

In the 1990’s, a number of inventions took place in the field of pattern recognition. The problem of pattern recognition, which traditionally followed the framework of Bayes and required estimation of distributions for the data, was transformed into an optimization problem involving minimization of the empirical recognition error [Juang(1985)]. This fundamental change of paradigm was caused by the recognition of the fact that the distribution functions for the speech 17

signal could not be accurately chosen or defined, and that Baye’s decision theory would become inapplicable under these circumstances. After all, the objective of a recognizer design should be to achieve the least recognition error rather than the best fitting of a distribution function to the given (known) data set as advocated by the Bayes criterion. The concept of minimum classification or empirical error subsequently spawned a number of techniques, among which discriminative training and kernel-based methods such as the support vector machines (SVM) have become popular subjects of study [B.H. Juang and Chou(1997)] [Vapnik(1998)].

This chapter presents a review of previous works in the area of linear and nonlinear speech processing, multi resolution analysis and wavelet transform, neural networks and statistical learning algorithms and is organized as follows. Section 2.2, section 2.3 and section 2.4 provides a summary of research findings in the area of traditional speech processing, wavelet transform and nonlinear speech processing respectively. Section 2.5 gives a review of previous works in the applications of neural network for speech recognition and section 2.6 contains review of previous work in the application of k-Nearest Neighborhood (k-NN) and SVM. Finally section 2.7 concludes this review.

2.2 Review on Traditional Features for Speech Recognition Linear Predictive Coding (LPC) and Mel Frequency Cepstral Coefficient (MFCC) features are known to be traditional basic speech features in the sense that the speech model used in many of these applications is the source-filter model which represent the vocal characteristics of the speech signal. The linear prediction (LP)

18

model for speech analysis and synthesis was first introduced by Saito and Itakura and Atal and Schroeder [Itakura and Saito(1970)] [Atal and Schroeder(1967)]. Saito and Itakura at NTT, Japan, developed a statistical approach for the estimation speech spectral density using a maximum likelihood method [Itakura and Saito(1970)]. Their work was originally presented at conferences in Japan and therefore, was not known worldwide. The theoretical analysis behind their statistical approach were slightly different than that of linear prediction, but the overall results were identical. Based on their statistical approach, Itakura and Saito introduced new speech parameters such as the partial autocorrelation (PARCOR) coefficients for efficient encoding of linear prediction coefficients. Later, Itakura discovered the line spectrum pairs, which are now widely used in speech coding applications.

In 1975, John Makhoul presented a tutorial review on Linear Predictive Coding [Makhoul(1975)], which gives an exposition of linear prediction in the analysis of discrete signals. The major part of this paper is devoted to all-pole models. The model parameters are obtained by a least squares analysis in the time domain. Two methods resulted, depending on whether the signal is assumed to be stationary or non stationary. The same results are then derived in the frequency domain also. The resulting spectral matching formulation allows for the modeling of selected portions of a spectrum for arbitrary spectral shaping in the frequency domain, and for the modeling of continuous as well as discrete spectra.

The method of linear prediction has proved quite popular and successful for use in speech compression system [Markel and A.H.Gray(1974)] [Itakura(1972)] 19

[Atal and Hanauer(1971)]. An efficient method for transmitting the linear prediction parameters has been found by Sambur using the techniques of differential PCM [Sambur(1975)]. Using this technique, speech transmission is employing fewer than 1500 bits/second. Further reduction in the linear prediction storage requirements can be realized at a cost of higher system complexity by transmission of the most significant eigenvectors of the parameters. It has been found that this technique in combination with differential PCM can lower the bitrate to 1000 bits/sec. Sambur and Jayant discusses several manipulations of LPC parameters for providing speech encryption [Sambur and Jayant(1976)]. They considers temporal rearrangement or scrambling of the LPC code sequence, as well as the alternative of perturbing individual samples in the sequence by means of pseudorandom additive or multiplicative noise. The latter approach is believed to have greater encryption potential than the temporal scrambling technique, in terms of time needed to break the security code. The encryption technique are assessed on the basis of perceptual experiments, as well as by means of qualitative assessment of speech spectrum distortion, as given by an appropriate distance measure.

2.3 Review on Wavelet Transform for Speech recognition Over the last decades, wavelet analysis have become very popular and new interests are emerging in this topic. It has turned to be a standard technique in the ‘area of geophysics, meteorology, audio signal processing and image compression [Hongyu Liao and Cockburn(2004)], [Soman and Ramachandran(2005)], [Mallat(2009)]. Wavelet Transform is a tool for Multi Resolution Analysis which can be used to efficiently represent the speech signal in the time-frequency plane. 20

Martin Vetterli [Vetterli(1992)] had compared the wavelet transform with the more classical short-time Fourier transform approach to signal analysis. In addition he also pointed out the strong similarities between the details of these techniques. Gianpaolo [Evangelista(1993)] explored a new wavelet representation using the transform based on a pitch-synchronous vector representation and its adaptation to the oscillatory or aperiodic characteristics of signals. Pseudo-periodic signals are represented in terms of an asymptotically periodic trend and aperiodic fluctuations at several scales. The transform reverts to the ordinary wavelet transform over totally aperiodic signal segments. The pitch-synchronous wavelet transform is particularly suitable to the analysis, rate-reduction in coding and synthesis of speech signals and it may serve as a preprocessing block in automatic speech recognition systems. Separation of voice from noise in voiced consonants is easily performed by means of partial wavelet expansions.

In [Xia and Zhang(1993)], the authors studied the properties of Cardinal Orthogonal Scaling Functions (COSF), which provide the standard sampling theorem in multiresolution spaces with scaling functions as interpolants. They presented a family of COSF with exponential decay, which are generalizations of the Haar functions. With these COSF, an application is the computation of Wavelet Series Transform (WST) coefficients of a signal by the Mallat algorithm. They also presented some numerical comparisons for different scaling functions to illustrate the advantage of COSF. For signals which are not in multiresolution spaces, they estimated the aliasing error in the sampling theorem by using uniform samples.

21

In [M Lang and R.O.Wells(1996)] Lang et al., had presented a new nonlinear noise reduction method using discrete wavelet transform.They employed thresholding in the wavelet transform domain following a suggestion by Coifman, using undecimated, shift-invariant, nonorthogonal wavelet transform instead of the usual orthogonal one. This approach can be interpreted as a repeated application of the original Donoho and Johnstone method for different shifts. The main feature of this algorithm is a significantly improved noise reduction compared to the original wavelet based approach. This holds for a large class of signals, and is shown theoretically as well as by experimental results.

A multi wavelet design criterion known as omnidirectional balancing using wavelet transform is introduced by James E. Fowler and Li Hua to extend to vector transforms the balancing philosophy previously proposed for multiwavelet based scalar-signal expansion [Fowler and Hua(2002)]. It is shown that the straightforward implementation of a vector wavelet transform, namely, the application of a scalar transform to each vector component independently, is a special case of an omnidirectionally balanced vector wavelet transform in which filter-coefficient matrices are constrained to be diagonal. Additionally, a family of symmetricantisymmetric multiwavelets is designed according to the omnidirectional balancing criterion. In empirical results for a vector-field compression system, it is observed that the performance of vector wavelet transforms derived from these omnidirectionally balanced symmetric-antisymmetric multiwavelets is far superior to that of transforms implemented via other multiwavelets and can exceed that of diagonal transforms derived from popular scalar wavelets.

22

O Farooq and S Dutta [Farooq and S.Datta(2004)]had proposed a subband feature extraction technique based on an admissible wavelet transform and the features are modified to make them robust to Additive White Gaussian Noise (AWGN). The performance of this system is compared with the conventional mel frequency cepstral coefficients (MFCC) under various signal to noise ratios. The recognition performance based on the eight sub-band features is found to be superior under the noisy conditions compared with MFCC features using this approach.

Elawakdy et al., proposed a speech recognition algorithm using wavelet transform. This paper discussed the combination of a feature extraction by wavelet transform, subtractive clustering and adaptive neuro-fuzzy inference system (ANFIS). The feature extraction is used as input of the subtractive clustering to put the data in a group of clusters. Also it is used as an input of the neural network in ANFIS. The initial fuzzy inference system is trained by the neural network to obtain the least possible error between the desired output (target) and the fuzzy inference system (FIS) output to get the final FIS. The performance of the proposed speech recognition algorithm (SRA) using a wavelet transform and ANFIS is evaluated by different samples of speech signals of isolated words with added background noise. The proposed speech recognition algorithm is tested using different isolated words obtaining a recognition ratio about 99%.

A multi resolution hidden markov model using class specific features is proposed by Baggenstoss [Baggenstoss(2010)]. He applied the PDF projection theorem to generalize the hidden Markov model (HMM) to accommodate multiple 23

simultaneous segmentations of the raw data and multiple feature extraction transformations. Different segment sizes and feature transformations are assigned to each state. The algorithm averages over all allowable segmentations by mapping the segmentations to a "proxy" HMM and using the forward procedure. A by-product of the algorithm is the set of a posteriori state probability estimates that serve as a description of the input data. These probabilities have simultaneously both the temporal resolution of the smallest processing windows and the processing gain and frequency resolution of the largest processing windows. The method is demonstrated on the problem of precisely modeling the consonant "T" in order to detect the presence of a distinct "burst" component. He compared the algorithm against standard speech analysis methods using data from the TIMIT Speech Database.

In short, from the literature survey it is shown that there is an emerging research trends in the study of application of Multi Resolution Analysis using Wavelet Transform for the human speech recognition for the fast few years. Thus Malayalam V/CV speech unit recognition using MRA based Wavelet Transform is of great importance to capture the non-stationary nature of the speech signal.

2.4 Review on Non-Linear Dynamical System Models for Speech Recognition Nonlinear speech processing is a rapidly growing area of research. Naturally, it is difficult to define a precise date for the origin of the field, but it is clear that there was a rapid growth in this area, which started in the mid-nineteen eighties. 24

Since that time, numerous techniques were introduced for nonlinear time series analysis, which are ultimately aimed at engineering applications.

Among the nonlinear dynamics community, a budding interest has emerged in the application of theoretical results to experimental time series data analysis in 1980’s. One of the profound results established in chaos theory is the celebrated Takens’ embedding theorem. Takens’ theorem states that under certain assumptions, phase space of a dynamical system can be reconstructed through the use of time-delayed versions of the original scalar measurements. This new state space is commonly referred to in the literature as Reconstructed State Space (RSS), and has been proven to be topologically equivalent to the original state space of the dynamical system.

Packard et al., first proposed the concept of phase space reconstruction in 1980 [Packard.N.H and Shaw.R.S(1980)]. Soon after, Takens showed that a delay coordinate mapping from a generic state space to a space of higher dimension preserves topology [Takens(1980)]. Sauer and Yorke have modified Taken’s theorem to apply for experimental time series data analysis [Sauer.T and Casdagli.M(1991)].

Conventional linear digital signal processing techniques often utilize the frequency domain as the primary processing space, which is obtained through the Discrete Fourier Transform (DFT) of a time series. For a linear dynamical system, representation of the signal appears in the frequency domain that takes the form of sharp resonant peaks in the spectrum. However for a nonlinear or chaotic system, the signal representation does not appear in the frequency domain, because the 25

spectrum is usually broadband and resembles noise. In the RPS, a signal representation emerges in the form of complex, dense orbits that form patterns known as attractors. These attractors contain the information about the time evolution of the system, which means that features derived from a RPS can potentially contain more or different information.

The majority of literature that utilizes a RSS for signal processing applications revolves around its use for control, prediction, and noise reduction, reporting both positive and negative results. There is only scattered research using RPS features for classification and /or recognition experiments.

In contrast to the linear source-filter model for speech production process, a large number of research works are reported in the literature to show the nonlinear effects in the physical process. Koizumi.T, Taniguchi.S et al., in 1985 showed that the vocal tract and the vocal folds do not function independently of each other, but that, there is in fact some form of coupling between them when the glottis is open [Koizumi.T and Hiromitsu.S(1985)]. This can cause significant changes in formant characteristics between open and closed glottis cycles [Brookes.D.M and Naylor.P.A(1988)].

Teager and Teager [Teager.S.M(1989)] have claimed that voiced sounds are characterised by highly complex airflows in the vocal tract, rather than well behaved laminar flow. Turbulent flow of this nature is also accepted to occur during unvoiced speech, where the generation of sound is due to a constriction at some point in the vocal tract. In addition, the vocal folds will themselves be responsible 26

for further nonlinear behaviour, since the muscle and cartilage, which comprise the larynx, have nonlinear stretching qualities [Fletcher and Munson(1937)].

Such non-linearities are routinely included in attempts to model the physical process of vocal fold vibration, which have focused on two or more mass models [Fletcher and Galt(1950)], [H. Dudley and Watkins(1939)], [Dudley(1940)] in which the movement of the vocal folds is modeled by masses connected by springs, with nonlinear coupling. Observations of the glottal waveform reinforce this evidence, where it has been shown that this waveform can change shape at different amplitudes [Schoentgen(1990)]. Such a change would not be possible in a strictly linear system where the waveform shape is unaffected by amplitude changes.

Extraction of invariant parameters from speech signal has attracted researchers for designing speech and speaker recognition systems. In 1988, Narayanan.N.K. and Sridhar C.S. [Narayanan.N.K and Sridhar.C.S(1988)] used the dynamical system technique mentioned in the nonlinear dynamics to extract invariant parameters from speech signal. The dynamics of speech signal is experimentally investigated by extracting the second order dimension of the attractor D2 and the second order Kolmogorov entropy K2 of speech signal. The fractal dimension of D2 and non-zero value of K2 confirms the contribution of deterministic chaos to the behavior of speech signal. The attractor dimension D2 and Kolmogorov entropy K2 are then used as a powerful tool for voiced / unvoiced classification of speech signals.

27

The dimension of the trajectories, or the dimension of the attractor is an important characteristic of the dynamic systems. The estimation of the dimension gives a lower bound of the number of parameters needed in order to model the system. The goal is to find if the system under study occupies all the state space or if it is most of the time in a subset of the space, called attractor. The correlation dimension [Tishby.N(1990)] is a practical method to estimate the dimension of an empirical temporal series.

There are a large variety of techniques found in the literature of nonlinear methods and it is difficult to predict which techniques ultimately will be more successful in speech processing. However, commonly observed methods in the speech processing literature are various forms of oscillators and nonlinear predictors, the latter being part of the more general class of nonlinear autoregressive methods. The oscillator and autoregressive techniques themselves are also closely related since a nonlinear autoregressive model in its synthesis form, forms a nonlinear oscillator if no input is applied. For the practical design of a nonlinear autoregressive model, various approximations have been proposed [Farmer.J.D and Sidorowich.J.D(1988)] [Casdagli.M and Gibson.J(1991)] [Abarbanel.H.D.I and Tsimring.L.S(1993)] [Kubin.G(1995)]. These can be split into two main categories: parametric and non parametric methods.

Phase space reconstruction is usually the first step in the analysis of dynamical systems. An experimenter obtains a scalar time series from one observable of a multidimensional system. State-space reconstruction is then needed for the indirect measurement of the system’s invariant parameters like, dimension, Lyapunov 28

exponent etc. Takens’ theorem gives little guidance, about practical considerations for reconstructing a good state space. It is silent on the choice of time delay (τ) to use in constructing m-dimensional data vectors. Indeed, it allows any time delay as long as one has an infinite amount of infinitely accurate data. However, for reconstructing state spaces from real-world, finite, noisy data, it gives no direction [Casdagli.M and Gibson.J(1991)]. Two heuristics have been developed in the literature for establishing a time lag [Kantz and Schreiber.T(2003)]. First one is the first zero of the autocorrelation function and the second one is the first minimum of the auto mutual information curve [Fraser.A.M and Swinney.H.L(1986)]. Andrew M Fraser and Harry L Swinney reported in their work that the mutual information is examined for a model dynamical system and for chaotic data from an experiment on the Belousov-Zhabotinskii reaction. An N log N algorithm for calculating mutual information (I) is presented. A minimum in ’I’ is found to be a good criterion for the choice of time delay in Phase Space Reconstruction from time series data. This criterion is shown to be far superior than choosing a zero of the autocorrelation function.

There have been many discussions on how to determine the optimal embedding dimension from a scalar time series based on Taken’s theorem or its extensions [Sauer.T and Casdagli.M(1991)]. Among different geometrical criteria, the most popular seems to be the method of False Nearest Neighbors. This criterion concerns the fundamental condition of no self-intersections of the reconstructed attractor [Kennel.M.B and Abarbanel.H.D.I(1992)].

29

Work by Banbrook, McLaughlin et al., [Banbrook and McLaughlin(1994)] Kumar et al.,

[Kumar.A and Mullick.S.K(1996)]

and Narayanan et al.,

[Narayanan.S.S and Alwan.A.A(1995)] has attempted to use nonlinear dynamical methods to answer the question: "Is speech chaotic?" These papers focused on calculating theoretical quantities such as Lyapunov exponents and Correlation dimension. Their results are largely inconclusive and even contradictory. A synthesis technique for voiced sounds is developed by Banbrook et al., inspired by the technique for estimating the Lyapunov exponents.

In a work presented by Langi and Kinsner [Langi.A and Kinsner.W(1995)], speech consonants are characterised by using a fractal model for speech recognition systems . Characterization of consonants has been a difficult problem because consonant waveforms may be indistinguishable in time or frequency domain. The approach views consonant waveforms as coming from a turbulent constriction in a human speech production system, and thus exhibiting turbulent and noise like time domain appearance. However, it departs from the usual approach by modeling consonant excitation using chaotic dynamical systems capable of generating turbulent and noise-like excitations. The scheme employs correlation fractal dimension and Takens embedding theorem to measure fractal dimension from time series observation of the dynamical systems. It uses linear predictive coding (LPC) excitation of twenty-two consonant waveforms as the time series. Furthermore, the correlation fractal dimension is calculated using a fast Grassberger algorithm [Grassberger and Procaccia(1983)].

30

The criterion in the False Nearest Neighbor approach for determining optimal embedding dimension is subjective in some sense that, different values of parameters may lead to different results [Cao.L(1997)]. He proposed in his work a practical method to determine the minimum embedding dimension from a scalar time series. It does not contain any subjective parameters except for the time delay for the embedding. It does not strongly depend on how many data points are available and it is computationally efficient. Several time series are tested to show the above advantages of the method. For real time series data, different optimal embedding dimensions are obtained for different values of the threshold value. Also with noisy data this method gives spurious results [Kantz and Schreiber.T(2003)].

Narayanan presented an algorithm for voiced/unvoiced speech signal classification using second order attractor dimension and second order kolmogorov entropy of the speech signals. The non-linear dynamics of the speech signal is experimentally analyzed using this approach. The proposed techniques were further used as a powerful tool for the classification of voiced/unvoiced speech signals in many applications [Narayanan(1999)].

In [N K Narayanan and Sasindran(2000)] Narayanan et al., investigated on the applications of phase space map and phase space point distribution parameter for the recognition of Malayalam vowel units. The presented features were extracted by utilizing the non-linear/chaotic signal processing techniques. Andrew et al., presented phase space feature for the classification of TIMIT corpus and demonstrated that the proposed technique outperform compared with frequency domain based MFCC feature parameters [Andre C Lingren and Povinelli(2003)]. 31

Petry et al.,

[A Petry and Barone.C(2002)] and Pitsikalis et al.,

[Pitsikalis.V and Maragos.P(2003)] have used Lyapunov exponents and Correlation dimension in unison with traditional features (cepstral coefficients) and have shown minor improvements over baseline speech recognition systems. Central to both sets of these papers is the importance of Lyapunov exponents and Correlation dimension, because they are invariant metrics that are the same regardless of initial conditions in both the original and reconstructed phase space. Despite their significance, there are several issues that exist in the measuring of these quantities on real experimental data. The most important issue is that these measurements are very sensitive to noise. Secondarily, the automatic computation of these quantities through a numerical algorithm is not well established and this can lead to drastically differing results. The overall performance of these quantities as salient features remains an open research question.

In [P Prajith and Narayanan(2004)], P Prajith et al., proposed a feature parameter by utilizing nonlinear or chaotic signal processing techniques to extract time domain based phase space features.Two sets of experiments are presented. In the first, exploiting the theoretical results derived in nonlinear dynamics, a processing space called phase space is generated and a recognition parameter called Phase Space Point Distribution (PSPD) is extracted. In the second experiment Phase Space Map at a phase angle p/2 is reconstructed and PSPD is calculated. The output of a neural network with error back propagation algorithm demonstrate that phase space features contain substantial discriminatory power.

32

Kevin M Lindrebo et al., introduced a method for calculating speech features from third-order statistics of sub band filtered speech signals which are used for robust speech recognition [Kevin M. Indrebo(2005)]. These features have the potential to capture nonlinear information not represented by cepstral coefficients. Also, because the features presented in this method are based on the third-order moments, they may be more immune to Gaussian noise than cepstrals, as Gaussian distributions have zero third-order moments.

Richard J Povinelli et al., introduced a novel approach to the analysis and classification of time series signals using statistical models of reconstructed phase spaces [Povinelli.R.J(2006)]. With sufficient dimension, such reconstructed phase spaces are, with probability one, guaranteed to be topologically equivalent to the state dynamics of the generating system, and, therefore, may contain information that is absent in analysis and classification methods rooted in linear assumptions. Parametric and non parametric distributions are introduced as statistical representations over the multidimensional reconstructed phase space, with classification accomplished through methods such as Bayes maximum likelihood and artificial neural networks (ANNs). The technique is demonstrated on heart arrhythmia classification and speech recognition. This new approach is shown to be a viable and effective alternative to traditional signal classification approaches, particularly for signals with strong nonlinear characteristics.

In [Prajith and Narayanan(2006)] P Prajith and N K Narayanan had introduced a flexible algorithm for pitch calculation by utilizing the methodologies developed for analyzing chaotic time series. The experimental result showed that the pitch 33

estimated using reconstructed phase space feature agrees with that obtained using conventional pitch detection algorithm.

Marcos Faundez-Zanuy compared the identification rates of a speaker recognition system using several parameterizations, with special emphasis on the residual signal obtained from linear and nonlinear predictive analysis [Zanuy(2007)]. It is found that the residual signal is still useful even when using a high dimensional linear predictive analysis. If instead of using the residual signal of a linear analysis a nonlinear analysis is used, both combined signals are more uncorrelated and although the discriminating power of the nonlinear residual signal is lower, the combined scheme outperforms the linear one for several analysis orders.

P Prajith introduced in his thesis the applications of nonlinear dynamical theory. As an alternate to traditional model of speech production a nonlinear system has been proposed. The problem of whether speech (especially vowel sounds) is chaotic has been examined through discussion of previous studies and experiments. Nonlinear invariant parameters for Malayalam vowels are calculated. The major invariant features include attractor dimensions and Kolmogorov entropy. The non-integer attractor dimension and non-zero value of Kolmogorov entropy confirm the contribution of deterministic chaos to the behavior of speech signal [Prajith(2008)].

In a recent study Kar proposed a novel criterion for the global asymptotic stability of fixed-point state-space digital filters under various combinations of quantization and overflow non linearities [Kar(2011)]. Yucel Ozbek et al., pro34

posed a systematic framework for accurate estimation of articulatory trajectories from acoustic data based on multiple-model dynamic systems via state-space representation [Yucel Ozbek and Demirekler(2012)]. The acoustic measurements and articulatory positions are considered as observable (measurement) and hidden (state) quantities of the system, respectively. To improve the performance of state space-based articulatory inversion they have used jump-Markov linear system (JMLS). Comparison of the performance of their method with the reported ones given in the literature shows that the proposed method improves the performance of the state-space approaches.

It is seen that the majority of literature that utilizes the nonlinear techniques for signal processing applications revolves around its use for control, prediction and noise reduction, reporting both positive and negative results. There is only scattered research using these methods for classification or recognition experiments. It is also important to notice that very less works are reported yet in nonlinear speech processing for Malayalam and no such work has been reported in other Indian languages. The succeeding session of this chapter is focused on the review of the applications of artificial neural network for speech recognition.

2.5 Review on Applications of ANN for Speech Recognition Artificial neural net (ANN) algorithms have been designed and implemented for speech pattern recognition by a number of researchers. ANNs are of interest because algorithms used in many speech recognizers can be implemented using highly parallel neural net architectures and also because new parallel algorithms

35

are being developed making use of the newly acquired knowledge of the working of biological nervous systems. Hutton.L.V compares neural network and statistical pattern comparison method for pattern recognition purpose [Hutton.L.V(1992)]. Neural network approaches to pattern classification problems complement and compete with statistical approaches. Each approach has unique strengths that can be exploited in the design and evaluation of classifier systems. Classical (statistical) techniques can be used to evaluate the performance of neural net classifiers, which often outperform them. Neural net classifiers may have advantages even when their ultimate performance on a training set can be shown to be no better than the classical. It is possible to be implemented in real time using special purpose hardware.

Personnaz L et al., presents an elementary introduction to networks of formal neurons [Personnaz.L and Dreyfus.G(1990)]. The state of the art regarding basic research and the applications are presented in this work. First, the most usual models of formal neurons are described, together with the most currently used network architectures: static (feedforward) nets and dynamic (feedback) nets. Secondly, the main potential applications of neural networks are reviewed: pattern recognition (vision, speech), signal processing and automatic control. Finally, the main achievements (simulation software, simulation machines, integrated circuits) are presented.

Willian Huang et al., presents some neural net approaches for the problem of static pattern classification and time alignment. For static pattern classification multi layer perceptron classifiers trained with back propagation can form arbitrary 36

decision regions, are robust, and are trained rapidly for convex decision regions. For time alignment, the Viterbi net is a neural net implementation of the Viterbi decoder used very effectively in recognition systems based on Hidden Markov Models (HMMs) [William Huang and Gold(1988)].

Waibel.A et al., [Weibel.A and Lang.K(1988)] proposed a time delay neural network (TDNN) approach to phoneme recognition, which is characterized by two important properties. Using a three level arrangement of simple computing units, it can represent arbitrary non-linear decision surface. The TDNN learns these decision surfaces automatically using error back propagation. The time delay arrangement enables the network to discover acoustic phonetic features and temporal relationships between them independent of position in time and hence not blurred by temporal shifts in the input. For comparison, several discrete Hidden Markov Models (HMM) were trained to perform the same task, i.e. the speaker dependent recognition of the phonemes "B", "D" and "G" extracted from varying phonetic contexts. The TDNN achieved a recognition rate of 98.5% correct compared to 93.7% for the best of HMMs. They showed that the TDNN has well known acoustic-phonetic features (e.g., F2-rise, F2-fall, vowel-onset) as useful abstractions. It also developed alternate internal representations to link different acoustic realizations to the same concept.

Yoshua Bengio and Renato De Mori used The Boltzmann machine algorithm and the error back propagation algorithm to learn to recognize the place of articulation of vowels(front, center or back), represented by a static description of spectral lines [Bengio and Mori(1988)]. The error rate is shown to depend on the 37

coding. Results are comparable or better than those obtained by them on the same data using hidden Markov Models. They also show a fault tolerant property of the neural nets, i.e. that the error on the test set increases slowly and gradually when an increasing number of nodes fail.

Mah. R.S.H and Chakravarthy.V examined the key features of simple networks and their application to pattern recognition [Mah.R.S.H and Chakravarthy.V(1992)]. Beginning with a three-layer back propagation network, the authors examine the mechanisms of pattern classification. They relate the number of input, output and hidden nodes to the problem features and parameters. In particular, each hidden neuron corresponds to a discriminant in the input space. They point out that the interactions between number of discriminants, the size and distribution of the training set, and numerical magnitudes make it very difficult to provide precise guidelines. They found that the shape of the threshold function plays a major role in both pattern recognition, and quantitative prediction and interpolation. Tuning the sharpness parameter could have a significant effect on neural network performance. This feature is currently under-utilized in many applications. For some applications linear discriminant is a poor choice.

Janssen et al., developed a phonetic front-end for speaker-independent recognition of continuous letter strings [Janssen.R.D.T and Cole.R.A(1991)]. A feedforward neutral network is trained to classify 3 msec speech frames as one of the 30 phonemes in the English alphabet. Phonetic context is used in two ways: first, by providing spectral and waveform information before and after the frame to be classified, and second, by a second-pass network that uses both acoustic features 38

and the phonetic outputs of the first-pass network. This use of context reduced the error rate by 50%. The effectiveness of the DFT and the more compact PLP (perceptual linear predictive) analysis is compared, and several other features, such as zerocrossing rate, are investigated. A frame-based phonetic classification performance of 75.7% was achieved.

Ki-Seok-Kim and Hee-Yeung-Hwang present the result of the study on the speech recognition of Korean phonemes using recurrent neural network models conducted by them [Ki-Seok-Kim and Hee-Yeung-Hwang(1991)]. The results of applying the recurrent multi layer perceptron model for learning temporal characteristics of speech phoneme recognition is presented. The test data consist of 144 vowel+consonant+vowel (V+CV) speech chains made up of 4 Korean monothongs and 9 Korean plosive consonants. The input parameters of the artificial neural network model used are the FFT coefficients, residual error and zero crossing rates. The baseline model showed a recognition rate of 91% for vowels and 71% for plosive consonants of one male speaker. The authors obtained better recognition rates from various other experiments compared to the existing multilayer perceptron model, thus showing the recurrent model to be better suited to speech recognition. The possibility of using the recurrent models for speech recognition was experimented upon by changing the configuration of this baseline model.

Ahn.R and Holmes.W.H propose a voiced / unvoiced / silence classification algorithm of speech using 2-stage neural networks with delayed decision input [Ahn.R and Holmes.W.H(1996)]. This feed forward neural network classifier is 39

capable of determining voiced, unvoiced and silence in the first stage and refining unvoiced and silence decisions in the second stage. Delayed decision from the previous frame’s classification along with preliminary decision by the first stage network, zero crossing rate and energy ratio enable the second stage to correct the mistakes made by the first stage in classifying unvoiced and silence frames. Comparisons with a single stage classifier demonstrate the necessity of two-stage classification techniques. It also shows that the proposed classifier performs excellently.

Sunilkumar and Narayanan investigated the potential use of zerocrossing based information of the signal for Malayalam vowel recognition. A vowel recognition system using artificial neural network is developed. The highest recognition accuracy obtained for normal speech is 90.62% [Sunilkumar(2002)], [R K Sunilkumar and Narayanan(2004)].

Dhananjaya et al., proposed a method for detecting speaker changes in a multi speaker speech signal [Dhananjaya.N and Yagnanarayana.B(2004)]. The statistical approach to a point phenomenon (speaker change) fails when the given conversation involves short speaker turns (< 5 sec duration). They used auto associative neural network (AANN) models to capture the characteristics of the excitation source that present in the linear prediction (LP) residue of speech signal. The AANN models are then used to detect the speaker changes.

In [P Prajith and Narayanan(2004)] P Prajith et al., presented the implementation of a neural network with error back propagation algorithm for the speech 40

recognition application with Phase Space Point Distribution as the input parameter. A method is suggested for speech recognition by utilizing nonlinear or chaotic signal processing techniques to extract time domain based phase space features. Two sets of experiments are presented in this paper. In the first, exploiting the theoretical results derived in nonlinear dynamics, a processing space called phase space is generated and a recognition parameter called Phase Space Point Distribution (PSPD) is extracted. In the second experiment Phase Space Map at a phase angle p/2 is reconstructed and PSPD is calculated. The output of a neural network with error back propagation algorithm demonstrate that phase space features contain substantial discriminatory power.

In [R K Sunilkumar and Narayanan(2004)] the speech signal is modeled using the zerocrossing base features of the signal. This feature is used for recognizing the Malayalam vowels and Consonant Vowel Units using Kolmogorov- Smirnov statistical test and multilayer feed forward artificial neural network. The average vowel recognition accuracy for single speaker using ANN base method is 92.62 % and for three female speaker is 91.48%. The average Consonant vowel recognition accuracy for single speaker is 73.8%. The advantage of this method is that the network shows better performance than the other conventional techniques and it takes less computation than the other conventional techniques of parameterization of speech signal like FFT, and Cepstral methods.

Xavier Domont et al., proposed a feed forward neural network for syllable recognition [Xavier Domont and Goerick(2007)]. The core of the recognition system is based on a hierarchical architecture initially developed for visual object 41

recognition. In this work, they showed that, given the similarities between the primary auditory and visual cortexes, such a system can successfully be used for speech recognition. Syllables are used as basic units for the recognition. Their spectrograms, computed using a Gammatone filter bank, are interpreted as images and subsequently feed into the neural network after a preprocessing step that enhances the formant frequencies and normalizes the length of the syllables.

P Prajith investigated in his work the application of Multi Layer Feed Forward Neural Network (MLFFNN) with error back propagation algorithm for the classification of Malayalam vowel units. To evaluate the credibility of the classifier he used reconstructed phase approach in combination with Mel Frequency Cepstral Coefficient(MFCC). An overall recognition accuracy of 96.24% is obtained from the simulation experiment and reported that a significant boost in recognition accuracy is obtained using ANN with the hybrid features [Prajith(2008)].

Anupkumar et al., [Anupkumar Paul and Kamal(2009)] studied Linear Predictive Coding Coefficients (LPCC) and Artificial Neural Network (ANN) for the recognition of Bangala speech. They presented the different neural network architecture design for the pattern at hand. It is concluded that neural networks having more hidden layers are able to solve the problems very easily. By comparing error curves and recognition accuracy of digits it is concluded that Multi Layer Perceptron with 5 layers is a more generic approach rather than Multi Layer Perceptron with 3 hidden layers.

42

Hanchate et al., investigated the application of Neural Networks with one hidden layer with sigmoid functions and the output layer with linear functions. There are 10 output neurons for all the networks while the numbers of hidden neurons vary from 10 to 70. The inputs of the network are the features of 4 selected frames with 256 samples per frame. Each frame is represented by 12 MFCC coefficients of the signal in the frame. A comparatively good recognition accuracy is obtained using this approach [D B Hanchate and Mourya(2010)].

In [Yong and Ting(2011)], Yong and Ting investigated the speaker independent vowel recognition for Malay children using the Time Delay Neural Network (TDNN). Due to less research done on the children speech recognition, the temporal structure of the children speech was not fully understood. Two hidden layers TDNN was proposed to discriminate 6 Malay vowels: /a/, /e/, /i/, /o/ and /u/. The speech database consisted of vowel sounds from 360 children speakers. Cepstral coefficient was normalized for the input of TDNN. The frame rate of the TDNN was tested with 10ms, 20ms, and 30ms. It was found out that the 30ms frame rate produced the highest vowel recognition accuracy with 81.92%. The TDNN also showed higher speech recognition rate compared to the previous studies that used Multilayer Perceptron.

The zerocrossing interval distribution of vowel speech signal is studied by Sunilkumar and Lajish [Sunilkumar and Lajish(2012)] using 5 Malayalam short vowel units. The classification of these sounds are carried out using multilayer feed forward artificial neural network. After analyzing the distribution patterns and the vowel recognition results, they reported that the zerocrossing interval dis43

tribution parameters can be effectively used for the speech phone classification and recognition. The noise adaptness of this parameter is also studied by adding additive white Gaussian noise at different signal to noise ratio. The computational complexity of the proposed technique is also less compared to the conventional spectral techniques which includes FFT and Cepstral methods, used in the parameterization of speech signal.

In [Battacharjee(2012)] Bhattacharjee discussed a novel technique for the recognition of Assamese phonemes using Recurrent Neural Network (RNN) based phoneme recognizer. A Multi-Layer Perceptron (MLP) has been used as phoneme segmenter for the segmentation of phonemes from isolated Assamese words. Two different RNN based approaches have been considered for recognition of the phonemes and their performances have been evaluated. MFCC has been used as the feature vector for both segmentation and recognition. With RNN based phoneme recognizer, a recognition accuracy of 91% has been achieved. The RNN based phoneme recognizer has been tested for speaker mismatch and channel mismatch conditions. It has been observed that the recognizer is robust to any unseen speaker. However, its performance degrades in channel mismatch condition. Cepstral Mean Normalization (CMN) has been used to overcome the problem of performance degradation effectively.

In this thesis the application of linear and non-linear dynamical system models and multi resolution analysis using wavelet transform features of V/CV speech units for the recognition using brain like computing algorithm namely Artificial Neural Networks is explored in detail. 44

2.6 Review on Statistical Learning Algorithms for Speech Recognition Support Vector Machines (SVM) are learning techniques that is considered as an effective method for general purpose pattern recognition because of its high generalization performance without the need of domain specific knowledge [Vapnik(1995)]. Intuitively, given a set of points belonging to two classes, a SVM finds a hyperplane that separates the largest possible fraction of points of the same class on the same side, while maximizing the distance from either class to the hyperplane. This is the optimal separating hyperplane which minimizes the risk of misclassifying not only the examples in the training set, but also the unseen example of the test set.

The main characteristics of SVM are that they minimize a formally proven upper bound on the generalization error. They work on high dimensional feature space by means of a dual formulation in terms of kernels. The prediction is based on hyperplanes in these feature spaces, which may correspond to quite involved classification criteria on the input data. The layer in the training data set can be handled by means of soft margins.

In a work done [Clarkson and Moreno(1997)] by Clarkson and Moreno, authors explores the issues involved in applying SVMs to phonetic classification as a first step to speech recognition. They presented results on several standard vowel and phonetic classification tasks and show better performance than Gaussian mixture classifiers. They also presented an analysis of the difficulties they

45

foresee in applying SVMs to continuous speech recognition problems. This paper represents a preliminary step in understanding the problems of applying SVMs to speech recognition.

As a preliminary analysis on speech signal analysis Anil K Jain et al., [Anil k Jain and Mao(2000)] presented a robust speech recognizer based on features obtained from the speech signal. The authors explored the issues involved in applying SVMs to phonetic classification. They presented results on several standard vowel and phonetic classification tasks and showed better performance than Gaussian mixture classifiers. They also present an analysis of the difficulties in applying SVMs to continuous speech recognition problems.

In a work presented by Aravindh Ganapathiraju et al., they addressed the use of a support vector machine as a classifier in a continuous speech recognition system. The technology has been successfully applied to two speech recognition tasks. A hybrid SVM/HMM system has been developed that uses SVMs to postprocess data generated by a conventional HMM system. The results obtained in the experiments clearly indicate the classification power of SVMs and affirm the use of SVMs for acoustic modeling. The oracle experiments reported in their work clearly show the potential of this hybrid system while highlighting the need for further research into the segmentation issue [Aravind Ganapathiraju and Picone(2004)].

In a work done by Tsang-Long et al., [Tsang-Long Pao and Li(2006)] SVM & NN classifiers and feature selection algorithm were used to classify five emotions from Mandarin emotional speech and compared their experimental results. The 46

overall experimental results reveal that the SVM classifier (84.2%) outperforms the NN classifier (80.8%) and detects anger perfectly, but confuses happiness with sadness, boredom and neutral. The NN classifier achieves better performance in recognizing sadness and neutral and differentiates happiness and boredom perfectly.

In [Jing Bai(2006)], to improve the learning and generalization ability of the machine-learning model, a new compound kernel that may pay attention to the similar degree between sample space and feature space is proposed. The author used the new compound kernel support vector machine to a speech recognition system for Chinese isolated words, non-specific person and middle glossary quantity, and compared the speech recognition results with the SVM using traditional kernels and RBF network. Experiments showed that the SVM performance with the new compound kernel is much better than traditional kernels and has higher recognition rates than ones of using RBF network in different SNRs, and is of shorter training time.

Sandhya Arora et al., [Sandhya Arora and Basu(2010)] discussed the characteristics of some classification methods that have been successfully applied to handwritten Devnagari character recognition and results of SVM and ANNs classification method, applied on Handwritten Devnagari characters. After preprocessing the character image, they extracted shadow features, chain code histogram features, view based features and longest run features. These features are then fed to Neural classifier and in Support Vector Machine for classification. In neural classifier, they explored three ways of combining decisions of four MLP’s, de47

signed for four different features.

In a work done by Zhuo-ming Chen et al., [Zhuo-ming Chen and tao Yao(2011)] they extracted a new feature(DWTMFC-CT) of the consonants by employing wavelet transformation, and explains that the difference of similar consonants can be described more accurately by this feature. The algorithm used for classification was multi-class fuzzy support vector machine(FSVM). In order to reduce the computation complexity caused by using the standard fuzzy support vector machine for multi-class classification, this paper propose an algorithm based on two stages. Experimental results shows that the proposed algorithm could get better classification results while reducing the training time greatly.

In [Ruben Solera-urena and de MariA(2012)], authors suggest the use of a weighted least squares (WLS) training procedure that facilitates the possibility of imposing a compact semiparametric model on the SVM, which results in a dramatic complexity reduction. Such a complexity reduction with respect to conventional SVM, which is between two and three orders of magnitude, allows the hybrid WLS-SVC/HMM system to perform real-time speech decoding on a connected-digit recognition task (Spanish database namely SpeechDat). The experimental evaluation of the proposed system shows encouraging performance levels in clean and noisy conditions.

In short, SVM have been widely used for speech recognition application for the last few years. It is because of its high generalization performance even without the domain specific knowledge. In this thesis application of non-linear and 48

wavelet based feature of V/CV speech units for recognition using SVM is studied in detail

2.7 Conclusion In summary, the current stage in the evalution of speech recognition research result from a combination of several elements, such as versatility of the database used, credibility of the different strategies of the feature selection, environmental conditions, the performance of different classifiers and their combinations etc. It is clear that no much well known attempts are reported towards in the recognition of Malayalam V/CV speech units and hence more research is needed to improve the recognition rates of V/CV units in both clean and noisy conditions. The studies performed in the following chapter of this thesis reveal that, the multi resolution analysis and non-linear dynamical system approach have a very good role in providing flexible information processing capability by devising methodologies and algorithms on a massively parallel system capable of handling infinite intra-class variations for representation and recognition of V/CV speech units.

49

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/263849718

Multi Lingual Speaker Identification on Foreign Languages using Artificial Neural Network Article · November 2012 DOI: 10.5120/9177-3584

READS

55

3 authors, including: Prateek Agrawal

Harjeet Kaur

Lovely Professional University

Lovely Professional University

16 PUBLICATIONS 16 CITATIONS

10 PUBLICATIONS 12 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, letting you access and read them immediately.

SEE PROFILE

Available from: Prateek Agrawal Retrieved on: 11 May 2016

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012

Multi Lingual Speaker Identification on Foreign Languages using Artificial Neural Network Prateek Agrawal, Harjeet Kaur, Gurpreet Kaur Lovely Professional University, Phagwara

ABSTRACT Based on the Back Propagation Algorithm, this paper portrait a method for speaker identification in multiple foreign languages. In order to identify speaker, the complete process goes through recording of the speech utterances of different speakers in multiple foreign languages, features extraction, data clustering and system training. In order to realize the purpose, a database has been prepared which contains one sentence in 8 different international languages i.e. Catalan, French, Finnish, Italian, Portuguese, Indonesian, Hindi, English spoken by 19 distinct speakers, both male and female, in each language. With total size of 760 speech utterances, the average performance of the system is 95.657%. Application of developed system is mainly used in speaker authentication in telephony security oriented applications where the normal conversations are of short durations and the tendency of the spokesperson is to switch language from one to another

economical and accepted systems for solving the problems of unauthorized use of computer and communications systems and multilevel access control [2]. This paper converges particularly on the problem of identification of spokesperson with simple isolated words spoken in eight different international languages as indicated in Figure 1. Speaker Recognition

Speaker Verification

Speaker Identification

Text Independant

Text Dependent

Keywords Artificial Neural Network (ANN), Back Propagation Algorithm (BPA), Cepstral Analysis, Multilingual Speaker Recognition, Power Spectral Density (PSD).

1. INTRODUCTION Different people can be differentiated on the basis of their speech because they all have different unique speech characteristics which can identify as well as distinguish one person from the other. These features are extracted from the speech utterances using different algorithms for these diverse features. From the signal processing point of view, speech can be characterized in terms of the signal carrying message information. The speech can be representated in the waveforms [1]. In order to build a truly multilingual acoustic model, a strong practical approach to multilingual speech recognition for multiple countries , where more than 200 internationally recognized languages are spoken across the countries, has to be followed. This multilingual model should then be adapted to the target language with the help of a language identification system. Based on the information extracted from the speech signal, it can have three different recognition systems itself: Speaker Recognition, Language Recognition and Speech Text Recognition [11, 12, 14]. Speaker recognition can be separated into two different phases - Speaker Verification and Speaker Identification. The purpose of a speaker verification system is to confirm whether an unidentified voice matches the voice of a speaker whose identity is being claimed. Speaker identification process includes the detection of an unidentified voice from a set of known voices. Speaker verification systems are primarily used in security access control whereas speaker identification systems are mainly used in criminal investigation [3]. Speech recognition encompasses a large number of complex applications such as speech driven consumer applications, speech commissioning in logistics, checking & recording in quality assurance work etc. Speaker identification and speaker verification are the most

Figure 1: Taxonomy of Speaker Recognition

1.1 Back Propagation Training One of the potential moves towards the solutions to the problem of speech recognition is ANN [6, 7]. Because of their ability to execute a eminent level of parallel computation, their exalted capability level of robustness and fault tolerance features, ANNs have advantages to handle both speech and speaker recognition. They can learn complex features from the data, due to the non-linear structure of artificial neuron [6, 8]. Different algorithms are being used for training purposes such as BPA, Radial Basis Function, Recurrent networks etc. In present piece of work, the BPA has been used to train the artificial neural network. The Recognition Component in present system is offered by Back Propagation (BP). The motivation behind using BPA is that when the data is increasing at high rates, the traditional approaches never compare results with the growing input size. The ANN performs well on all the available databases and is known to be an excellent tool for such problems. The back-propagation training algorithm is an iterative gradient algorithm designed to minimize the mean square error between the actual output of a multilayer feed-forward perceptron and the desired output. It requires continuous differentiable non-linearity [16].

2. Previous Work There are number of practical works have been presented where most existing Automatic Speaker Recognition (ASR) systems are used. Authors in [15] proposed the use of a multilayer perceptron (MLP), which is trained using the backpropagation technique to be the engine of an automated digit recognition system using voice. Example for inputting credit card numbers, phone numbers etc. In view of only practically existing speaker recognition systems, most of the practical applications are of the small vocabulary or isolated word type.

36

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012 Medium- and large vocabulary systems perform well in laboratories but not in real life [3]. Statistical methods like Hidden Markov Model (HMMs), Harmonic Product Spectrum (HPS) algorithm have been used for research multilingual speaker identification system [4]. ANNs have been used to identify the speakers in single language [3]. Text dependent Voice Recognition System [16] worked on eight regional languages containing 904 speech utterances to achieve efficiency of 95.354 % was presented. No research work has been carried out for speaker recognition with multiple foreign languages.

3. Features of Speech Speech features play an important and useful role to differentiate one speaker from another [13]. In speaker independent speech recognition, premium is placed on extracting features that are somewhat invariant to changes in the speaker [5]. In this work the basic features of speech like Linear Prediction Coefficient (LPC), Linear Prediction Cepstral Coefficients (LPCC), Average Power Spectral Density, Cepstrum Coefficient, Line Spectral Frequency have been extracted using MATLAB.

3.1 Linear Prediction Coefficients LPC determines the coefficients of a forward linear predictor by reducing the prediction error in the least squares sense. It has applications in filter design and speech coding. [a,g] = lpc(x,p) finds the coefficients of a pth-order linear predictor (FIR filter) that predicts the current value of the real-valued time series x based on past samples.

3.2 Linear Prediction Cepstral Coefficients (LPCC) Linear frequency cepstral coefficient is found by using Discrete Fourier Transformation (DFT) by formula

where i represents number of coefficients and k is then number of DFT.

3.3 Average power spectral density This is another important feature of speech signal. PSD is intended for continuous spectra. The integral of the PSD over a given frequency band computes the average power in the signal over that frequency band. In contrast to the meansquared spectrum, the peaks in the spectra do not reflect the power at a given frequency. Average PSD of any signal can be calculated as the ratio of total power to the frequency of the signal. [17]

3.4 Cepstrum Coefficients The cepstral coefficients are the coefficients of the Fourier transform representation of the logarithm magnitude spectrum of the most distinctive feature that helps to differentiate the speakers.. Consider a sequence, x(n), having a Fourier transform X(ω). The cepstrum, cx(n), is defined by the inverse Fourier transform of Cx(ω), where Cx(ω) = logeX (ω). [18]

3.5 Line Spectral Frequency The LPC to LSF/LSP Conversion block takes a vector or matrix of linear prediction coefficients (LPCs) and converts it to a vector or matrix of line spectral pairs (LSPs) or line spectral frequencies (LSFs). When converting LPCs to LSFs, the block outputs match those of the poly2lsf function. [18]

4. Approach Presented work has been done in two main parts, first multilingual speech analysis/synthesis and second creation of neural network model as a classifier for speaker identification. Multilingual Speech Analysis/Synthesis consists of collection of speech utterances, pre-processing of speech utterances, features extraction, clustering of entire featured data. This is the first sub-process of this work and it consists of sentences of different languages taken from a public domain. Sound wave files (.wav) are created by using microphone connected with Personal computer at sampling rate 44,100 KHz and 16 bit per sample with stereo channel. For recording of sentences Free Sound Recorder 9.3.1 has been used and for preprocessing software Cool Record Edit Pro. For developing the speech database one sentence ( Now this time your turn ) is recorded from 19 speakers (11 male and 9 female). The sentence is translated in 8 different languages English, Hindi, Catalan, Finnish, Portugese, French, Italian and Indonesian as shown in Table I. The sentence is taken in such a way that in each word a good combination of consonant and vowel appear. The reason behind this is, whenever we pronounce any letter a vowel sound is always generated .For the sentence considered all five words are used in all eight languages, collectively 760 words are recorded in eight different international languages. Table I: Languages and Sentences SNo

Language

Sentence

1.

English

Now this time your Turn

2.

Hindi

Abb is samay tumhari baari

3.

Catalan

Ara aquesta vegada el torn

4.

French

Maintenant cette fois votre tour

5.

Finnish

Nyt talla kertaa sinun vuorosi

6.

Portugese

Agora esta vez transformer oseu

7.

Italian

Questa volta il tuo tourno

8.

Indonesian

Sekarang kali ini giliran anda

Pre-processing includes resampling sound files from Stereo channel to Mono channel by clipping words from the sentence of particular language, reducing noise from the file (Figure 2 (a) and (b)). Silence and noise are removed from speech signal and then splitting each sentence into individual words in each language for every user, in this way speech database of 760 words is created.

37

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012 Next step is extracting the features from these speech signals with the help of MATLAB (A computational language). Five types of features are extracted – LPC, LPCC, Cepstrum coefficient, , average PSD and LSF. Out of these features, cepstrum coefficient is the feature with maximum knowledge. Figure 4 shows how the features extracted from speech samples have been stored in speech database in MATLAB. The columns indicate speech utterances and the rows indicate features. Eight languages per speaker and five words each language makes it forty utterances for each speaker. Therefore for nineteen speakers, total number of utterances would be 760. With fourteen extracted features, the matrix size is 14x760. After obtaining these exclusive attributes from every utterance, the system is trained by using BPA with single hidden layer. Numbers of hidden neurons are being kept less than the target numbers. The training parameters for momentum, maximum epochs, non linear function, umber of hidden layers, number of neurons per layer, number of targets, training parameter are illustrated in Table II. Figure 2: (a) Data before resampling (Stereo Channel)

Figure 2: (b) Data after resampling (Mono Channel) The .wav files containing complete sentences are split into individual .wav files containing one word each as shown in Figure 3.

Table II: Training Parameters SNo

Parameter

Value

1.

Momentum

0.001

2.

Train Function

trainlm

3.

Training Parameter

0.30

4.

Maximum Epochs

1000

5.

Non-Linear Function

log-sigmoid

6.

Number of Hidden Layers

1

7.

Number of neurons in Hidden Layers

40

8.

Number of Target

19

Subsequent to the completion of the training, the next move is to simulate trained network and to check whether the class of actual output and target output is same. For simulation, a sample data is taken from the input set for testing and it is imitated with the trained network. A target matrix shown in Figure 5, mapped with the dimensions of input matrix is created by assigning each column with a unique value ranging between 0 and 1. Figure 6 shows Levenberg-Marquardt (trainlm) neural network training with 1000 iterations and it took 52 minutes and 51 seconds to get trained

5. Results

Figure 3: Splitting of sentences into individual word Splitting is done so that features of individual words can be extracted and stored in the database so that when matching is done, each and every word can be matched and this can only be done if features of individual words are already stored in speech database. This improves accuracy of the system.

When the proposed system is trained with Artificial Neural Network using Back Propagation Algorithm, keeping the number of Layers fixed, the system was able reach the performance of 95.657 % giving a total of 33 errors on a total number of 760 utterances (Input data) with 14 features. Table III shows the final training results in terms of Speakers, Number of input utterances, Number of errors and the percentage efficiency for nineteen speakers.

38

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012 Table III: Training Result Data

Speaker Number

Number of Input Utterances

Number of Errors

Efficiency (%)

1

40

0

100

2

40

1

97.5

3

40

2

95

4

40

1

97.5

5

40

1

97.5

6

40

0

100

7

40

3

92.5

8

40

0

100

9

40

4

90

10

40

2

95

11

40

3

92.5

12

40

0

100

13

40

2

95

14

40

3

92.5

15

40

1

97.5

16

40

3

92.5

17

40

1

97.5

18

40

2

95

19

40

4

90

Σ=760

Σ=33

Σ=95.657

The regression graph is shown below in Figure 7. Y is the output matrix obtained after training the input matrix under the defined conditions and parameters. T is the target matrix. Figure 8 gives the graphical representation of Mean square error w.r.t number of iterations. When the system is trained with ANN BPA as used by researchers, [16] for almost same problem with limited Indian Languages only, the performance was 95.354% with 42 errors. With the introduction of foreign languages and more words, the average performance of our system is of 95.657% giving a total of 33 errors.

6. Conclusion When using Back Propagation Algorithm, keeping the number of Layers fixed, the system was able reach the performance of 95.657 % giving a total of 33 errors on a total number of 760 utterances (Input data) with 14 features. Due to PC configuration problem number of neurons and number of layers could not be increased beyond a certain limit. Mean square error decreases with number of iterations and becomes almost stable after about 1000 iterations. Adding more features to system improves the user identification rate for the system. The result shows that BPA can be used for multi language system. This research focuses on text dependent speaker recognition. The desired goal of a system which can understand a text independent expression uttered by different speakers using various languages in different environments can be the further enhancement of this research.

Figure 4: Matrix showing extracted features from speech signals

39

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012

Figure 5: Target Matrix

Figure 6: Levenberg-Marquardt (trainlm) neural network training

40

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012

Figure 7: Regression Graph

Figure 8: Mean square error w.r.t number of iterations

41

International Journal of Computer Applications (0975 – 8887) Volume 57– No.13, November 2012

7. REFERENCES [1] Love,C. and Kinsne,W. , A Speech Recognition System Using A Neural Network Model For Vocal Shaping, Department of Electrical and Computer Engineering, University of Manitoba Winnipeg, Manitoba, Canada R3T- 2N2, March 1991, 198 pp. [2] Chougule,S. and Rege,P., Language Independent Speaker Identification, IEEE 3rd international conference, ieeexplore pp 364-368, May 2006. [3] Mak,M.W.; Allen,W.G. and Sexton,G.G., Speaker Identification Using Radial Basis Functions University of Northumbria at Newcastle, U.K., IEEE 3rd International Conference, ieeexplore pp. 138-142, 1993 [4] Bum,D., Experiments On Neural Net Recognition Of spoken And Written, Text IEEE Transaction on Acoust.. Speech and Signal Proc., vol. ASSP-36. no. 7. pp. 11621168, 1988. [5] Deiri,Z.and Botros ,N., LPC-Based Neural Network For Automatic Speech Recognition, Proc. IEEE Engineering in Medicine and Biology Soc., IEEE Cat. no. 90 CH2936-3. pp. 1429-1430, 1990 [6] Philip P. Wassermann,, Neural Computing: Theory and Practice , VNR, New York, 1989 [7] Richard P. Lipmann, Review of Neural Networks for Speech Recognition, Neural Computation 1, pp.1-38, Massachusets Institute of Technology, 1989. [8] Zebdum, RS; Guedes, K; Vellasco, M.; Pacheco, M.A,. Short Term Load Forecasting Using Neural Nets , Proceedings of the International Workshop on Artificial Neural Networks, LNCS No. 930, Springer Verlag, Torremohos, Spain, JuneJ995 [9] Zurada ,J. M., Introduction to W c i a l Neural Systems , West Publishing Company, 1992.

the IEEE International Advance Computing Conference, ieeexplore, pp 223-227, March 2009, Patiala, India [11] K. Santhosh, C. Mohandas ,V. P. and H. Li, Multilingual Speech Recognition: A Unified Approach , Proceedings of Interspeech 2005, (Eurospeech - 9th European Conference on Speech Communication and Technology), Lisboa, Sept. 2005. [12] S. C. Kumar; L. Haizhou, Language identification System for Multilingual Speech Recognition Systems , Proceedings of the 9th International Conference Speech and Computer (SPECOM 2004), St. Petersburg, Russia, Sept. 2004. [13] S. Hanwu, M.Bin, L. Haizhou, An Efficient Feature Selection Method for Speaker Recognition , International Symposium on Chinese Spoken Language Processing, December 16-19, 2008, China. [14] M. Bin, G.Cuntai, L.Haizhou, L.Chin-Hu, Multilingual Speech Recognition with Language Identification , International Conference on Spoken Language Processing (ICSLP), DENVER-COLORADO, Sept. 16-20, 2002. [15] N. Mohini, P.Manoj, P. Vijay, M.K. Prabhat, Vocal digit recognition using Artificial Neural Networks ,Dept. of Computer Engineering, VPCOE, Baramati in (2010 2nd International Conference on Computer Engineering and Technology) Page(s): V6-88 - V6-91 [16] A. Prateek, S. Anupam and T. Ritu, Multi Lingual Speaker Recognition Using Artificial Neural Networks , Advances in Computational Intelligence, Springer VerLlog,, 2009 , pages 1-9. [17] http://www.sp4comm.org/webversion.html [18] http://www.mathworks.in/help/toolbox/dsp/ref/lpctofrom cepstralcoefficients.html

[10] K. Rahul, S. Anupam, T Ritu, Fuzzy Neuro Systems for Machine Learning for Large Data Sets , Proceedings of

42

International Journal of Advent Research in Computer and Electronics (IJARCE) Vol.1, No.7, November 2014 E-ISSN: 2348-5523

Analysis on Software Project Scheduling and Staffing using ACO and EBS Prof. Dadaram Jadhav1, Akshada Paygude2, Aishwarya Bhosale3, Rahul Bhosale4 Dept. of Computer Engineering1234, Trinity College of Engineering, Pune, India1234 Email:[email protected]@gmail.com2 Abstract–In this world of technology, developing the software at the pace of changing requirements from clients is challenging and important for software engineers. A software project is a huge people intensive activity and the related work is mainly dependent on the technology we are using and people involved in developing the software. Therefore a proper model for software project planning and staffing has to focus on project task scheduling as well as human resource allocation. Managing both of these problems simultaneously and manually is a difficult job. Existing systems either suffer from a large search space, flexibility in human allocation and simplicity. Our proposed

1. INTRODUCTION In this aura of automation, software companies are facing a growing need for introducing automation in their software development life cycle. The highlights of a software project are resource allocation, techniques, project workload and schedule. A project manager alone cannot estimate all of these efficiently. Handling these critical activities manually can lead to poor performance and costly software project. It was reported that the major reason behind the unsuccessful software projects was inefficient planning of project tasks and staffing. Different from projects in other fields, resources in software projects can usually be allocated in more flexible way than those in construction or manufacturing projects. Project task requires employees with different skills. Expertise of employees notably influences efficiency of project execution. The tools based on traditional project management techniques like Program Evaluation and Review Technique termed as PERT, Resource constraint project scheduling problem termed as RCPSPdo not consider appointing multiskilled employees[1]. Our proposed system represents a plan by a task list using ACO and planned employee allocation matrix using EBS.We

system takes into account the problems of employee allocation as well as task scheduling by using Ant Colony Optimization (ACO) algorithm and Event Based Scheduler (EBS). The EBS regards the beginning time of the project, task preemption, adjusts the work load and coordination between the tasks and the employees. Different from existing approaches, the proposed methodology is characterized by flexibility in human resource allocation, task preemption and resource conflict. Keywords: Software planning and staffing, Employee allocation, Ant colony optimization, Event based Scheduler. will also be considering the cost minimization using COCOMO model and more stable workload assignments compared to other existing approaches. In this paper we develop a productive and practical system for Software Planning and Scheduling Problem(SPSP) and employee allocation with task preemption coordinately which is sparsely considered in traditional software management tools. 2. LITERATURE SURVEY Software project scheduling and staffing consist of four stages.

13

International Journal of Advent Research in Computer and Electr ctronics (IJARCE) Vol.1, No.7, November 2014 E-ISSN: 2348-5523 Usually task scheduling an and employee allocation are regarded as two separatee activities a and leave the job of staffing to be donee manually by the project manager resulting in poo oor management. A model either focuses on staffingg or o task scheduling but does not consider the importanc ance of coordinating both of these activities. A perso son may have to stop his current work and join som me other critical task so task preemption also has to be b considered as it has a significance impact on red educing the time and cost of a software project.

Fig. 1. Stages of SPSP The first stage comprises of collec ecting requirements from the clients regarding differe rent functions and other needs. After gathering the nee eeds from clients a proper plan has to be made for or the project like infrastructure, technology, manpow wer, schedules and more importantly what tasks are to t be included in constructing the actual softwaree project. So the second stage involves making aactual plans for defining the tasks and the related workload. wo The third stage is Task Ordering, to build a task ta list we have to determine order of task dependingg upon the priority of a task. To do this we will ill be considering positioning models. The last stagee is the assignment of employees to the task depending ng upon their skills and working hours. An employee al allocation matrix is used in this stage. Since one task ccan be assigned to several employees, one employe yee can undertake several tasks simultaneously and skill sk proficiency is considered it is important to define ne related tasks and employee allocation matrix appropri priately. 3. EXISTING SYSTEM Software management is usually co conducted by using traditional techniques like the program pro evaluation and review technique (PERT),, tthe critical path method (CPM) , and the resource-co constrained project scheduling problem (RCPSP) mo odel. A software project is a people intensive projec ject therefore these models even being important and he helpful prove to be inadequate for modeling uniquee characteristics of today’s software projects. A single le project task may require employees with different skills sk so appointing employees to the suitable tasks is tough for project managers. Resource allocation is not n considered by techniques like PERT and CPM and an RCPSP do not consider allocation of employees w with various skills.

g systemDisadvantages of Existing (1) Space complexity ty is high & lacks flexibility (2) Inefficient resourc rce allocation (3) Poor management nt performance (4) Allocate the sam me task different groups of employees in diffe fferent periods (5) Time consuming M: 4. PROPOSED SYSTEM TheSPSP model is the basic ba model required by all software projects, the mode del consist of: 1) Description of employee eesIt considers employees’ s’ information of wages, skills& working constrai raints. Using this we can allocate employees to suita itable task. 2) Description of taskThe tasks in software proj rojects can be anything such as designing, coding, testin sting etc. A Task precedence Graph (TPG) is used for or describing the tasks and ordering them. The attrib ibutes related to a task are work effort in person-mon onths, skill set of employees, maximumheadcount (numb mber of employees working on same task), deadline and an penalty. The precedence constraint is a task can only nly start when all of its direct predecessor’s tasks have finished. fin 3) Planning objectivePlan has to determine the he start time and the finish time of each task and tthe working hours of all employees.

4.1Ant Colony Optimizatio tion – The ACO algorithm is used for preparing a preordered task list and builds bu solutions depicting the foraging behavior of ants. a The pheromone is considered as a record off the past search experience of ants for guiding to make ke decisions. For finding the order of tasks we havee two types of pheromone models: 1) absolute positio ition and 2) relative position.

14

International Journal of Advent Research in Computer and Electronics (IJARCE) Vol.1, No.7, November 2014 E-ISSN: 2348-5523 In our system we are using the absolute position model which indicates the desirability of putting a task to a certain position using summation rule instead of using relative position as it is not necessary that two tasks will always be related to each other. In this way the task list is ordered. (1) Solution construction – Construction of task list and employee allocation matrix.

Fig. 2. An example of Task List

Fig. 4. Flowchart of the ACO algorithm

Fig. 3. Employee Allocation Matrix [1] (2) Pheromone management – pheromone values are updated according to the performance of the solutions built by ants. (3) Daemon actions – Daemon actions are optional and are used in improving performance. E.g.Local search procedure.

4.2Event Based Scheduler The Event Based Scheduler combines the task list representation and the employee allocation matrix so both the problems of task scheduling and human resource allocation are addressed. The concept of EBS is to manage the employee allocation at events without affecting the allocation at non-events. Modeling of resource conflict, task preemption and flexibility in human resource allocation is enabled by the strategy of proposed system. In addition EBS only makes new assignments at events; it is able to keep the implementation of tasks in a more stable manner. 5. CONCLUSION Our proposed system deals with the problems encountered by traditional systems. It uses the ACO to solve the complicated planning problem.[2] The EBS is effective as it helps in coordinating the tasks and allocation of employees and adjusts the planned working hours to actual working hours enabling task preemption and minimizing cost and the time required for the whole project. ACKNOWLEDGEMENT We here taking the opportunity in thanking all the people who helped is for completing out project, it would not have been possible to complete this project without them. Firstly we would thank our utter

15

International Journal of Advent Research in Computer and Electronics (IJARCE) Vol.1, No.7, November 2014 E-ISSN: 2348-5523 supportive guide, Prof. Dadaram Jadhav, for their ever ending guidance at all times and providing the guideline without any complaints at any time and also all of the ideas and effort they put into our project. We would also like to thank Prof. S. B. Chaudhari (HOD Comp.) for his interest & co-operation for the project. Expressing our sincere gratitude to TCOER, Pune, for providing us the facilities like Library & Internet. Also thanking all the people who have helped directly or indirectly in completing our project.

[9] AvinashMahadik, “An Improved Ant Colony Optimization Algorithm for Software Project Planning and Scheduling, ” International Journal of Advanced Engineering and Global Technology, Vol-2, Issue-1, January 2014

REFERENCES [1] Wei-Neng Chen and Jun Zhang, “Ant Colony Optimization for Software Project Scheduling and Staffing with an Event-Based Scheduler,” IEEE Transactions on software engineering, VOL. 39, NO. 1, JANUARY 2013 [2]Ms.Minal C.Toley, Prof.V.B.Bhagat, “An Application of Ant Colony Optimization for Software Project Scheduling with Algorithm In Artificial Intelligence, ”International Journal of Application or Innovation in Engineering & Management (IJAIEM) Volume 3, Issue 2, February 2014 [3]J. Duggan, H. Byrne, and G.J. Lyons, “A Task Allocation Optimizer for Software Construction,” IEEE Software, vol. 21,no. 3, pp. 76-82, May/June 2004. [4]R.-G.Dingand X.-H.Jing,“Five Principles ofProjectManagementinSoftwareCompanies,” Project Management Technology (in Chinese), vol. 1, 2003. [5] A. Barreto, M. de O. Barros, C.M.L. Werner, “Staffing a Software Project: A Constraint Satisfaction and Optimization-Based approach,”Computers & Operations Research, vol. 35, pp. 3073-3089,2008. [6]K.N.Vitekar,S.A.Dhanawe,D.B.Hanchate, “Review of Solving Software Project SchedulingProblem with Ant Colony Optimization,” International Journal of Advanced Research in Electrical, Electronics and Instrumentation Engineering,Vol. 2, Issue 4, April 2013 [7] A. Shtub, J.F. Bard, and S. Globerson, “Project Management: Processes, Methodologies, and Economics”, second ed. Prentice Hall, 2005. [8] O. Bellenguez and E. Ne´ron, “Methods for the Multi-Skill Project Scheduling Problem,” Proc. Ninth Int’l Workshop Project Management and Scheduling, pp. 66-69, 2004.

16

Volume 5, Issue 9, September 2015

ISSN: 2277 128X

International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: www.ijarcsse.com

Scale and Rotation Invariant Shape Features for Animal Identification Pavitra S Herakal* Department of CSE (PG) NMIT Bangalore, India

Shilpa Ankalaki Asst. Prof, Department of CSE (PG) NMIT Bangalore, India

Jharna Majumdar Prof. & Head, Department of CSE (PG) NMIT Bangalore, India

Abstract— Nowadays animal identification is an emerging area, due to extinction and endangering of animals. In order to identify an animal in an image, the image has to be described or represented by certain features. Shape is an important visual feature of an image. We can identify the animals using shape properties of different animals like dog, deer, camel, elephant etc. The main objective of this paper is to identification of animal. First label the multiple animals in image by using component labeling method. After labeling each animal by using image operation extract scale and rotation invariant shape features of the animal images, then creation of feature vectors for every animal in animal database to get the feature database. Finally using K-means clustering method used for identification of animal that achieves efficient animal identification System. Keywords- Image pre-processing, Component Labeling Method, scale and rotation invariant features, K-means Clustering Method. I. INTRODUCTION Animals play a vital role in our environment. In today’s generation day by day numbers of animals are decreasing. Identifying the animals through machine learning techniques. Using this application we can identify the animal and prevent from dangerous animal entering in the residential area, avoiding collision of vehicles and animals on the roads [9]. Animal identification system involves mainly 5 steps those are Capture the animal image, Image Preprocessing, Component labeling, image Feature Extraction, Classification or Clustering. Capture the animal image step is collection of animal images is from variety of animals. Pre-processing used to remove the noise part in image. It separates every object in the image. Image Feature Extraction step for animal identification uses some of scale and rotation invariant features methods used to classify the animals. After extracting each animal by using scale and rotation invariant shape features of the animal images, creation of feature vectors for every animal in animal database to get the feature database. Finally using K- means clustering method for identification of animal that achieves efficient animal identification System. II. PROPOSED METHODOLOGY In this paper a methodology has been proposed which will learn the shape features of the animal and stores it in a database and finally identifies it when given as a query. The proposed methodology consists of 2 main phases 1) Learning Phase 2) Identification Phase Fig 1 shows the Schematic diagram of diagram of the proposed method.

Fig 1: Overall System Diagram © 2015, IJARCSSE All Rights Reserved

Page | 151

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159 A. Learning Phase This is the phase where the proposed system learns about its database. In this phase the unknown animal are collected and stored in the database. Features which are invariant to translation, rotation and scaling are extracted and are stored in the database. These features are then clustered such that features that are most similar are grouped together. Learning phase involves the following steps ―data collection, pre-processing, feature extraction, feature normalization and Clustering Algorithms‖. Identification Phase: This is the phase where the proposed system identifies the query animals by comparing their features with the features in the database and retrieves those animals whose features match exactly or the nearest matched animals are retrieved. 1) Image Acquisition Image of the animal is acquired in a simple set up of Machine Vision System, a set of 40 animal images have been captured. The captured leaves should have white background. Figure 2 shows sample database of input animal database and figure 3 shows database of image with multiple animals

Fig 2 Example of animals Stored in Database

Fig 3: Input image with multiple animals Stored in Database 2)

Pre- Processing The raw data, depending on the data acquisition type is subjected to a number of pre processing steps to make it usable for subsequent processing. Pre processing aims to generate image data that is easy for the animal Identification system and can operate quickly and accurately. The acquired animal image which is in RGB format is converted to a gray scale image using the Equation (1): Gray = 0.2989 * R + 0.5870 * G + 0.1140 *B (1) Where R, G, B correspond to the Red, Green and Blue color of the pixel, respectively [3-4]. Fig 4 shows steps involved in pre-processing .

Fig 4 Steps Involved In Pre-Processing © 2015, IJARCSSE All Rights Reserved

Page | 152

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159 Next step is Thresholding, where objects of interest are separated from the background. Output of thresholding operation is a binary image in which pixels belonging to the object are represented by one gray level and all other pixels belonging to background are represented by another gray level. In the present work, Otsu’s thresholding [2] method is used to calculate the threshold value to convent the grayscale image into binary image. The Binarization processing is shown in Equation (2).

1 g ( x, y)  t d ( x, y)   0 g ( x, y)  t

(2)

g(x,y) is the gray value of a pixel, t is the threshold calculated by the OTSU method and d(x,y) is the binary value of the pixel. The binary image obtained after thresholding the connected component algorithm is applied to label the each animal from multiple animal image then contour extraction to determine the boundary of the object. Connected Component Labelling [6] Algorithm: Input: Binary Image Output: Components First scan Step1: Initialize an array Eq_label with labels [0-255] and curr_ label =0 For each foreground pixel encountered Step 2: check its top and left neighbor Step 3: if the top and left neighbor is a background pixel then Step 4: curr_label++; Step 5: Else if the top or the left neighbor is a background pixel then Step 6: Assign the label of the pixel which is not background Step 7: Else if label of the top pixel is not equal to the label of the left pixel and both the pixels are not background pixels then Step 8: Scan the Eq_label array from the beginning and replace the label of the top pixel by the label of the left pixel in the Eq_label array Second scan Step 9: For each foreground pixel encountered replace it with the label in the Eq_label array. 3)

Feature Extraction Shape Feature extraction involves the extraction of counter based feature methods [1] are Perimeter, Centroid Distance Method, Extreme Distance Ratio method, Diagonal Ratio Method, Perimeter area ratio, Diagonal Ratio Method and region based feature methods are Area, Waddell’s Ratio, Regional moment of Inertia, Shape square matrix, Area Ratio, Extent, [10] Axis Ratio of all of whom represent the shape of the animal.  Area: Animal Area can be defined as total number of black pixel in the binary image  Perimeter : Animal Perimeter is the number of pixels along the closed contour of the animal  Waddell’s Ratio

     

Waddell’s ratio = Drainage- Basin Ratio Drainage basin ratio= pi*Area/Perimeter Axis Ratio: Minor Axis Length / Major Axis Length. Area Ratio: Image area / image Filled Area. Perimeter area ratio: Image Perimeter / Image Area. Rectangularity: Rectangularity=As/Ar Rectangularity = (area)/(area of bounding box).

(3)

Centroid Distance Method: Convert the gray image into binary image then Find the Centroid of the image (xi), Find the distance between each boundary pixel to centroid using Euclidian distance. Find maximum distance [11]. Centroid Distance Method=maximum distance*size of the image /all distances

Algorithm of Centroid Distance Method: Step1: Convert the gray image into binary image Step2: Find the Centroid of the image(xi) Step3: Find the distance between each boundary pixel to centroid using Euclidian distance. Step4: Find maximum distance Step5: Centroid distance Method=maximum distance*size of the image /all distances © 2015, IJARCSSE All Rights Reserved

Page | 153

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159 Figure 6 shows representation of extreme distance method and Equation 4 is Central Point Determination formulas=

(4) 

Square Shape Matrix: figure 5(a),(b),(c) shows flow of Square Shape Matrix.

Fig 5 a) Original Shape Region (b) Square Model (c) Reconstruction Of The Shape Algorithm of Square Shape Matrix: Step1: convert grey scale image to binary image. Step2: Image is divided into number of blocks (64) Step3: Count the number of black pixels in each block Step 4: If the block contains the pixels more than the threshold value then make entire block black otherwise make entire block as white. Step5: repeat step 3 and 4 for all for all blocks. To make this feature scale invariant divide number of blocks covered by object and total number of blocks 

Extreme Distance Ratio method :

Fig 6 Representation of extreme distance method Extreme distance ratio =D11/(D11+D22+D33+D44) (5)  Extreme distance method Algorithm of Extreme distance method Step1:convert grey scale image to binary image. Step2:image height is divided into number(3) of parts. Step3:scan from left to right and right to left direction and take the extreme black pixel coordinates(x,y) for each part. Step 4: find out the distance between both extreme points using Euclidian distance formula. Step5: repeat step 3 and 4 for all parts. Step 6: compute the extreme distance ratio by the formula Figure 6 shows representation of extreme distance method Extreme distance ratio = distance of individual part (6) sum of distance of all parts 

Diagonal Ratio Method: Fig 7 shows Representation of Diagonal Ratio Method

Fig 7. Representation of Diagonal Ratio Method © 2015, IJARCSSE All Rights Reserved

Page | 154

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159  Diagonal Ratio Method Algorithm of Diagonal Ratio Method: Step1: convert grey scale image to binary image. Step 2: draw the bounding box around the image. Step3: find the centroid of bounding box. Step 4: find the distance from centroid to each corner of bounding box using Euclidian distance formula. Step 5: find the distance from centroid to extreme diagonal point of the image in all directions. Step 6: compute the diagonal ratio by the formula Diagonal ratio =



Regional moments of inertia: Regional moment of inertia describes pixel distribution information of the animal at different positions on its vertical axis [8]. Animals preserve steady orientation and are bilaterally symmetric concerning their vertical axis; P xx quantities for four different regions are used as a descriptor, where Pxx is the variable that holds the pixel distribution value. It is scale invariant. Fig.8 shows Regional moment of inertia of animal. Pxx can be calculated using the Equation (3). (7) Where Xi : Vertical coordinates of the animal. X centroid: Vertical coordinates of centroid animal image. N : Number of black pixels in each region Figure 8 shows representation of regional moment of inertia Table 1 shows result analysis of the scale invariant features.

Fig 8. Regional Moment of Inertia TABLE I RESULT ANALYSIS OF ALL SCALE INVARIANT FEATURES OF CAMEL Image Size 128x128 256x256 300x300 400x400 Features 0.168955 00.168363 0.164513 0.159568 Drainage Basin Ratio 0.367947 0.365234 0.365000 0.364544 Extreme Distance Ratio(D1,D2,D3) 0.281722 0.287109 0.286667 0.287141 0.318212 0.318359 0.318333 0.318352 1.655662 1.666120 1.677450 1.690798 Diagonal Ratio Method 1.020143 1.023510 1.022956 1.023696 (R1,R2,R3,R4) 2.274244 2.270559 2.267824 2.260657 1.073696 1.073423 1.075344 1.073383 0.395020 0.395231 0.394167 0.394219 Area Ratio 1.000000 1.000000 1.000000 1.000000 Axis Ratio 0.407659 0.401265 0.399475 0.394219 Extent 0.087764 0.084954 0.087546 0.082354 Circularity 1.394994 1.398644 1.400586 1.401991 Centroid distance Method 0.018042 0.018899 0.019032 0.019116 ROI 0.104465 0.105418 0.105532 0.105714 0.288340 0.288409 0.288408 0.288603 0.589153 0587282 0.587028 0.586522 0.385000 0.383789 0.398101 0.391200 Shape Matrix © 2015, IJARCSSE All Rights Reserved

Page | 155

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159 Moment Invariant Features The Invariant Moments are widely applied because of their unique property of ―translation, rotation and scale invariance‖[5].In our proposed work Hu moments were used. Moment Invariants were first introduced by Hu. He proposed 6 orthogonal invariant moments and one skew invariant moment [3] based on algebraic moments. A two

dimensional central moment is given by The 7 invariant moments introduced by Hu are as follows

TABLE 2 RESULT ANALYSIS OF HUE MOMENT INVARIANT FEATURES OF CAT Features Original(128x128) 0.292599 0.008760 0.0006357 0.001170 0.000003 0.000006 0.000001 0.293915 0.009995 0.006047 0.000913 0.000002 0.000004 0.000001 0.296275 0.012214 0.005786 0.000774 0.000002 0.000008 0.000001 0.299969 0.014211 0.005673 0.000738 0.000002 0.000008 0.000001 0.292599 0.008760 0.0063860 0.001182 0.00002 0.000019 0.000001 256x256 0.293764 0.009892 0.0062 8 0.00965 0.00003 0.00004 0.000001 0.296653 0.012028 0.005938 0.000839 0.00002 0.000005 0.000001 0.298666 0.014205 0.005600 0.000720 0.00002 0.000003 0.000001 4) Clustering Clustering is an unsupervised machine learning technique. The purpose of clustering is to put data points into related groups without having prior knowledge of the group definitions where every group is termed as cluster. In proposed work clustering is used to cluster database animals in groups where animals in one cluster are similar between them and are dissimilar to the animals belongs to the other clusters [7]. K –Means Algorithm INPUT: Set of feature values. OUTPUT: Clusters of feature values. Step1: Place K points into the space represented by the objects that are being clustered. These points represent initial group centroid. Step2: Assign each object to the group that has the closest centroid. Step3: When all objects have been assigned, recalculate the positions of the K centroid. Step4: Repeat Steps 2 and 3 until the centroid no longer move. This produces a separation of the objects into groups from which the metric to be minimized can be calculated. B. Identification Phase In this phase, animals which are to be identified are placed in from of the camera and its image is acquired, it is then pre-processed and then component labeling method is used to detect multiple animals, features of each animal from the query image is extracted, they are then normalized and matched with the database based on the distance measure and the animals are identified. The query input can be a single animal or multiple leaves. III. EXPERIMENTAL RESULTS A. Results of Component Labeling The query image consisting of multiple animals is given as input to the proposed system, after preprocessing Component labeling method is applied to identify multiple animals. Figure 9(a) shows sample input image to the application and figure 9(b) shows results of connected component labeling. © 2015, IJARCSSE All Rights Reserved

Page | 156

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159

Fig 9(a) Input Image

Fig 9(b) Results of Component Labeling B. Results of animal Identification Figure 10(a)(b)(c)(d)(e) shows the results of multiple animal identification. As seen in the figure 4 animal have been identified.

Fig 10(a) Input Image

Fig 10(b) Identified animal: Deer © 2015, IJARCSSE All Rights Reserved

Page | 157

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159

Fig 10(c) Identified animal: Dog

Fig 10(d) Identified animal: horse

Fig 10(e) Identified animal: elephant Figure (10a) Query Image containing 4 animals, 10(b-e) shows the results of identified animal © 2015, IJARCSSE All Rights Reserved

Page | 158

Herakal et al., International Journal of Advanced Research in Computer Science and Software Engineering 5(9), September- 2015, pp. 151-159 IV. CONCLUSION AND FUTURE WORK In this paper of animal identification has been concerned with the two challenging steps in which are scale and rotation invariant retrieval of features and Identification. In proposed system different scale and rotation invariant features extraction methods those are describes the shape of the animal and moment invariants method are implemented. In Identification phase the K-means clustering algorithm is used to group the similar animals having the similar shape feature values. The proposed methods restrict the animal identification to the simple animal with white background. Future scope is identification of animal with complex background, accuracy and performance improvement of animal identification. ACKNOWLEDGEMENT The authors acknowledge Prof. N R Shetty, Director, Nitte Meenakshi Institute of Technology and Dr. H C Nagaraj, Principal, Nitte Meenakshi Institute of Technology for providing the support and infrastructure to carry out our research. REFERENCES [1] Mingqiang Yang, Kidiyo Kpalma, Joseph Ronsin. ― A Survey of Shape Feature Extraction Techniques.‖ PengYeng Yin. Pattern Recognition, IN-TECH, pp.43-90, 2008. [2] https://en.wikipedia.org/wiki/Otsu's_method. [3] S. G. Wu, F. S. Bao, E. Y Xu, Y-X. Wang, Y-F. Chang, & Q-L.Xiang, ―A Leaf Recognition Algorithm for Plant Classification Using Probabilistic Neural Network‖, IEEE 7th International Symposium on Signal Processing and Information Technology, Cairo, 2007. [4] J. Du, X. Wang, and G. Zhang, ―Leaf shape based plant species recognition,‖ Applied Mathematics and Computation, vol. 185-2, pp. 883-893, February 2007. [5] Zhihu Huang, Jinsong Leng ―Analysis of Hu's Moment Invariants on Image Scaling and Rotation‖ , Edith Cowan University,2010,IEEE [6] Luigi Di Stefano, Andrea Bulgarelli , ―A Simple and Efficient Connected Components Labeling Algorithm‖, Italy [7] Aastha Joshi , Rajneet Kaur, ―A Review: Comparative Study of Various Clustering Techniques in Data Mining‖ , International Journal of Advanced Research in Computer Science and Software Engineering, March 2013 [8] David Knight, James Painte, Matthew Potter, ―Automatic Plant Leaf Classification for a Mobile Field Guide‖. [9] Mukesh B Rangdal, and Dinesh B. Hanchate, ―Animal Detection Using Histogram Oriented Gradient,‖ International Journal on Recent and Innovation Trends in Computing and Communication, vol. 2, feb 2014. [10] João Ferreira Nunes Pedro Miguel Moreira João Manuel R. S. Tavares ―Shape Based Image Retrieval and Classification‖ CISTI ,2010 [11] Dengsheng Zhang and Guojun Lu ―Shape Retrieval Using Fourier Descriptors with Different Shape Signatures ― Université Européenne de Bretagne, France – INSA 2007

© 2015, IJARCSSE All Rights Reserved

Page | 159