Computers and Software

International Review on

Computers and Software (IRECOS) Contents

.

Scheduling Strategies in Grid Computing Environment: a Survey by D. I. George Amalarethinam, P. Muthulakshmi

1134

Performance Analysis Comparative Study of Fingerprint Recognition Systems by Mohammed Saeed Jawad, Fares Al-Shargie, Murthad Al-Yoonus, Zahriladha bin Zakaria

1154

A Comparative Analysis on Markov Model (MM) Association Rule Mining (ARM), Association Rule Mining-Statistical Features (ARM-SF), Boosting and Bagging Model (BBM) to Impervious Web Page Prediction by Sampath Prakasam, Amitabh Wahi, Ramya Duraisamy

1163

A Rough Set Based Classification Model for Grading in Adaptive E-Assessment by G. S. Nandakumar, V. Geetha, B. Surendiran, S. Thangasamy

1169

Evaluation and Analysis of Novel Mobile Agent Model in Network Fault Management by S. Kavitha, K. V. Arul Anandam

1178

Data Access Prediction and Optimization in Data Grid Using SVM and AHL Classifications by R. Kingsy Grace, R. Manimegalai

1188

Fuzzy Segmentation and Feature Extraction for an Efficient Identification of Mine-like Objects by G. Suganthi, Reeba Korah

1195

RVARM Algorithm for Classifying Event Based Video Sequences by D. Pushpa Ranjini, D. Manimegalai

1201

An Enhanced Backpropagation Neural Network for Multimodal Biometric System by K. Krishneswari, S. Arumugam

1208

Arabic Handwriting Text Offline Recognition Using the HMM Toolkit (HTK) by H. El Moubtahij, A. Halli, K. Satori

1214

Context Aware Adaptive Fuzzy Based Quality of Service Over MANETs by A. Ayyasamy, K. Venkatachalapathy

1220

Review on CBIR Trends and Techniques to Upgrade Image Retrieval by B. Thenkalvi, S. Murugavalli

1227

A Novel Framework for Alias Detection in Web Search Using Extreme Learning Machine (ELM) Approach by Subathra M., Nedunchezhian R.

1241

A Framework for SMS Spam and Phishing Detection in Malay Language: a Case Study by Cik Feresa Mohd Foozy, Rabiah Ahmad, Faizal M. A.

1248

Developing of Natural Language Interface to Robot - an Arabic Language Case Study by Ayad T. Imam, Thamer Al-Rousan, Ashraf Odeh

1256

Design of High Speed Modulo Multipliers by A. Rosi, R. Seshasayanan, A. Nisha

1263

(continued)

Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

QoS-Based Parallel GRASP Algorithm for RP Selection in PIM-SM Multicast Routing and Mobile IPv6 by Y. Baddi, M. D. Ech-Cherif El Kettani

1271

Fast Localization of the Optic Disc in Retinal Images Using Intensity and Vascular Information by J. Benadict Raja, C. G. Ravichandran

1282

Optimized Network Selection Using Aggregate Utility Function in Heterogeneous Wireless Networks by C. Amali, Dhanasree Jayaprakash, B. Ramachandran

1293

Rainfall Intensity Classification Method Based on Textural and Spectral Parameters from MSG-SEVIRI by Y. Mohia, S. Ameur, M. Lazri, J. M. Brucker

1302


International Review on Computers and Software (I.RE.CO.S.), Vol. 9, N. 7 ISSN 1828-6003 July 2014

Scheduling Strategies in Grid Computing Environment: a Survey D. I. George Amalarethinam1, P. Muthulakshmi2 Abstract – Grid Computing is a technology that integrates various forms of resources that contribute themselves in a geographically distributed environment. The heterogeneous resources are utilized by large-scale business and scientific applications. The resource management complexities increase due to the heterogeneity of resources and scalable nature of the grid environment. To manage the issues efficient resource management techniques are needed. Proper scheduling strategies can increase the efficiency of resource management systems. Scheduling is the process of allocating the jobs to the appropriate resources. This paper presents a package of scheduling strategies and factors that influence grid scheduling. The study finds algorithms that are data intensive, computation intensive, communication intensive, application intensive and so on. Generally these algorithms emphasis minimum execution time, load balancing, fault tolerance, task duplication, network behavior, job migration, rescheduling, dependencies, etc. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Parallel Processing, Distributed Computing, Grid Computing, Scheduling Policies, Scheduling Algorithms, Heuristics, Hybrid Scheduling Algorithms

I.

Introduction

It is required to have high performance super computers to process petabytes of information in a short span of time. Utilizing the super computers is very expensive. The idea to aggregate and use large number of distributed resources to solve complex problems lead to the technology called grid computing. Grid computing is a form of parallel and distributed computing system composed of heterogeneous resources owned by various organizations and connected by different networks of various characteristics. The environment is a virtual super computer that helps us to access multiple resources simultaneously to work out big complex applications at an economical rate. Foster et al [1] defined grid computing as a coordinated resource sharing, problem solving, dynamic, multi-institutional collaborations. Parallel processing is a promising approach to meet the computational requirements of a large number of current and emerging applications [2]. Grid computing fulfills the computational requirements of extremely large problems related to science, engineering, research and commerce by facilitating to share the heterogeneous distributed resources in a parallel fashion. Resource management is always a big challenge in such environments due to the heterogeneity of resources, networks, vendors, policies and clients. Efficient resource management systems are desired to achieve high performances. To improve the performance of resource management systems efficient grid scheduling algorithms are required. Scheduling refers to the way of assigning the applications to run on available resources.

Job scheduling is a process of establishing a queue to run a sequence of programs over a period of time. Task scheduling is the process of mapping tasks to a selected group of resources which may be distributed in multiple administrative domains. But the problem lies in developing an algorithm to suit various forms of applications. Generally applications to be solved in a grid environment are interdependent task models known as work flows. In work flow model, the applications may be partitioned into heap of tasks that are expected to run parallel. Hence coordination is expected among tasks. Grid computing helps to identify and filter resources to suit the application; also it synchronizes tasks and resources. An appropriate scheduling can enhance the functional capability of the grid system. The primary objective of grid scheduling is to minimize the completion time of parallel applications by carefully allocating the tasks to the appropriate resources. Scheduling can be classified into two main categories called static scheduling and dynamic scheduling. In static scheduling, there is no job failure and resources are assumed to be available all the time. But this is not applicable for dynamic scheduling. There are two main aspects that determine the dynamicity of the grid scheduling namely, (i) the dynamicity of job execution, which refers to the situation when job execution can be failed due to some resource failures or job execution could be stopped due to the arrival of high priority jobs in the system (when the case of preemptive mode is considered) (ii) the dynamicity of resources, which refers to the behavior of resources. The resources can join or leave the grid in an unpredictable way. The workload of the resources can significantly


1134

D. I. George Amalarethinam, P. Muthulakshmi

vary over time and the local policies to use the resources can be changed based on utilization, vendors and so on. The above mentioned factors decide the behavior of the Grid Scheduler (GS) to be static or dynamic. This paper presents various aspects of scheduling and the factors that influence scheduling. In the paper processors, resources, computing nodes are used interchangeably. Similarly process, tasks, job are used alternatively. The paper is structured on sections as follows, section II shows taxonomy of grid, section III discusses grid scheduling, section IV lists various types of scheduling policies and section V presents a collection of traditional scheduling algorithms. In section VI, a discussion on various scheduling algorithms that are proposed with combined factors is presented and finally the paper is concluded in section VII.

II.

APPLICATIONS AND SERVICES OPERATIONAL SUPPORT, UTILITY MANAGEMENT IN BUSINESS SOLUTIONS AND OTHERS

MIDDLE WARE SOFTWARE RESPONSIBLE FOR RESOURCE MANAGEMENT, JOB SCHEDULING, SECURITY, TASK EXECUTION

Taxonomy of Grid

The organization of the grid infrastructure involves (1) fabrication level, (2) middleware, (3) applications and services as shown in Fig. 1. Fabrication level aggregates the physical components. Middleware is associated with software responsible for resource management, job scheduling and task execution. Services will be utilized by users and offered by vendors and it includes applications. These applications consist of operational support, utility management, and business solutions/tools. There are many types of grids based on the factors like the structure of the organization, the resource used in the grid and the services offered by the grid. They can be classified as: (1) Departmental Grids are used to solve problems for a particular group (ex. Smart Grid), (2) Enterprise Grids have been established by the enterprises consisting of resources and they offer services to customers (Sun Enterprise Grid), (3)Extraprise Grids are established between companies, partners and customers (ex. Amazon), (4) Global Grids hold itself over the internet (ex. White Rose Grid), (5) Computational Grids are solely for the use of providing access to computational resources (ex. SETI@home [3]), (6) Data Grids are highly data intensive and are optimized for data oriented operations (Biomedical informatics Research Network), (7) Utility Grids are Grid Resource Providers, generally they provide access to resources, (8) National Grids are nation run grids managed by the country to plan during disaster etc (ex. GridPP, GARUDA).

III. Grid Scheduling The objective of grid computing is to provide transparent and efficient access to remotely distributed resources. Proper resource management techniques are required to handle enormous resources and tasks. Scheduling is the process of mapping tasks and resources. Scheduling decides the resource that results in a minimum execution for a task to run itself. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

FABRICATION HARDWARE RESOURCES LIKE COMPUTERS, SENSORS, NETWORKING DEVICES

Fig. 1. Taxonomy of Grid

Scheduling is needed to coordinate the accessibility of available resources by the tasks. Implementing an efficient scheduling in resource management system increases the performance of the system. Due to dynamicity and heterogeneity of resources and tasks in grid environment, it is very hard to devise a scheduling algorithm that could prove itself in all situations. Grid Brokers are the interface between the vendors and users of grid. Grid Brokers System maintains the information of resources and their policies and also the application and their requirements. Grid Schedulers use this information to decide the task to be executed by a particular resource. Generally Grid Schedulers use scheduling algorithms like First Come First Served algorithm, Shortest Path First algorithm, Round Robin technique, Least Recently Serviced algorithm to schedule jobs. Grid scheduling is influenced by the factors like job priority, availability of resource, reservation of resources, specific resource requirement and cost for execution. A careful scheduling minimizes the cost and time of execution. Therefore additional features are always encouraged to avoid cost and time overrun. Keeping check points and mentoring is a mechanism, in which the place where the process has stopped or malfunctioned due the reason like resource failure can be identified. The mechanism of check pointing helps in resuming the jobs where it has stopped and may lead to the completion of the job at the earliest (without restarting jobs from the beginning).

International Review on Computers and Software, Vol. 9, N. 7

1135


This avoids the cost to be spent on executing the job from the beginning. Fault due to load imbalance and other reasons result in the migration of jobs to other nodes. Grid system is a decentralized system, in which failure detection should not affect the computation and therefore a simultaneous rescheduling is needed. Table I shows some of the factors that influence the work flow in grids.

IV.

possible to find the expected execution time in advance. The jobs would be entering dynamically and the scheduler has to work hard in making decisions to allocate resources. This is the situation where the load balancing is a factor to be considered seriously. In dynamic scheduling, the system need not possess the run time behavior of the application before it runs and it is an advantage over static scheduling. IV.7. Independent Scheduling

Scheduling Policies

IV.1. Immediate Mode [4] In immediate mode scheduling, tasks are scheduled as soon as they enter the system. Further they will not be waiting for the next time interval. Tasks will be scheduled when the scheduler gets activated. Scheduling will be easier if the job arrival rate is less than the number of resources that are available. Enough availability of resources could start the execution of tasks immediately.

In independent scheduling, the applications would be able to partition themselves into almost independent parts (or loosely coupled), which can be scheduled independently. IV.8. Centralized Scheduling [5]

In batch mode scheduling, tasks are grouped into batches and are allocated to the resources by the scheduler. The results of processing are usually obtained at a later time. Batch scheduling could take better advantage of job and resource characteristics in deciding the allocation of a job to the resource.

In dynamic scheduling scenarios, the responsibility for making global scheduling decisions may lie with one centralized scheduler, or may be shared by multiple distributed schedulers. The advantage of having centralized scheduling is the ease of implementation, but lacks scalability, fault tolerance and sometimes performance. In centralized scheduling, there is more control over resources and the scheduler has knowledge of the system by monitoring the resource state. Therefore it is easier to obtain efficient schedules. This type of scheduling would suffer from limited scalability and is not suitable for large-scale grids.

IV.3. Preemptive Scheduling [4]

IV.9. Co-Operative Scheduling [5]

In the preemptive mode, the current execution of the job can be interrupted and it could be migrated to other resource for execution. Preemption can be useful if job priority is considered as one of the constraints.

Many schedulers will participate in cooperative scheduling and each grid scheduler has the responsibility to carry out its own portion of the scheduling task, but all schedulers are working towards common system-wide reach. A feasible schedule is achieved through the cooperation of procedures, rules, and grid users.

IV.2. Batch Mode [4]

IV.4. Non Preemptive Scheduling [4] In the non-preemptive mode, a task is expected to be completed in the particular resource, where it started its execution (the resource cannot be taken away from the task; similarly the task cannot be migrated to other resources). IV.5. Static Scheduling [5] In static scheduling, the information regarding the available resources and the tasks in an application is assumed to be known in advance by the time when the application is scheduled. Assigned task cannot be changed from the resource until it gets over. To the scheduler’s point of view, it is easier to adopt. IV.6. Dynamic Scheduling [5] In dynamic scheduling, the task allocation is done onthe-go when the application is executed, where it is not


IV.10.

Non–Cooperative Scheduling [5]

In non-cooperative cases, individual schedulers act as autonomous entities and they arrive at decisions regarding their own optimum objects. These decisions are independent of other decisions taken for the rest of the system. IV.11.

Decentralized Scheduling [6]

In decentralized or distributed scheduling, there is no central entity to control the resources. The autonomous grid sites make it more challenging to obtain efficient schedulers. In decentralized scheduling, the local schedulers play an important role. The scheduling requests (either by local users or other grid schedulers) is sent to local schedulers which manage and maintain the state of the job queue. Although decentralized schedulers


1136


could be less efficient than centralized schedulers, this type of scheduling is more realistic for real grid systems of large-scale. IV.12.

The top level is called the meta-level, a super manager in this level would control group merging/partitioning. In this level, tasks are not scheduled directly, but scheduler will be reconfigured according to the characteristics of tasks and resources. This is called meta-scheduling. The mediate level is called the group level, where the manager in each group collaborates with each other and allocates tasks for the workers in the group. The purpose of establishing collaborations between managers is to improve the load balance across groups. Specifically, the managers ensure the workload in the local ready lists is roughly equal for all groups. The bottom level is called the intra-group level, where the workers in each group perform self-scheduling. The hierarchical scheduler behaves between centralized and distributed schedulers in order to complete executions.

Adaptive Scheduling [7]

The changes in grid computing environment during the recent times require adaptive scheduling techniques, which will take the account of both the current status of the resources and predictions for their future. This is done to detect and avoid performance deterioration. Rescheduling can also be seen as a form of adaptive scheduling in which running jobs can be migrated to more suitable resources. IV.13.

Hierarchical Scheduling [8]

In hierarchical scheduling, the computation is done at three levels.

TABLE I CLASSIFICATION OF FACTORS AND THEIR ATTRIBUTES INFLUENCING THE WORK FLOW IN THE GRID SYSTEM Entities Characteristics Behaviors Complexity of the Problem Criteria to Solve the Problem

FACTORS INFLUENCING THE WORK FLOW IN A GRID SYSTEM

Availability of the Resources

Network Behavior

Nature of the Tasks Nature of the Resources Application/Problem Demand

Low, Medium, High Communication Time Computation Time Static Dynamic

Uniform, Varying Uniform, Varying Load Balancing Task Duplication Load Balancing Task Duplication Check Pointing Job Migration Rescheduling

Topology Centralized Network Decentralized Network Data Transfer Rate Bandwidth Dependent, Independent Homogeneous, Heterogeneous Resource Intensive

Uniform, varying

Application Intensive Data Intensive

Mapping Dependency

Data Processing

Priority Based Random Allocation Relative Algorithms Data Availability

Data Duplication/ Replication Access from the Source Others Resource Failure Idle Resources Faulty Resources Multi Site Policies Data Loss Task Duplication Task Grouping


Availability Computing Speed Data Dependency Data Transfer Rate Storage and Locality Loss And Backup Data Transfer Rate

Readily Available Transfer Rate Shortest Distance Path Shortest Cost Path Memory Availability Shortest Path Check Pointing Rescheduling Job Migration Fault Rate Restrictions On Resources Recovery Through Backup


1137


IV.14.

List Scheduling [9]

A list scheduling heuristic prioritizes workflow tasks and schedules them based on their priorities. The basic idea of list scheduling is to make an ordered list of processes by assigning them some priorities and execute the following steps repeatedly until a valid schedule is obtained. Steps: 1) The process with the highest priority for scheduling is selected from the list. 2) Select a resource to execute the process chosen in previous step. The priorities are determined statically before the scheduling process begins. Process of high priority is selected in the first step and the best resource is selected during the second step. Some of the list scheduling strategies are Highest Level First, Longest Path, Longest Processing Time, Critical Path Method etc. IV.15.

Master Slave Scheduling [10]

In master slave scheduling model, each job has to be processed sequentially in three stages. Multiple masters and slaves are considered. In the first stage, the preprocessing task runs on a master resource. During the second stage, the slave task runs on a dedicated slave resource. In the last stage, the post processing task runs again on a master resource, possibly different from the master resources selected in the first stage.

V.

Traditional Scheduling Algorithms V.1.

First Come First Served

U. Schwiegelshoh and R. Yahyapour [11] analyzed the First Come First Served (FCFS) method for parallel processing. The resource with the least waiting queue time can be given to the task that enters the system. This is called as Opportunistic Load Balancing (OLB) [12] or myopic algorithm. It is found to be very simple, and not optimal. OLB assigns tasks to the available resources in an arbitrary order, irrespective of the task's expected execution time on that resource [13]-[15]. It is noted that the intuition behind OLB is to keep all resources as busy as possible. Simplicity is the advantage of using OLB. Expected task execution time of a task is not considered. It is stated that sometimes the allocation results in poor makespans. V.2.

Min-Min

R. Armstrong et al. [13] stated that the Min-Min heuristic begins with the set of all unmapped tasks having the set of minimum completion times. Then task with the overall minimum completion time from the set is selected and assigned to the corresponding resources. The newly mapped task is removed from the set, and the process repeats until all tasks are mapped (i.e., until the set is found to be empty). Min-Min is based on the Minimum Completion Time (MCT).


However, Min-Min considers all unmapped tasks during each mapping decision and MCT considers only one task at a time. Min-Min maps the tasks in the order that changes the resource availability status. The expectation is that a less makespan can be obtained if more tasks are assigned to the resources that complete at the earliest. V.3.

Minimum Completion Time

R. F. Freund et al. [14] said that the Minimum Completion Time (MCT) assigns each task in an arbitrary order to the machine with the expected minimum completion time. It follows arbitrary order to map tasks [13]. This causes some tasks to be assigned to resources that do not have the minimum execution time (MET). The intuition behind MCT is to combine the benefits of OLB and MET (i.e., avoiding the circumstances in which OLB and MET perform poor). V.4.

Backfilling

Ahuva W. Mu’alem and Dror G. Feitelson [16] proposed the idea of back filling in distributed environments. Backfilling is a scheduling optimization technique that allows a scheduler to make better use of available resources by running jobs out of order. The scheduler prioritizes the jobs in the queue. The prioritized jobs are kept at the front of the list. The priority can be given based on the factors like First Come First Served, Minimum Execution Time and Maximum Execution Time. Backfilling allows small jobs to initiate before larger jobs in the queue are allocated. This can happen when the resources for the larger jobs are not available at the time. It is required that all jobs service times to be known in advance. Stavrindis et al. [17] stated that the service time can be provided either as estimation by users during job submission or predictions made by the system based on historical data. V.5.

Round Robin Technique

Ruay-Shiung Chang et al. [18] said that Round Robin (RR) algorithm focuses on fairness. RR uses the ring structure as its queue to store jobs. Each job in a queue has the same execution time and will be executed in turns (rounds). If a job is unable to get completed during its first turn, it will be stored back in the queue and will be waiting for the next turn to come for continuing its execution. Each job will be executed in turns and they would not be required to wait for any other jobs to get completed. RR will be taking a very long time to complete all the jobs when load is high. The priority scheduling algorithm could assign a priority value to the jobs and the same is used to dispatch jobs. The priority value of each job will be depending on memory size, CPU time and so on.


1138


The disadvantage of this algorithm is that it may cause an indefinite blocking or starvation if the requirement of a job is never being satisfied. V.6.

Earliest Deadline First

Earliest Deadline First (EDF) [19] or Least Time-toGo is a dynamic scheduling algorithm used in real-time operating systems. The jobs are placed in a priority queue as they enter the system. Whenever a scheduling event occurs (task finishes, new task enters, etc.) the queue is searched for the process closest to the deadline of the job that has been completed earlier. The found job can be scheduled next for execution. EDF is an optimal scheduling algorithm on preemptive uni-processors. V.7.

V.10. Genetic Algorithm

Minimum Execution Time

R. F. Freund et al. [14] investigated that in contrast to OLB, Minimum Execution Time (MET) uses an arbitrary order to assign tasks to resources with the best expected execution time for the tasks, regardless of the availability of resources. The motivation behind MET is to give the best resource for each task. This can cause a severe load imbalance across resources. In general, this heuristic is suitable for high computing environments [20]. V.8.

Max Min

O. H. Ibarra et al. [21] stated that Max-Min heuristic is very similar to Min-Min. The Max-Min heuristic begins with a set of all unmapped tasks. Minimum completion time is calculated for all the tasks. The task having maximum completion time among the tasks is selected and assigned to the suitable resource (hence said to be Max-Min). The mapped task is removed from the set and the process repeats until all tasks are mapped (i.e., until the set is found to be empty). Naturally, Max-Min attempts to minimize the penalty incur from performing tasks with longer execution times. It is assumed that the meta task to be mapped would be with more tasks of very less execution times and one task of very long execution time. Mapping the task with the longer execution time to its best resource would allow the task to be executed concurrently with the remaining tasks (with shorter execution times). This is considered to be a better mapping than a MinMin mapping. In Min-Min allocation all smaller tasks are given priority to execute earlier than the longer running task. In spite of several resources remain idle. Hence it is said that Max-Min heuristic can give a balanced load among the resources and result in better makespans. V.9.

M. Dorigo et al. [22] presented Ant Colony Optimization (ACO) algorithm for grids. ACO is found to be dynamic in nature. ACO is a heuristic algorithm with efficient local search for combinatorial problems. ACO imitates the behavior of real ant colonies in nature to search for food and to connect each other by pheromone laid on paths traveled [23]. Pheromone is a chemical substance that could be laid and used by ants to search food and traverse the path of the ants that travel. In ACO algorithm, the pheromone value (the execution time on processors) is used to allocate the tasks to resources. ACO has been used to solve NP-hard problems such as traveling salesman problem [24], graph coloring problem [25], vehicle routing problem [26], and so on.

Ant Colony Optimization

A good schedule should accommodate its scheduling policy according to the dynamicity of the entire environment and the nature of jobs.


GA was proposed by J. H. Holland in 1975. Prodan and Fahringer [27] used Genetic Algorithm (GA) in grid system for scheduling. It is an evolutionary technique for large space search. Braun. R. et al. [20] have given out the general procedure of GA search and is as follows: (i) Population Generation: A population is a set of chromosomes, representing a possible solution, which can be a mapping sequence between tasks and resources. The initial population can be generated by heuristic algorithms like Min-Min. (ii) Chromosome Evaluation: Each chromosome is associated with a fitness value, which could be the makespan of task on a resource. The goal of GA search is to find the chromosome with optimal fitness value. (iii) Crossover and Mutation Operations: Crossover operation is used to select random points in a random pair of chromosomes to exchange gene values. This is used to exchange the assignments between tasks and resources. Mutation selects a random task in the solution and assigns it to a new random resource. (iv) Evaluation of the modified chromosome: The chromosomes from this modified population are evaluated again. The above defined steps complete an iteration of the GA. The GA could be stopped when any one of the following situation is met: (i) predefined number of evolutions are made, (ii) no improvement in recent evaluations, (iii) all chromosomes converge to the same mapping, (iv) a cost bound is met. GA is the most attracted and popular Nature’s Law heuristic algorithm used in optimization problems. V.11. Simulated Annealing Y. Liu [28] investigated Simulated Annealing (SA) and described it as a search technique based on the actual process of annealing. Annealing is the thermal process of obtaining lowenergy crystalline states of a solid. First level of processing starts with melting the solid. The temperature is increased to melt the solid. Then the temperature is slowly decreased.


1139


Due to this the particles of the melted solid arrange themselves locally in a stable ground state. SA theory states that if temperature is lowered sufficiently, the solid will reach thermal equilibrium, which is an optimal state. The state of thermal equilibrium is related to optimal task-machine mapping (optimization). The temperature is the total completion time of a task (cost function) and the change of temperature is the process of dynamicity in mapping tasks and resources. If the next temperature is higher than the previous one it can be accepted with certain probability (though it is worse). This is because the acceptance of some worse states provides a way to break out local optimality which occurs often in local search. V.12. Particle Swarm Algorithm Particle Swarm Optimization (PSO) was proposed by James Kennedy and R. C. Eberhart in 1995 [29]. The algorithm imitates the social behavior of organisms such as bird flocking and fish schooling. PSO algorithm is not only a tool for optimization but also a tool for representing socio cognition of human and artificial agents based on principles of social psychology. PSO as an optimization tool provides a populationbased search procedure in which individuals called particles change their position (state) over time. In a PSO system, particles fly around in a multidimensional search space. During its fly, each particle adjusts its position according to its own experience and according move of neighboring particle. PSO system combines local search methods and global search methods with an expectation to balance exploration and exploitation. V.13. Game Theory The grid scheduling problem is modeled using Game Theory (GT) by L. Young et al [30], which is a commonly used technique to solve economic problems. Each task is modeled as a player whose available strategies are the resources on which the task can run. The payoff for a player is defined as the sum of benefit values for running the task and all communication reaching the task. GT is known as a job scheduling game that models a scenario in which multiple selfish users wish to utilize multiple processing resources. Each user has a single job and the user needs to choose a single machine to process it. The aim of each user is to run his job as fast as possible. Job scheduling problems consist of a set of resources and a set of jobs. Each job is associated with a vector corresponding to its size on each resource (i.e., the processing time of job on a particular resource). Players in the game are related to the jobs in the system. The strategy set of each player is the set of machines. The total load on each resource is the sum of processing times of the jobs that chose the particular resource.


Usually each job tries to minimize the total load on its chosen resource. The standard objective function is to minimize the load of most-loaded resource. V.14. Tabu Search R. Braun et al [20] implemented Tabu Search (TS) in their work, which is about short-hop procedure to find the nearest local minimum solution with in the solution space. TS is a solution space search algorithm that keeps track of the regions of the solution space which have been searched already and TS would not allow to repeat the same [31], [32]. TS is a meta-strategy for guiding known heuristics to overcome local optimality and has now become an established optimization approach that is rapidly spreading to many new fields. The method can be viewed as an iterative technique which explores a set of problem and leads to solutions by repeatedly making moves from one solution to another that are located in the neighborhoods. These moves are performed with the aim of efficiently reaching an optimal solution by minimizing some of the objective functions. V.15. Fuzzy Algorithms Zhou et al. [33] used Fuzzy Logic (FL) techniques to design an adaptive FL scheduler, which utilizes the FL control technology to select the most suitable computing node in the grid environment. A fuzzy set can contain elements having varied degrees of memberships. There is fuzziness in all preemptive scheduling algorithms. A Fuzzy Neural Network was proposed by Yu et al [34] to develop a high-performance scheduling algorithm. The algorithm uses FL techniques to evaluate the grid system load information and adopt the Neural Networks (NN) to tune the membership functions. Artificial Neural Network (ANN) is data-driven modeling tool that could capture and represent complex and non-linear input/output relationships. They are recognized as powerful and general technique for machine learning because of their nonlinear modeling abilities and robustness in handling noise ridden data. Hao et al. [35] presented a grid resource selection based on NN that aims to achieve Quality of Service.

VI.

Investigation of Scheduling Algorithms with Combined Factors and Simulators for Scheduling

VI.1. Modified OLB, Modified MET, Modified MCT, Modified Min-Min, Modified Min-Max It is found by L. Y. Tseng et al. [36] that most of the studies on scheduling algorithms relate the mapping of task and computer by the Expected Time to Complete (ETC) matrix.


1140


In the matrix each row represents the ETC value for the tasks to be completed on computer and each column lists the ETC value for each computer to execute the task. It is stated that the prediction of completion time of each task is difficult, but the ETC matrix is utilized for its heuristic feasibility. The Computer Availability Time (CAT) of each computer is accumulated and is used for finding expected execution time of tasks. They represented it through a matrix comprised of tasks and computers, row represents tasks and column represents resources. The matrix values can be the minimum completion time of the task on the resource. They highlighted three issues upon several observations. First, most studies took an assumption of high speed network communication environment and the communication time is not considered. Second, the importance of each task is not notable. And finally, the deadline for each task is not considered. In their study they have considered factors such as: communication cost, weight and deadline. Pinedo and Chao in 1999 proposed an algorithm called Apparent Tardiness Cost Setups (ATCS) algorithm that satisfied the requirements. But ATCS schedules all the jobs to single resource. It is stated that ATCS can be modified to suite grid environments. The factors considered by the algorithm are Weighted Shortest Processing Time First (WSPTF), Minimum Slack First (MSF), Shortest Setup Time First (SSTF), due date range, due date tightness, set up time severity. ATCS value is calculated by using all the factors said above and the job having maximum ATCS value shall be executed first. In their work, they divided the job into several tasks, since tasks can be dispatched to idle resources that could get better performance in grid environment. Every task has its weight and deadline and they remain the same for all the resources. Deadline is generated randomly between ranges. The Apparent Tardiness Cost Setups-Minimum Completion Time (ACTS-MCT) algorithm is found to be the revised form of ATCS. Followed the ACTS-MCT, they proposed modified OLB, modified MET, modified MCT, modified Min-Min, modified Max-Min. Deadline is added as an extra constraint added to the original forms of the above said algorithms. They compared the proposed algorithms with ACTS-MCT. VI.2. Modified FCFS That Uses Backfilling Sofia K. Dimitriadou et al. [37] proposed Modified FCFS that uses Backfilling and is applied in problems that deal with resource allocation. They stated that the grid jobs that enter the system are parallel jobs called gangs. Job scheduling is done at two levels, viz., grid level and local level. At both the levels the jobs would be competing for the same resources. The goal is to provide services to all provided the local jobs are treated as higher priority jobs and hence their waiting time is minimized. The local jobs are simple and sequential. They require a single resource for execution.


A gang consists of number of tasks that are to be processed simultaneously. Thus each task must be allocated to different resource. The Grid Scheduler (GS) has a queue where the jobs will be stored before they are dispatched. In order for a gang to start execution, all required resources must be available. When FCFS policy is applied, the local jobs would get blocked by the gangs and execution time may be extended. Delay in execution is caused as the availability of idle resources are not utilized. To avoid such a kind of fragmentation, backfilling is applied. Backfilling allows small jobs to initiate before larger jobs (which would wait for the resources that are currently unavailable to execute themselves). Although this scheduling technique dramatically improves utilization, it also requires that all jobs’ service times be known earlier. Stavrinidis et al. [17] stated in their work that predictions can be made with the help of earlier jobs done. Tchernykh et al. [38] said that the poor estimates do not affect the overall performance. The modified FCFS (FCFS and backfilling) allows the smaller jobs to execute to avoid the fragments of idle time. VI.3. Adaptive Hierarchical Scheduling Policy J. H. Abawajy [39] proposed a hierarchical scheduling policy that concentrates on I/O and service demands of parallel jobs in both homogeneous and heterogeneous systems. It is observed that the paper dealt with enterprise grid computing environments. The paper focused on the ability to deploy commodity computational power, network, storage resources and also how to share and manage resources. It is said by Thain et al. [40] that the grid computing schedulers must bring jobs and data in close proximity in order to satisfy throughput, scalability and policy requirements. The algorithm Integrates the job assignment, affinity scheduling, self scheduling approaches to assign resources to parallel jobs. The working of the algorithm speaks about two queues where the unscheduled jobs/tasks are maintained in the first queue (Queue(jobs)) and the second queue is to store the unsatisfied Request For Computation (RFC) (Queue(RFC)). A Boolean function is used to evaluate a particular node is root or leaf. When a non root node is in neutral state, it sends RFC to its parent RFC. If parent is in neutral state while receiving RFC, it generates its RFC and sends to its parent on the next level of the cluster tree. The process is recursively done until the RFC reaches either System Scheduler (SS) or a node with unassigned computation. If RFC reached SS, the request is backlogged and there should not be any job to be scheduled. If SS or a node with unassigned computation is found, space-sharing (job/task assignment) policy is followed by the scheduler to assign jobs. The assignment of job/task components allows a set of ideal jobs/tasks to transfer from parent node to a child node. Then the affinity scheduling is applied. If there is no computation


1141


with affinity condition then transfer rate is set one (from parent to a child node). At last each job is partitioned into task and is assigned to processors on demand, where each processor is allowed to follow time sharing policy. VI.4. Player’s Minimum Completion Time Joanna Kolodziej et al. [41] proposed the algorithm, which is devised for the computational grids and are concerned with high performance computations ensured with the request level security during the data transfer. The proposed work is an improved version of GT model for secure scheduling presented by Joanna Kolodziej et al. [42]. In their work they considered Independent Job Scheduling Problem. Maheswaran et al. [12] proposed a model in which tasks are processed in batch mode. They defined a non co-operative symmetric game model without additional synchronization mechanisms. It is stated that the game scenario is simpler than the one presented by Kwok et al. [43], where an additional mechanism is needed. In the proposed model, the players/users of the game are non co-operative, where the usage privileges are the same for all the users. Each user would be trying to choose an optimal way of matching the task to resources to minimize the total cost of scheduling. The problem of minimizing the game cost consists of two co-operating functional hierarchical units (main unit and subordinate unit). Main unit solves the global level problem of the cost minimization function and the subordinate unit solves the local level problems of the cost minimization function. They proposed modified MCT in the name of Player’s MCT, which is incorporated with the concepts of GT for security assistances. The working of the said algorithm begins as follows; every task is assigned to the machine getting the Earliest Completion Time (ECT). The schedule forms the main unit and is scanned and then minimum completion time is found for tasks of individual user. The scenario of the algorithm states that there can be a population of schedules and ready time of resources in the main unit. There can be many users in the population and each user can have many tasks. Now minimum completion time algorithm is applied to the tasks of individual user and the same is computed for schedules received from main unit until no schedule remain in the main unit. Then the minimum of all MCT is found and sent to the main unit. VI.5. Iterative Divisible Load Theory Monir Abdullah et al. [44] proposed algorithm that aims to design a load distribution model by taking communication and computation time into consideration. It is said that the Divisible Load Theory (DLT) [45] is a powerful tool for modeling data intensive computational problems incorporating communication and computational issues [46]. They said that the load scheduling problem is to


decompose the large data set into smaller data sets across virtual sites. The algorithm is all about finding minimum makespan to schedule a large volume loads within multiple sites. The proposed algorithm started by dividing the large volume of divisible loads using either of the Adaptive Divisible Load Theory (ADLT) models [47], [48], then the makespan is calculated using the cost model defined in the same paper. If all the nodes finish computation at the same time then the current time will be taken as makespan, otherwise the sum of the processing time of the entire nodes is calculated and an average is taken between the sum of the processing time and the number of virtual sites involved. Now the load is calculated and redistributed based on the average in an iterative manner, then the makespan is calculated in the same iterative manner until all the nodes finish its execution. Finally, all nodes will take the new load based on the new averages and for the last node the rest of the load is given without considering the average. Therefore, it is expected that the last node will be finish at the same time as the others. VI.6. Genetic Algorithm Based Integrated Job Scheduling Algorithm Chao-Chin Wu et al. [49] proposed a GA based integrated job scheduling strategy for grid systems that support four different fault tolerance mechanisms. The considered fault tolerance mechanisms are; i) Job Retry (JRT): The JRT mechanism is the simplest fault-tolerance technique, which will re-execute the failed job from the beginning on the same computational node. ii) Job Migration without Check Pointing (JMG): The JMG mechanism will move the failed job to another computational node and re-execute the job from the beginning on the later computational node. iii) Job Migration with Check Pointing (JCP): The JCP mechanism will record the state of the job periodically during run time. If the job fails, it is moved to another computational node and resumed the execution from the last checkpoint. iv) Job Replication (JRP): The JRP mechanism replicates a job to multiple computational nodes. The mechanism is expected to have high success rate. If one of those replicas would have completed, then all other replicas would stop their execution to save the computing power. It is said that each computational site in a grid system supports one of the three mechanisms. In favor to JRP, the scheduler will allocate multiple computational sites to execute a certain job concurrently. Further, the scheduler can execute a certain job by any combination of these four different fault tolerance mechanisms. A new chromosome encoding approach is proposed in the paper. A chromosome is a list of variable length integer sequences. Each integer sequence represents a gene, whose length depends on the type of adopted fault tolerance mechanism. The integer sequence represents the job identity (job


1142


id) and the node identity (node id). Based on the node identities, the job will be moved to the next node for reexecution whenever a failure is encountered. They used Roulette Wheel Selection Method [50] to select chromosomes based on the fitness value. The chromosome with highest fitness value will be the selected for the next generation. They employed cut and splice operator [51] to pick two points in two chromosomes and exchange chromosomes after the points. Since the length of the chromosome is variable in this proposed approach, the mutation operation randomly selects a gene in a chromosome and then mutates its value. VI.7. A Hybrid Scheduling Algorithm to Minimize the Cost With Deadline Constraint Xin Liu et al. [52] proposed the algorithm in which they tried to obtain minimum cost by perturbing the schedule of some tasks from minimum time solution. They proposed min-time algorithm to find the minimum completion time and the min-cost algorithm to find the minimum cost. The proposed algorithm is a hybrid scheduling algorithm to minimize execution time and cost. The tasks stand first in the list follow minimum time algorithm and the remaining tasks follow minimum cost algorithm. This is called as perturbation degree. The proposed algorithm stated that the task from the list is allowed to evaluate the minimum completion time. If it is greater than the deadline, then there is no possibility of getting feasible solution else if the minimum completion time is less than the deadline then binary search is used recursively for getting largest perturbation degree, such that the current or the next perturbation degree is smaller than the deadline. The cost and perturbation degree obtained will be returned as schedule with minimum cost. Hence the task is said to be finished before deadline. VI.8. Balanced Ant Colony Optimization Algorithm Ruay-Shiung Chang et al. [18] stated that the Balanced Ant Colony Optimization Algorithm (BACO) inherits the basic ideas from Ant Colony Optimization (ACO) algorithm to decrease the computation time of jobs executing in Taiwan UniGrid [53] environment. The load of each resource is considered. BACO changes the pheromone density according to the resources status by applying the local and global pheromone update functions. The aim is to minimize the completion time for each job while balancing the system load. They related the ant system to the grid system and the relationships are: An ant in the ant system is a job in the grid system and pheromone value on a path is the weight for a resource in the grid system. A resource with a larger weight value means that the resource has a better computing power. The scheduler collects data from Information Server and uses the data to calculate a weight value of a


resource. The pheromone (weight) of each resource is stored in the scheduler and can be used as the parameters for BACO algorithm. Finally, the scheduler selects a resource by a scheduling algorithm and it sends jobs to the selected resource. It uses the Application Programming Interface (API) of Globus Toolkit [54]. VI.9. Pro-Active Failure Handling Within Scheduling Algorithms B. T. Benjamin Khoo et al. [55] stated pro-active mechanisms in which the failure consideration for the grid is made before the scheduling of a job. They introduced three pro-active failure handling strategies which allow existing scheduling algorithms to be modified to avoid job failures. The strategies are; Site Availability based Allocation (SAA), Node Availability based Allocation (NAA), and Node and Site based Allocation (NSA). They compared pro-active failure mechanism with that of passive failure handling mechanism. They implemented the strategies in Backfill algorithm (BF), Replication Algorithm (REP) and Multi-Resource Scheduling Algorithm (MRS). VI.10.

Time and Cost Improvement Algorithm

Hamid Mohammadi Fard and Hossein Deldari [56] proposed an algorithm called Time and Cost Improvement (TCI) algorithm that presents a new list heuristic combining the greedy approach. The algorithm is devised to optimize both makespan and cost. This algorithm consists of three phases: 1. Level Sorting phase sorts tasks at each level in order to group the tasks that are independent to each other, 2. Task Prioritization phase assigns priority to a task that uses Average Computation Cost (ACC), Data Transfer Cost (DTC) and Rank of Predecessor Task (RPT), 3. Resource selection phase computes the difference between the execution times of various processors and selects the appropriate resource to execute tasks. VI.11.

Contention - Aware Task Duplication Scheduling Algorithm

Oliver Sinnen et al. [57] proposed a contention- aware task duplication scheduling algorithm. The algorithm employed the strategies from both contention model and duplication model. The insertion technique is used in the algorithm for tentative scheduling and to remove redundant tasks, redundant edges. All the predecessors are not duplicated, rather task from which the data arrived at the latest is duplicated and the task is called the critical parent. Also the predecessor of the critical parent is considered for duplication. To know the data ready time of any task, a tentative scheduling of edges on the communication link is done.


1143


VI.12. Opportunistic Algorithm Luiz Meyer et al. [58] proposed a work called as opportunistic algorithm that adopted the observational approach to assign jobs to grid sites. The algorithm concentrated on the dynamic characteristics of grid environment. The algorithm is implemented using virtual data systems. A database is employed to log the location of the site and the status of the work flow. A site selector program is responsible for choosing the execution site for a job based on the performance of the site in its previous execution. Round Robin technique is followed among sites when no job is completed for a longer time. Another database is used to track the submission and execution of jobs. The algorithm uses a queue monitor called MonitQueue to monitor and remove the jobs that could not perform well. VI.13. Multi-Site Scheduling Algorithm Weizhe Zhang et al. [59] proposed a work in the name of multi-site scheduling algorithm. In which they used the first come first served policy for priority strategy and backfilling to improve system utilization. They stated that the decision-making in mapping the tasks to resources increases the complexity in a heterogeneous environment. It is found by the authors that the execution time of a job is reduced by multi-site reservation i.e., a job is reserved in order to acquire better execution time even if it starts running immediately. They proposed two resource selection algorithms namely, (i) optimal adaptive multi-site resource selection, (ii) greedy adaptive multisite resource selection. Both the algorithms are adaptive to variety of resources. In the resource selection, when job execution time is longer than the time required for it, then the job is reserved for the resource. VI.14.

Task Duplication Based Efficient Multi-Objective Scheduling

D. I. George Amalarethinam et al. proposed a multi objective algorithm called Task Duplication Based Efficient Multi-objective Scheduling (TDB-EMOS) [60]. The factors considered are makespan, resource utilization and communication time depending on the precedence constraint. This algorithm uses the duplication mechanism as its key component and is an extension of EDOS [61]. It starts working with the duplication of the parent task with respect to the number of children. Intermediate task duplication is done for tasks that have maximum number of tasks to travel to reach the exit task. It is found that the communication time is reduced after duplication. VI.15. Throughput Constrained Latency Optimization Vydyanathan et al. proposed an algorithm called throughput constrained latency optimization (TCLO) [62]. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

The algorithm is focusing on latency and throughput. The algorithm employed task duplication mechanism and is designed for a k-port communication model. The presented algorithm is an extended work of the authors’ previous work [63]. As an extension, the authors adapted the heuristics for more realistic k-port communication model and they concentrated on the communication contention while estimating the latency and throughput. This algorithm is designed to generate two schedules for non-pipelined and pipelined schedules respectively. The non-pipelined schedule is produced out of priority based list scheduling heuristic. The pipelined schedule is done in three phases. In the first phase, the schedule that meets the throughput requirement is obtained using the satisfying throughput heuristic and is assumed that the processors availability is unbound. In the second phase, the processor reduction heuristic is used and the limited number of processors is utilized. The third phase focused on minimizing the latency through latency minimization heuristic. VI.16.

Avoiding Useless Duplication

K. Asakurae et al. proposed an algorithm that emphasis on avoiding useless duplication [64]. This algorithm used limited number of idle processors. It is said that the algorithm follow traditional Bounded Number of Processors (BNP). The algorithm consists of three phases namely: (i) BNP scheduling phase, (ii) Task fill phase, (iii) Task duplication phase. The algorithm uses minimum number of processors to map the available tasks. The idle processors are used for duplication. The depth first search is adapted to select the tasks to be duplicated. The traversal is done from tail to head task(s). The head task is preferred to duplicate only when the listed conditions are obeyed by the task. The task in best predecessor group should not be duplicated to execute in any idle time slots that occurs between the tasks execution time of tasks from best predecessor group. VI.17.

Pw-Tabu Algorithm

Dang Minh Quan et al. proposed an algorithm entitled pw-Tabu Algorithm [65]. The algorithm takes the workflow and grid resources as its input parameters. A set of reference configuration is created after the jobs resource requirements are matched with the resource configuration of the Resource Management System (RMS). The RMS and jobs are monitored to generate a reference solution set. A parallel algorithm will be reducing the execution time of each solution in the reference set. The configurations are distributed to several nodes for parallelism. The tasks that could improve the quality of reference configuration set are divided into subtasks and processed in parallel manner. Finally, the improved solutions are collected to find the best solution.


1144


The overall processing is done in master-slave processing mode, where the master sends data and the slave process. VI.18. Local Node Fault Recovery Mechanism Suchang Guo et al. proposed a mechanism called the Local Node Fault Recovery (LNFR) [66], which concentrates on service reliability and fault recovery in grid scheduling. A greater amount of resource and time is wasted when an execution is terminated by a node on failure. They said that LNFR helps in resuming the execution from failure and saves the migration expense. Distributed management of fault tolerance can be achieved when fault recovery models are available at grid resources pool. The authors divide the nature of resource failures into recoverable and unrecoverable failures. In recoverable failure may cause interruption on execution of tasks and can be resumed by using the state information of the task. In the unrecoverable failures the execution of the tasks will be terminated. They classified as hardware failure, software failure and communication link failure. VI.19.

Dynamic Constraint Algorithm

Prodan R. et al. Dynamic Constraint criterion scheduling programming. It uses knapsack problems. VI.20.

presented an algorithm called Algorithm [67], a general biheuristic based on dynamic the principle of multiple choice

Data (MIMD), superscalar, pipeline, dataflow and vectors. The algorithm describes the reconfigurable characteristics. The reconfigurable characteristics are described as the requirement expectation on communication type that the topology establishes between tasks. The heterogeneous characteristics rely on tasks of the DAG (ie., computation) and the reconfigurable characteristics rely on the edges (ie., communication). Matching degree is the execution speed of task running on a processing unit. The heterogeneity matching matrix consists of values of matching degree between task and processing unit. The re-configurability coupling matrix contains the values of communication capability between different interconnections. Each value is called as coupling degree and is calculated using communication workload, bandwidth of the network, actual communication time. The coupling degree is used to configure the interconnection. The algorithm uses heterogeneity matching matrix and reconfigurability coupling matrix to schedule tasks. VI.22. Genetic Algorithm Based Scheduling Algorithm for Parallel Heterogeneous Multiprocessor System Mohammed Sadeq Garshasbi and Mehdi Effatparvar proposed a GA based scheduling algorithm for parallel heterogeneous multiprocessor system [70]. The algorithm aims at load balancing among multi-processor systems. The fitness function is calculated using the total response time of tasks in graph with respect to the available processors. VI.23.

Duplication Based Bottom-Up Scheduling Algorithm

A Scheduling Model Based on Ant Colony Optimization Technique

Doruk Bozdag et al. presented an algorithm called Duplication based Bottom-Up Scheduling Algorithm (DBUS) [68]. The traversal of Directed Acyclic Graph (DAG) is done in bottom-up fashion. It is said that the algorithm limits the number of duplications. The proposed algorithm is an insertion based algorithm that allows the tasks to schedule at the first available time slot that matches the available resources. The algorithm consists of two phases: (i) Listing phase that uses critical path based prioritization to assign tasks to nodes, (ii) Scheduling phase in which the tasks are scheduled in bottom up manner formed by the listing phase. The task is scheduled to a resource based on its latest start time and the resource cannot be further scheduled for task’s children.

Saeed Molaiy and Mehdi Effatparvar proposed a scheduling model based on Ant Colony Optimization Technique [71]. The scheduling strategy is found to be the time dependency between task execution and communication. The number of ants in the model is related to the loading status of the resource. In the proposed model each node is expected to have information about the other nodes. The model is fully distributed. The nodes and the resources are found to be autonomous. The ant that passes/traverses any node on its path gets an update of node’s information. When the control is transferred from communication mode to computation mode and the relative updates are done in the ant table and the node table. Values from the ant table and the node table are used for scheduling.

VI.21. Scheduling Algorithm for Heterogeneous and Reconfigurable Computing

VI.24. Dynamic Cluster Based Job Scheduling

Yiming Tan et al. proposed a scheduling algorithm for heterogeneous and reconfigurable computing [69]. The heterogeneity of tasks is identified as Single Instruction Multiple Data (SIMD), Multiple Instruction Multiple

Reza Fotohi and Mehdi Effatparvar proposed a dynamic cluster based job scheduling for grid computing environment [72]. The proposed method sorts the jobs based on the burst time. The jobs are distributed between



1145


clusters according to the range. The time quantum is calculated for CPUs, which is square of CPU burst time. The resources belonging to the cluster having lower quantum time value is chosen for execution. VI.25.

Multiple Priority Queues Genetic Algorithm

Yuming Xu et al. proposed an algorithm called Multiple Priority Queues Genetic Algorithm [73]. The algorithm uses Genetic Algorithm to assign priority to each subtask. Earliest Finish Time (list heuristics) is employed to search a process to map the task for execution. Roulette Wheel selection method is used to select population for next generation. Heuristic crossover, heuristic mutation and elitism are used to assign priority for the tasks. Heterogeneous Earliest Finish Time (HEFT) [74] algorithm is used to find the priority of resource to execute the task. HEFT is a list and rank based scheduling algorithm. VI.26.

Earliest Deadline First with Shortest Remaining Time

Dipti Sharma and Pradeep Mittal proposed an algorithm called Earliest Deadline First with Shortest Remaining Time [75]. Scheduling is done based on the deadline of jobs and remaining time. The job having smallest earliest deadline can be given priority to get the resources. It is observed that the algorithm works under pre-emptive mode. When a job with smallest deadline enters the system than the job that is executed then the current execution is stopped. The processor would be assigned to execute the job with smallest deadline. When the processors are not satisfying the deadline constraints then the job will get the processor that has shortest remaining time. VI.27. Bi-Criteria Scheduling Algorithm P. Keerthika and N. Kasthuri proposed Bi-criteria Scheduling Algorithm (BSA) that focused on pro-active fault tolerance mechanism [76]. The algorithm ensures that the job dispatched for execution would not get failed. Job is allocated based on fitness value, which calculated using the completion time of jobs and failure rate of resources. VI.28. Grid Computing Life in Indian Terrain Sanjeev Puri and Harsh Dev discussed grid computing life in Indian terrain [77]. Several issues upon handling large number of users, handling plenty of information in grid environment are discussed. Working principles of cross certification that supports trusted sharing of resources and ensures confidentiality and privacy across domains are discussed as core components of the paper. The importance of distributed authorization and


domain credentials are also listed. Challenges prevail in ensuring security in grid computing is seriously discussed the paper. VI.29.

Cuckoo Optimization Algorithm

Rajabioun proposed Cuckoo optimization algorithm [78]. The cuckoo search is to find the optimized nest for the eggs of the cuckoo birds. The nature’s behavior is related to the scheduling problem. Generally, cuckoo seeks other birds nest for hatching their eggs. During this search, cuckoo may destroy the eggs of the other birds to increase the hatching rate of their own. This is related to that of searching a best resource to allocate jobs in a job allocation problem. VI.30.

Double Auction-Inspired Meta-Scheduling Mechanism

Saurabh Kumar Garg et al. proposed a double auctioninspired meta-scheduling mechanism for scheduling parallel applications in grid environment [79]. This mechanism is proposed as inspired from the auction principles for resource allocation. The mechanism is a sequence of three stages, (i) gathering stage, (ii) valuation stage and (iii) matching stage. During the first stage, information about resources and jobs are gathered. The information can be slot availability in the queue, wait time, etc. The resource valuation and job valuation are done in the second phase. The resource valuation is done based on queue load, demand and supply of the resources, etc. The job valuation is done based on CPUs need, deadline and run time. The matching between the jobs and resources is done during the third phase. The matching between job and resource is done based on valuations arrived in the previous phase. If a job is not getting a proper resource then it is retained to schedule in the next cycle. The algorithm uses Double Auction (call auction), where the sellers submit their offers (asks) and buyers submit their requests (bids) to the auctioneer. In the matching process, auctioneer decides the resource to execute the job. Call auctions would be best if the resource allocation can be continued for a long time. Call auctions can be used when processors are found as substitutes for one another. The mechanism relates the queue slots to asks and jobs to bids and behaves like a call auction metascheduler. In the call auction, the matching is happen between the maximum bid and the minimum ask, i.e., the earliest deadline job is matched to the fastest queue that has sufficient processors. VI.31. Power-Aware Task Clustering Algorithm Lizhe Wang proposed power-aware task clustering algorithm [80]. The aim is to reduce the power consumption during parallel tasks execution. Dynamic Voltage Frequency Scaling (DVFS) technique is used to develop the algorithm. DVFS technique is proven to be


1146


feasible to cut down the processor power consumption [81]. The power aware task clustering algorithm begins by marking all edges as unvisited and each task is considered as individual cluster. Then it sorts the edges with respect to communication time in non-increasing order. The edge with high communication cost is made zero if the power consumption is not increased and the edge is marked examined. The clusters are merged and ordered with respect to the longest path from the particular job to the exit job. The algorithm is repeated until no edge remains unexamined. VI.32.

Energy Efficient Elastic Scheduling

Xiaomin Liu et al. proposed Energy Efficient Elastic Scheduling strategy for independent tasks [82]. The algorithm is found very flexible to the system work load. When the system load is heavy, the supply voltages of new task and the tasks in the local queues can be increased. The voltages get reduced if system load is insignificant. The algorithm begins by calculating the start and finish time of each task on each computing node at the lowest voltage. If the computing nodes are not meeting the user’s expectations with the set voltage, then the voltage is raised to meet the user’s expected finish time. If the increased voltage is not able to meet the constraints of new task then the task from a local queue is given priority. If the new task meets the constraints then it is allocated to the computing node. VI.33.

Community Aware Scheduling Algorithm

Ye Huang et al. proposed an algorithm called Community Aware Scheduling Algorithm (CASA) [83]. It consists of two phases called job submission phase and dynamic scheduling phase. During the job submission phase, the node that gets a job starts behaving as a requester node and generates a message request for the submitted job. The message request may include the information like required processing elements, required kind of operating system and estimated execution time. The request message is forwarded to discover the remote nodes in the decentralized environment for job delegation. The nodes that receive the delegation request may accept jobs are called responder nodes. Among the responder nodes, the node having more processing elements, faster processing speed is chosen for job execution. The second phase keeps the previous scheduling decisions to be optimized by adjusting the scheduler according to the changes in the computing environment. VI.34.

Adaptive Energy Efficient Scheduling Algorithm

Wei Liu et al. proposed adaptive energy efficient scheduling algorithm [84]. The algorithm combines Dynamic Voltage Scaling (DVS) and adaptive task duplication technique. The algorithm has two phases. The first phase consists of an adaptive threshold-based


task duplication strategy that selectively replicates the predecessor of a task. During the second phase, an effective measure is taken to schedule tasks on DVSenabled processors to reduce processor energy. This avoids the slack time due to tasks dependency. VI.35. Bacterial Foraging Optimization Technique Based Hyper-Heuristic Resource Scheduling Algorithm Rajni and Inderveer Chana proposed a hyper-heuristic resource scheduling algorithm [85]. The algorithm uses bacterial foraging optimization technique [86]. A hyperheuristic is a generic, fast method that can produce acceptable quality solutions. The bacterial foraging is a process in which the bacteria moves in search of food in a region and decides to get in to the region or to move to another region to get better food. The process comprised of four mechanisms namely: (i) chemotaxis, the process of simulating the movement of bacteria, (ii) swarming, the process of repelling the nearby cells as the particular location could not have more than one cell (generally, cells could be departed at specific distance), (iii) reproduction, after certain number of chemotaxis, a reproduction step is done. The fitness value of the bacteria is used for reproduction. The least value bacteria would die and the most valued bacteria is found to be healthier and splits into two cells (asexual reproduction), (iv) elimination and dispersal, elimination may happen due to rise in temperature or a part of the bacteria may move suddenly to some other regions. Dispersal may make the bacteria to move to good sources of food that are very close to the region. The proposed algorithm begins by collecting a resource list then it initializes the job list. After this a random feasible solution is obtained. Each bacterium is a hyper-heuristic agent which will be having an initial solution and an access to the evaluation function. At each decision a low level heuristic is chosen and the fitness function is calculated using the above said mechanisms. The swim of bacteria results in a best schedule that reduces the cost. VI.36.

Besom Scheduling Algorithm

Sucha Smanchat et al. proposed an algorithm called Besom scheduling algorithm [87]. The proposed algorithm is a scheduling technique for parameter sweep workflow. It is said that a parameter sweep workflow is executed numerous times with different input parameters to determine the influence of the parameters as individuals and in combinations. The algorithm is of three phases namely, (i) instance generation, (ii) task prioritizing, (iii) resources selection. During the instance generation phase, the task instances in the work flow are added to the unscheduled task list, an instance of the parameter is used for combining parameters. During the


1147


task prioritizing phase, the available grid resources are updated. The dependent tasks information is updated only if parents finished their executions. These dependent tasks are executed based on dependencies in parameter sweep workflow. In the resource selection phase, the tasks prioritized by the previous phase are allocated to the sources. These steps would be executed until no more tasks remain unscheduled. VI.37.

Optimized Hierarchical Load Balancing Algorithm

Ramya and Shalini Thomas proposed an algorithm called Optimized Hierarchical Load Balancing Algorithm [88]. The algorithm uses the clustering concept for grouping the resources and is found to be an optimized algorithm for computational grid. The algorithm consists of three phases namely; (i) resource discovery, (ii) resource allocation and load analysis, (iii) job execution. During the first phase, the available resources are identified and filtered based on the average computational power of clusters. In the second phase, the mapping of job and resource is done and is carried out by checking the average load of the cluster and resources. Then, the cluster having fastest average computing power is chosen by the scheduler to assign the job. During the third phase, mapped jobs are executed on resources. The average load of a resource is calculated using the factors like CPU utilization of resource, utilization of the resource and the utilization of the network. A balance threshold is set to check if the cluster is overloaded or under loaded. When a cluster is found overloaded then it is not considered for the current allocation. Finally, the updates are done locally and globally to get the latest status of resources. VI.38. Scheduling Based on Genetic Algorithm That Uses Tournament Selection Method to Get Initial Population

VI.39.

Mohammad BSoul et al proposed a simulator called MICOSim [90] for scheduling jobs in grid computing environment. The simulator consists of four basic interacting components namely: (i) the system, (ii) the entity, (iii) the entity strategy, (iv) the scenario. The system manages the entities and it has a vector can could contain the events to be executed in an order and occurrence of time. An entity can send and receive jobs and bids. The entity could be instances like users, resources and brokers. The entity strategy deals with user strategy, broker strategy and resource strategy. User strategy includes parameters like deadlines, price, etc. Broker strategy includes parameters like waiting time, decision, etc. The resource strategy includes parameters like resource utilization price, possible completion time of jobs, expiry dates, etc. Finally, the scenario deals with characteristics like kind of job, length of job, number and speed of resources, number of brokers, number of users, etc. VI.40.


Job Scheduling Algorithm Based on Cuckoo Optimization Technique

Maryam Rabiee and Hedieh Sajedi proposed a job scheduling algorithm that uses cuckoo optimization technique [91]. The fitness function is evaluated as the sum of all minimum execution time of tasks on resources along with a penalty co-efficient. When more than one resource try to execute an operation in a parallel fashion the value of fitness function will increase due to an increase in penalty co-efficient. In order to prevent the said situation the fitness function is calculated as the maximum of sum of the finish time and penalty coefficient. VI.41.

Anis Gharbi et al. proposed a genetic algorithm [89] for scheduling parallel jobs to resources. The initial population is generated by choosing jobs to be executed on resources which have the smallest execution time. The parent population is selected based on tournament selection method. It is stated that three chromosomes are randomly chosen from the population and one is finalized based on its fitness. One point cross over operator is applied and genes are exchanged. Mutation operation is applied for every gene in the chromosome. A list of five heuristics is implemented as mutation operators. The feasibility is checked after mutation. A few replacements are encouraged at a time and the children are allowed to compete with the parents to get the best candidates. The authors implemented this algorithm in a real world problem. The working principle of the algorithm is used to schedule complicated tasks in a production company that performs complex operations in its workshop.

Micosim, a Simulator to Schedule Jobs

Popular File Replicate First Algorithm

Lee et al. proposed an algorithm called Popular File Replicate First algorithm [92]. The algorithm uses an adaptive data replication technique. The algorithm is developed for star topology data grid. It is stated that the file access behaviors like geometric, uniform distributions are considered to find the correctness of the algorithm. The algorithm is devised to track the access behaviors of users and to replicate the popularly used files. VI.42.

Job Data Scheduling based on Bee Colony Optimization

Taheri et al. proposed an algorithm called Job Data Scheduling that uses Bee Colony Optimization [93]. The algorithm aims to schedule jobs to resources to reduce makespan and to replicate data sets on storage resources to reduce the transfer time of files containing data. It is stated that the algorithm is found fit in making decisions to achieve efficiency.


1148


Task Duplication

Heterogenic Network Behavior

Grid Type

Data

Memory To Store Data

Other Factors

-

-

-

-

Deadline, Tardiness

-

-

-

-

-

Gang Scheduling, Wait Time

-

Y

Enterprise Grid

Y

Y

I/O Requests, Load Balancing

-

Y -

Computational Grid Data Intensive

Y

Y

-

-

Computational Grid

-

-

Data Intensive Computational Grid

Not Readily Available

Y

-

-

-

-

-

-

-

-

Y

Y

Communication Intensive

Security Assurance Fault Tolerance Mechanism, Job Failure Lack Of Wavelength Bad Resources, Band Width Job Failure, Backfilling Greedy Approach

-

-

Network Topology

Y

Scheduling Tasks Across Grid Sites (Multisites)

-

FCFS, Backfilling, Greedy Approach

Y

VI.3

-

-

VI.4 VI.5

Y -

Y

VI.6

-

Y

VI.7

Y

Y

Affinity Scheduling, Self Scheduling Job Migration Without Check Pointing -

VI.8

Y

-

-

-

Y

-

Y

-

Y

VI.10

Y

Y

-

VI.11

-

Y

-

VI.12

-

-

-

-

-

VI.13

-

Y

-

-

Y

VI.14

-

Y

-

Y

-

-

-

-

VI.15

-

Y

-

Y

-

Data Intensive

Data Transfer Rate

-

VI.16

-

Y

-

Y

-

-

-

-

VI.17

Y

Y

-

-

Y

Data And Computation Intensive

Message (Data) Configuration

-

VI.18

Y

Y

Fault Recovery on Job Migration

-

Y

Service Grid

-

-

VI.19 VI.20

Y -

Y Y

-

Y

-

Scientific Grid -

-

-

VI.9

VI.43.

Job Migration

VI.2

Rescheduling

Y Conside red Message Passing

VI.1

Negligible

Computation/ Communication Time

-

Computation/ Communication Cost

Algorithms Referred from this section

TABLE II COMPARISON OF VARIOUS FACTORS IN SCHEDULING ALGORITHMS

-

Uses Third Party Transfer To Send Files Considers Data Transportation With Computational Grid Respect To Bandwidth And Latency Scientific Grid

Pre-Fetching Based Dynamic Data Replication Algorithm

VI.44.

Saadat presented an algorithm called pre-fetching based Dynamic Data Replication algorithm [94]. It is given that it is a prediction based algorithm. The algorithm predicts the future requirement and replicates the data at an earlier time even before it is requested. It is said that the prediction is done based on the history of file accessed status.


-

Rank , Insertion Approach Communication Contention, Homogeneous Processors Limited Number of Processors Master-Slave Process, Resource Management System, Tabu Search Ant Colony Optimization, Resource Management System Greedy Approach Insertion Based

Modified Dynamic Hierarchical Replication Algorithm

N. Mansouri and Gh. Dastghaibyfard proposed an algorithm called Modified Dynamic Hierarchical Replication Algorithm [95]. This is an improved version of Dynamic Hierarchical Replication Algorithm [96]. The Dynamic Hierarchical Algorithm replicates the data and stores in the sites which accessed the particular data sets very frequently. Replicating the data set in multiple sites is not encouraged.


1149


The Modified Dynamic Hierarchical Replication algorithm states the factors to choose the best replica. Choosing a replica is described as the first part of the algorithm. The factors to be considered while choosing a replica are storage access latency, distance between the nodes and processing capability of the CPU. The second part of the algorithm is replica placement, which ensures the storage of replica in the best site than in all requested sites. The third part of the algorithm is replica management, which aims at providing the storage space for the file selected for replication. If space is found then the replica is stored, otherwise existing replica files would be deleted based on some constraints described as rules. Hence the replica can be stored. The algorithm works towards reducing cost and execution time, effective usage of network and storage resource, maintaining distance information, etc. VI.45.

VI.48.

Lindberg et al. [100] introduced a set of eight heuristics to minimize the schedule length and energy utilization to manage memory requirements of tasks. The heuristics include evolutionary algorithms and list based scheduling algorithm. The algorithms discussed are analyzed based on some factors and a few observations are given in Table II.

VII. Conclusion Investigations are made on various factors that influence scheduling in grid system. This is an effort made to find and present the state of art scheduling policies, algorithms and also to illustrate how they are combined with other ideas to provide an extensive work in the related area.

Qos Aware Dynamic Replication Algorithm

Andronikou et al proposed an algorithm called QoS aware dynamic replication algorithm [97]. The algorithm considers the infrastructure and need of data. It is stated that the algorithm is scalable and supports fast execution. The number of replicas required is found based on data request, the content of the data and the QoS expected. It increases and decreases the replicas based on the dynamic behavior of the grid system.

References [1]

[2]

[3]

[4]

VI.46.

Eight Heuristics and List Based Scheduling Algorithm

Novel Adaptive Decentralized Job Scheduling

Kamalam and Murali Baskaran proposed a scheduling algorithm called Novel Adaptive Decentralized Job Scheduling [98]. The algorithm uses Divisible Load Theory and Least Cost Method. The algorithm aims to provide maximum and effective resource utilization. They stated that the entire grid system is treated as a collection of clusters and each cluster is a collection of computing resources. The cluster server is called as co-coordinator node and others are called worker nodes. The coordinator node finds the worker node having minimum completion time to complete a job. The loads of the clusters are monitored dynamically to achieve load balancing.

[5]

[6]

[7]

[8]

[9]

[10]

VI.47.

Fault Tolerant Min-Min Algorithm

Keerthika and Kasthuri proposed an algorithm called Fault Tolerant Min-Min algorithm for scheduling static tasks [99]. They consider the following factors that may lead to execution failures. Execution failures may happen due network failure, unavailability of necessary software to complete execution, resources overload, etc. This algorithm is devised to handle fault and to prevent failures. The algorithm calculates the fault rate of all resource before scheduling tasks. It is stated that resources are not encouraged to be idle for a long time.


[11]

[12]

[13]

[14]

I Foster, C Kesselman, S Tuecke, The Anatomy of the Grid: Enabling Scalable Virtual organizations, International Journal of Supercomputer Applications, 2001. Hwang K., Advanced Computer Architecture: Parallelism, Scalability, and Programmability, Mc. Graw Hill, Inc., New York, 1993. D P Anderson, J Cobb, E Korpela, M Lebofsky, D Werthimer, Seti@home: An Experiment in Public-Resource Computing, Communications of the ACM, 2002. Fatos Xhafa, Ajith Abraham, Computational Models and Heuristic Methods for Grid Scheduling Problems, Future Generation Computer Systems, 2010. T Casavant and J Kuhl, A Taxonomy of Scheduling in General Purpose Distributed Computing Systems, IEEE Transactions on Software Engineering, 1988. M Arora, S K Das, R Biswas, A Decentralized Scheduling and Load Balancing Algorithm for Heterogeneous Grid Environments, Proceedings of International Conference on Parallel Processing, 2002. L Lee, C Liang, H Chang, An Adaptive Task Scheduling System For Grid Computing, Proceedings of the 6th IEEE international Conference on Computer and information Technology, 2006. Yinglong Xia, Viktor K Prasanna, James Li, Hierarchical Scheduling of DAG Structured Computations on Many Core Processors with Dynamic Thread Grouping, Job Scheduling strategies for parallel processing, Lecture Notes in Computer Science, 2010. J Yu, R Buyya and K Ramamohanrao, Workflow Scheduling Algorithms for Grid Computing, Metaheuristics for scheduling in distributed computing environments, Springer link, 2008. S Sahni, Scheduling Master-Slave Multiprocessor Systems, IEEE Transactions on Computers, 1996. U Schwiegelshohn, R Yahyapour, Analysis Of First-Come-FirstServe Parallel Job Scheduling, Proceedings of the 9th SIAM Symposium on Discrete Algorithms, 1998. M Maheswaran, S Ali, H J Siegel, D Hensgen and R F Freund, Dynamic Matching and Scheduling of a Class of Independent Tasks onto Heterogeneous Computing Systems, Journal of Parallel and Distributed Computing, 1999. R Armstrong, D Hensgen, T Kidd, The Relative Performance of Various Mapping Algorithms is Independent of Sizable Variances In Run-Time Predictions, 7th IEEE Heterogeneous Computing Workshop, 1998. R F Freund, M Gherrity, S Ambrosius, M Campbell, M Halderman, D Hensgen, E Keith,T Kidd, M Kussow, J D Lima, F Mirabile, L Moore, B Rust, and H J Siegel, Scheduling Resources in Multi-User, Heterogeneous, Computing


1150


[15] [16]

[17]

[18]

[19]

[20]

[21]

[22] [23] [24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

[32] [33]

[34]

[35]

[36]

Environments with Smartnet, 7th IEEE Heterogeneous Computing Workshop, 1998. R F Freund and H J Siegel, Heterogeneous Processing, IEEE Transactions on Computers, 1993. Ahuva W Mualem, Dror G Feitelson, Utilization, Predictability, Workloads, and User Runtime Estimates in Scheduling the IBM SP2 with Backfilling. IEEE Transactions on Parallel and Distributed Systems, 2001. Stavrinidis G, H D Karatza, Performance Evaluation of Gang Scheduling in Distributed Real-Time Systems with Possible Software Faults, Proceedings of the International Symposium on Performance Evaluation of Computer and Telecommunication Systems, 2008. Ruay Shiung Chang, Jih Sheng Chang, Po Sheng Lin, An Ant Algorithm for Balanced Job Scheduling in Grids, Future Generation Computer Systems, 2009. C L Liu and J W Layland, Scheduling Algorithm for Multi Programming in a Hard Real Time Environment, Journal of the Association for Computing Machinery, 1973. R Braun, H Siegel, N Beck, L Boloni, M Maheswaran, A Reuther, J Robertson, M Theys, B Yao, D Hensgen, R Freund, A Comparison of Eleven Static Heuristics for Mapping a Class of Independent Tasks onto Heterogeneous Distributed Computing Systems, Journal of Parallel and Distributed Computing, 2001. O H Ibarra, C E Kim, Heuristic Algorithms for Scheduling Independent Tasks on Non-Identical Processors, Journal of Association of Computer Machines, 1977. M Dorigo, C Blum, Ant Colony Optimization Theory: A Survey, Theoretical Computer Science, 2005. M Dorigo, Ant Colony Optimization, http://www.acometaheuristic.org. M Dorigo, L M Gambardella, Ant Colony System: A Cooperative Learning Approach to the Traveling Salesman Problem, IEEE Transactions on Evolutionary Computation, 1997. E Salari, K Eshghi, An ACO Algorithm for Graph Coloring Problem, Congress on Computational Intelligence Methods and Applications, 2005. Xiaoxia Zhang, Lixin Tang, CT-ACO-Hybridizing Ant Colony Optimization with Cycle Transfer Search for the Vehicle Routing Problem, Congress on Computational Intelligence Methods and Applications, 2005. R Prodan, T Fahringer, Dynamic Scheduling of Scientific Workflow Applications on the Grid using a Modular Optimisation Tool: A Case Study, The 20th Symposium of Applied Computing, 2005. Y. Liu, Survey on Grid Scheduling, http://www.cs.uiowa.edu/~yanliu, (for Ph.D Qualifying Exam), 2004. Kennedy J, Eberhart R, Particle swarm optimization, Proceedings of IEEE International Conference on Neural Networks, 1995. L Young, S McGough, S Newhouse, and J Darlington, Scheduling Architecture and Algorithms within the ICENI Grid Middleware, Proceedings of UK e-Science All Hands Meeting, 2003. I De Falco, R Del Balio, E Tarantino, R Vaccaro, Improving Search by Incorporating Evolution Principles in Parallel Tabu Search, IEEE Conference on Evolutionary Computation, 1994. F Glover and M Laguna, Tabu Search, Kluwer Academic, 1997. Jiayi Zhou, Kun-Ming Yu, Chih-Hsun Chou, Li-An Yang, ZhiJie Luo, A Dynamic Resource Broker and Fuzzy Logic Based Scheduling Algorithm in Grid Environment, 8th International Conference Proceeding of Adaptive and Natural Computing Algorithms, 2007. K M Yu, Zh J Luo, Ch H Chou, Ch K. Chen, J Zhou, A Fuzzy Neural Network Based Scheduling Algorithm for Job Assignment on Computational Grids, Lecture Notes in Computer Science, Springer, 2007. Xianwen Hao, Yu Dai, Bin Zhang, Tingwei Chen, Lei Yang, QoS-Driven Grid Resource Selection Based on Novel Neural Networks, First International Conference proceeding of Advances in Grid and Pervasive Computing, 2006. Li-Ya Tseng, Yeh-Hao Chin, Shu-Ching Wang, A Minimized


[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44]

[45]

[46]

[47]

[48]

[49]

[50] [51]

[52]

[53] [54] [55]

[56]

[57]

[58]

Makespan Scheduler with Multiple Factors for Grid Computing Systems, Expert Systems with Applications, 2009. Sofia K Dimitriadou, Helen D Karatza, Multi-Site Allocation Policies on a Grid and Local Level, Electronic Notes in Theoretical Computer Science, 2010. Tchernykh, A, Ramirez J M, Avetisyan A, Kuzjurin N, Grushin D, Zhuk S, Two Level Job-Scheduling Strategies for a Computational Grid, Parallel Processing and Applied Mathematics, Springer Verlag, 2006. J H Abawajy, Adaptive Hierarchical Scheduling Policy for Enterprise Grid Computing Systems, Journal of Network and Computer Applications, 2009. Thain D, Bent J, Arpaci Dusseau A, Arpaci Dusseau R, Livny M. Gathering at the Well: Creating Communities for Grid I/O, Proceedings of Super Computing, 2001. Joanna Kolodziej, Fatos Xhafa, Meeting Security and User Behavior Requirements in Grid Scheduling, Simulation Modeling Practices and Theory, 2010. Joanna Kolodziej, Fatos Xhafa, A Game Theoretic and Hybrid Genetic Meta-heuristic Model for Security Assured Scheduling of Independent Jobs in Computational Grids, Proceedings of International Conference on Complex, Intelligent and Software Intensive Systems, 2010. Kwok Y K., Hwang K., and Song S, Selfish Grids: GameTheoretic Modeling and NAS/PSA Benchmark Evaluation, IEEE Transactions on Parallel and Distributing Systems, 2007. Monir Abdullah, Mohamed Othman, Hamidah Ibrahim, Shamala Subramaniam, Optimal Workload Allocation Model for Scheduling Divisible Data Grid Applications, Future Generation Computer Systems, 2010. Y C Cheng, T G Robertazzi, Distributed Computation with Communication Delays, IEEE Transactions on Aerospace and Electronic Systems, 1988. S Kim, J B Weissman, A Genetic Algorithm Based Approach for Scheduling Decomposable Data Grid Applications, IEEE Proceeding of the International Conference on Parallel Processing, 2004. Monir Abdullah, Mohamed Othman, Hamidah Ibrahim, Shamala Subramaniam, Adaptive Divisible Load Model for Scheduling Data-intensive Grid Applications in Computational Science, Lecture Notes in Computer Science, Springer Verlag, 2007. Monir Abdullah, Mohamed Othman, Hamidah Ibrahim, Shamala Subramaniam, A2DLT: Divisible Load Balancing Model for Scheduling Communication-intensive Grid Applications, Lecture Notes in Computer Science, Springer Verlag, 2008. Chao Chin Wu, Ren Yi Sun, An Integrated Security-aware Job Scheduling Strategy for Large-scale Computational Grids, Future Generation Computer Systems, 2010. M Srinivas, L M Patnaik, Genetic Algorithms: A Survey, IEEE Transactions on Computers, 1994. Song S, Kwok Y K, Hwang K, Risk-resilient Heuristics and Genetic Algorithms for Security-assured Grid Job Scheduling, IEEE Transactions on Computers, 2006. Xin Liu, Chunming Qiao, Wei Wei, Xiang Yu, Ting Wang, Weisheng Hu, Wei Guo, and Min-You Wu, Task Scheduling and Light Path Establishment in Optical Grids, Journal of Light Wave Technology, 2009. Taiwan Unigrid project portal site, http://www.unigrid.org.tw. GlobusToolkit-v4, http://www.globus.org/toolkit/downloads/4.0.4/. B.T. Benjamin Khoo, Bharadwaj Veeravalli, Pro-active Failure Handling Mechanisms for Scheduling in Grid Computing Environments, Journal of Parallel Distributed Computing, 2010. Hamid Mohammadi Fard,Hossein Deldari, An Economic Approach for Scheduling Dependent Tasks in Grid Computing, The 11th IEEE International Conference on Computational Science and Engineering, 2008. O Sinnen, A To, M Kaur , Contention-Aware Scheduling with Task Duplication, Journal of Parallel and Distributed Computing, 2011. Luiz Meyer, Doug Scheftner, Jens Vöckler, Marta Mattoso, Mike Wilde, Ian Foster,An Opportunistic Algorithm for Scheduling Workflows on Grids, High Performance Computing for Computational Science, 2006.


1151


[59] Weizhe Zhang, Cheng, A M K, Mingzeng Hu, Multisite CoAllocation Algorithms for Computational Grid, 20th International Parallel and Distributed Processing Symposium, 2006. [60] D I George Amalarethinam, F Kurus Malai Selvi, A Task Duplication Based Efficient Multi-Objective Workflow Scheduling Algorithm, International Journal of Advanced Research in Computer Science, 2012. [61] D I George Amalarethinam, F Kurus Malai Selvi, An Efficient Dual Objective Grid Workflow Scheduling Algorithm, International Journal of Computer Applications, 2011. [62] Vydyanathan N, Catalyurek U, Kurc T, Sadayappan, P, A Duplication Based Algorithm for Optimizing Latency Under Throughput Constraints for Streaming Workflows, 37th International Conference on Parallel Processing, 2008. [63] Vydyanathan N, Catalyurek U, Kurc T, Sadayappan, P, Saltz J, Toward Optimizing Latency Under Throughput Constraints for Application Workflows on Clusters, 13th International European Conference on Parallel and Distributed Computing, 2007. [64] K Asakura, B Shao, T Watanabe, A Task Duplication Based Scheduling Algorithm for Avoiding Useless Duplication, PDPTA, 2006. [65] Dang Minh Quan, Yang, L T, Parallel Mapping with Time Optimization for SLA-aware Compositional Services in the Business Grid, IEEE Transactions on Services Computing, 2011. [66] Suchang Guo, Hong-Zhong Huang, Zhonglai Wang, Min Xie, Grid Service Reliability Modeling and Optimal Task Scheduling Considering Fault Recovery, IEEE Transactions on Reliability, 2011. [67] Prodan R, Wieczorek M, Bi-Criteria Scheduling of Scientific Grid Workflows, IEEE Transactions on Automation Science and Engineering, 2010. [68] Bozdag D, Catlyurek U, Ozguner F,A Task Duplication Based Bottom-Up Scheduling Algorithm for Heterogeneous Environments, 20th International Parallel and Distributed Processing Symposium, 2006. [69] Yiming Tan,Guosun Zeng,Shuixia Hao, Performance Analysis for Heterogeneous & Reconfigurable Computing Based on Scheduling, International Journalof Education and Management Engineering, 2011. [70] Mohammad Sadeq Garshasbi, Mehdi Effatparvar,High Performance Scheduling in Parallel Heterogeneous Multiprocessor Systems Using Evolutionary Algorithms, International Journal of Intelligent Systems and Applications, 2013. [71] Saeed Molaiy,Mehdi Effatparvar,Scheduling in Grid Systems using Ant Colony Algorithm, International Journal of Computer Network and Information Security, 2014. [72] Reza Fotohi, Mehdi Effatparvar,A Cluster Based Job Scheduling Algorithm for Grid Computing, International Journal of Information Technology and Computer Science, 2013. [73] Y. Xu et al., A genetic algorithm for task scheduling on heterogeneous computing systems using multiple priority queues, Information. Sciences, 2014. [74] Haluk Topcuoglu, Salim Hariri, Min-You Wu, PerformanceEffective and Low-Complexity Task Scheduling for Heterogeneous Computing, IEEE Transactions on Parallel and Distributed Systems, 2002. [75] Dipti Sharma and Pradeep Mittal, Job Scheduling Algorithm for Computational Grid in Grid Computing Environment, International Journal of Advanced Research in Computer Science and Software Engineering, 2013. [76] P. Keerthika and N. Kasthuri, An Efficient Grid Scheduling Algorithm with Fault Tolerance and User Satisfaction, Mathematical Problems in Engineering, 2013. [77] Sanjeev Puri,Harsh Dev,Deliberate, Secure Grid Computing Blueprint Design in Indian Context, International Journal of Computer Network and Information Security, 2012. [78] Rajabioun. R, Cuckoo optimization algorithm. Applied soft computing, 2011. [79] Saurabh Kumar Garg, Srikumar Venugopal, James Brober, Rajkumar Buyya, Double auction-inspired meta-scheduling of parallel applications on global grids, Journal of Parallel and Distributed Computing, 2013.


[80] Lizhe Wang, Samee U. Khan, Dan Chen, Joanna Kołodziej, Rajiv Ranjan, Cheng-zhong Xu, Albert Zomaya, Energy-aware parallel task scheduling in a cluster, Future Generation Computer Systems, 2013. [81] Chung-Hsing Hsu, Wu chun Feng, A feasibility analysis of power awareness in commodity-based high-performance clusters, CLUSTER, 2005. [82] Xiaomin Zhu, Rong Ge, Jinguang Sun, Chuan He, 3E: Energyefficient elastic scheduling for independent tasks in heterogeneous computing systems, The Journal of Systems and Software, 2013. [83] Ye Huang, Nik Bessis, Peter Norrington, Pierre Kuonen, Beat Hirsbrunner, Exploring decentralized dynamic scheduling for grids and clouds using the community-aware scheduling algorithm, Future Generation Computer Systems, 2013. [84] Wei Liu, Wei Du, Jing Chen, Wei Wang, Guo Sun Zeng, Adaptive energy-efficient scheduling algorithm for parallel tasks on homogeneous clusters, Journal of Network and Computer Applications, 2014. [85] Rajni, Inderveer Chana, Bacterial foraging based hyper-heuristic for resource scheduling in grid computing, Future Generation Computer Systems, 2013. [86] K.M. Passino, Biomimicry of bacterial foraging for distributed optimization and control, IEEE Control Systems Magazine, 2002. [87] Sucha Smanchat, Maria Indrawan, Sea Ling, Colin Enticott, David Abramson, Scheduling parameter sweep workflow in the Grid based on resource competition, Future Generation Computer Systems, 2013. [88] Ramya R and Shalini Thomas, An Optimal Job Scheduling Algorithm in Computational Grids, Special Issue of International Journal of Computer Applications, 2012. [89] Anis Gharbi et al, An Effective Genetic Algorithm for a Complex Real-World Scheduling Problem, International Journal of Mechanical, Industrial Science and Engineering, 2013. [90] Bsoul, M. , Phillips, I. , Hinde, C. , MICOSim: A simulator for modelling economic scheduling in Grid computing, World Academy of Science, Engineering and Technology, International Science Index 68, 2012. [91] Maryam Rabiee and Hedieh Sajedi, Job Scheduling in Grid Computing with Cuckoo Optimization Algorithm, International Journal of Computer Applications, 2013. [92] M. C. Lee, F. Y. Leu, and Y. Chen, PFRF: An adaptive data replication algorithm based on star topology Data Grids, Future Generation Computer Systems, 2012. [93] J. Taheri, Y. C. Lee, A. Y. Zomaya and H. J. Siegel, A Bee Colony based optimization approach for simultaneous job scheduling and data replication in Grid environments, Computers & Operations Research, 2013. [94] N. Saadat and A. M. Rahmani, PDDRA: A new pre-fetching based dynamic data replication algorithm in Data Grids, Future Generation Computer Systems, 2012. [95] N. Mansouri and Gh. Dastghaibyfard, Improving Data Grids Performance by using Modified Dynamic Hierarchical Replication Strategy, Iranian Journal of Electrical & Electronic Engineering, 2014. [96] N. Mansouri and G. H. Dastghaibyfard, A dynamic replica management strategy in Data Grid, Journal of Network and Computer Applications, 2012. [97] V. Andronikou, K. Mamouras, K. Tserpes, D. Kyriazis and T. Varvarigou, Dynamic QoS aware data replication in Grid environments based on data importance, Future Generation Computer Systems, 2012. [98] Kamalam, G.K. and V.Murali Bhaskaran, Novel adaptive job scheduling algorithm on heterogeneous grid resources, American Journal Applied Sciences, 2012. [99] Keerthika, P. and N. Kasthuri, An efficient fault tolerant scheduling approach for computational grid, American Journal of Applied Sciences, 2012. [100] P. Linberg, J. Leingang, D. Lysaker, S. Khan, J. Li, Comparison and analysis of eight scheduling heuristics for the optimization of energy consumption and makespan in large-scale distributed systems, The Journal of Supercomputing, 2012.


1152


Authors’ Information 1

Department of Computer Science, Jamal Mohamed College, Trichirappalli, Tamil Nadu, India. 2

Department of Computer Science, Faculty of Science and Humanities, SRM University, Kattankulthur Campus, Chenni, Tamil Nadu, India. Dr. D. I. George Amalarethinam is an Associate Professor of Computer Science and Director of MCA Programme. He has 26 years of experience in teaching and 17 years of experience in research. He has presented and published more than 80 research articles in the international/national conferences and journals. He has chaired many technical sessions and delivered invited talks in national/international conferences. He is an editorial board member in leading international journals. His interests include Parallel Computing, Grid Computing, Distributed Databases and Ad hoc Networks. P. Muthulakshmi is an Assistant Professor of Computer Science. She is interested in Parallel Processing and Distributed Computing, Grid Computing, Data Structures and Algorithms.



1153


Performance Analysis Comparative Study of Fingerprint Recognition Systems Mohammed Saeed Jawad1, Fares Al-Shargie2, Murthad Al-Yoonus3, Zahriladha bin Zakaria2 Abstract – Forensic applications, such as criminal investigations, terrorist identification, and national security, require a strong fingerprint identification system. This paper proposes four methods, namely, canny filter, Gabor filter, dual-tree complex wavelet transform (DTCWT), and principal component analysis (PCA), to obtain a high fingerprint recognition rate. Frequency domain filtering is used to enhance fingerprint images. In canny filter, feature extraction based on the gray level co-occurrence matrix (GLCM) is computed from the gradient and coherence images. Fingerprint features are extracted and stored through the eight different orientations of Gabor filter. The redundancy and shift invariance of DTCWT is useful for obtaining highresolution images with preserved edges. PCA is used to extract the statistical features of fingerprints by reducing their dimensions and complexity. The proposed methods improved the efficiency of fingerprint recognition by combining GLCM-based feature extraction with a Knearest neighbors classifier. Co-occurrence matrices are used to extract features from the fingerprint image because they are composed of regular texture patterns. The proposed methods increased the recognition rate and reduced complexity and time. The false accepted rate, false rejected rate, and total success rate were improved by the proposed methods compared with those of existing algorithms. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Canny Filter, Gabor Filter, DTCWT, PCA, GLCM, KNN

Traditional security systems are also prone to inaccuracies compared with biometrics, which is accurate. Fingerprint recognition is one of the most reliable personal identification methods. This method significantly influences forensic applications, criminal investigations, terrorist identification, and national security. The fingerprint is a patterned impression consisting of ridges, which are single curved segments and valleys (the region between two adjacent ridges). The uniqueness of the fingerprint depends on the pattern of ridges and the location of minutiae. Minutia points are local ridge characteristics at ridge bifurcation or ridge ending (Figs. 1).

Nomenclature PCA DTCWT FFT KNN GLCM TSR FRR FAR

Principal Component Analysis Dual-Tree Complex Wavelet Transform Fast Fourier Transform K-Nearest Neighbors Gray Level Co-Occurrence Matrix Total Success Rate False Rejected Rate False Accepted Rate

I.

Introduction

Reliable personal authentication has become an important human–computer interface activity in the increasingly digitized world. National security, ecommerce, and access to computer networks have become common and vital tools for establishing the identity of a person. Biometry is the science of identifying an individual based on certain physiological or behavioral characteristics. Fingerprint, palm print, iris, cornea, face, voice print, gait, and DNA are several biological features used in biometrics. Biometric identifiers are permanently associated with the user and are thus more reliable than token- or knowledge-based authentication methods. Biometrics offers several advantages over traditional security measures. Biometrics-based security systems are secure and accurate compared with traditional passwords.

(a)

Figs. 1. Minutia points. (a) Ridge ending and (b) bifurcation

Fingerprints have five major classes, namely, arch, tented arch, left loop, right loop and whorl [1].  Arch: Ridges enter from one side and rise to form a small bump before going down to the opposite side. No loops or delta points are present.


1154

(b)

Mohammed Saeed Jawad, Fares Al-Shargie, Murthad Al-Yoonus, Zahriladha bin Zakaria

 Tented arch: This is similar to the arch, except that at least one ridge has high curvature, which has one core and one delta point.  Left loop: One or more ridges enter from one side, curve at the back, and exit on the same side they entered. Core and delta are present.  Right loop: This is the same as the left loop, but in the opposite direction.  Whorl: This contains at least one ridge that makes a complete 360-degree path around the center of the fingerprint. Two loops (same as one whole) and two deltas can be found. Fingerprints in databases are non-uniformly distributed in these classes.

Arch

Left loop

Tented Arch

Right loop

Whorl

Unclassified

Figs. 2. Fingerprint classes

This paper proposes methods for a fingerprint recognition system based on canny filter, Gabor filter, principal component analysis (PCA), and dual-tree complex wavelet transform (DTCWT). Detecting the edges of fingerprint images facilitates the location of the core, ridges, and minutia points. The canny edge detection algorithm detects optimal edges in digital image pre-processing. Canny uses a Gaussian filter to smooth the image and remove noise. The first-order derivatives in an image are computed by using the gradient, whereas second-order derivatives are obtained through the Laplacian [2]. The advantage of the canny algorithm is that the first derivatives are computed in the x and y directions and combined into four directional derivatives. The points where these directional derivatives are the local maxima are the candidates for assembly into edges. However, the most significant new dimension of the canny algorithm is its attempt to assemble individual edge candidate pixels into contours usually formed by applying a hysteresis threshold to the pixels. However, a pixel with a gradient higher than the upper threshold is accepted as an edge pixel. A pixel below the lower threshold is rejected, whereas a pixel gradient between thresholds is accepted only if it is connected to a pixel above the high threshold value [3]. Therefore, this method is less likely to be fooled by noise and more likely to detect true weak edges than other methods [4]. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

This paper proposes an approach to estimating local ridge orientation by using image convolution based on canny algorithm. The proposed algorithm can detect fingerprint features even at different orientations. This approach is simple compared with the minutia point pattern matching algorithm. Gabor filters have both frequency-selective and orientation-selective properties and optimal joint resolution in both the spatial and frequency domains. Thus, Gabor filters are suitable bandpass filters for removing noise and preserving ridge and valley structures [5]. Hong, Wan, and Jain (1998) introduced the use of Gabor filters in image enhancement. Lee et al. [6] used Gabor filter-based features for fingerprint recognition. This method required an additional step for detecting the reference point (core point) in the fingerprint image. Since then, several studies have improved and modified Gabor filters for fingerprint image enhancement. These studies include Yang, Liu, Jiang, and Fan (2003) and Wang, Jian, Huang, and Feng (2008) [7]. DTCWT is an efficient tool for various image processing applications, such as de-noising, edge detection, restoration, enhancement, and compression. DTCWT is an enhanced version of the discrete wavelet transform (DWT), with important additional properties. The main advantage of DTCWT over the original DWT is that the complex wavelets are approximately shift-invariant and have separate sub-bands for positive and negative orientations. For 2-D image signals, filter banks with a pair of trees are used in the columns and rows of images [8]. In case of real 2D filter banks, the three high-pass filters have orientations of 0° , 45° , and 90° ; for the complex filters, the six sub-band filters are oriented at ±15° , ±45° , and ±75° .

II.

Previous Work

Image-based automatic fingerprint matching is proposed in [9]. Fingerprint images were matched based on features extracted from the wavelet domain. The feature vector is an approximation of the image energy distribution over different scales and orientations. A 95.2% recognition rate is achieved through this method [9]. Islam (2010) proposed fingerprint detection based on canny filter and DWT. The raw image was convolved with a Gaussian filter and thereby produced a blurred version of the original image. This version was unaffected by any noisy pixel to any significant degree [10]. The Canberra distance metric was used to determine the similarity between texture classes; an accuracy rate of above 95% was obtained. The accuracy obtained was between 80% and 90%. Hong, Wan, and Jain (1998) introduced the use of Gabor filters in image enhancement. Lee et al. [5] used Gabor filter-based features for fingerprint recognition. Their method required an additional step for detecting the reference point (core point) in the fingerprint image. Since then, several studies have improved and modified Gabor filters to enhance fingerprint images.


1155


These studies include Wang, Jian, Huang, and Feng (2008) [7]. Mudegaonkar (2011) proposed an approach to fingerprint identification using a Gabor filter bank. The proposed filter-based algorithm uses a bank of Gabor filters to capture local and global details in a fingerprint as a compact fixed-length finger code [11]. Identification is based on the Euclidean distance between the two corresponding finger codes, obtaining an accuracy of 98.22%. Carsten (2012) proposed fingerprint recognition using Curved Gabor filter are applied to the curved ridge and valley structure of low-quality fingerprint images [12]. The features were extracted from Gabor filter. Euclidian distance was used for recognition.

4) The image is multiplied by the filter mask. 5) The result is transformed back to the spatial domain. Normalization adjusts the gray-level values of an input fingerprint image to obtain the specific mean and variance. Normalization is a pixel-wise operation. This method does not change the clarity of the ridge and furrow structures in the fingerprint image. Normalization can be used: 1) To obtain images with similar characteristics. 2) To remove the effect of sensor noise. 3) To reduce variation in gray levels along ridges and valleys. 4) To improve image quality.

III. Methodology III.2. Principal Components Analysis

III.1. Pre-processing Image Enhancement Using Fast Fourier Transform Enhancements primarily aim to improve the clarity of ridge structures in the recoverable region and mark unrecoverable regions. In this technique, the image is divided into small blocks (32 × 32 pixels). The fast Fourier transform (FFT) is performed according to the following equation: ( , )=

( , )

{

}

(1)

where u = 0, 1, 2, ..., 31 and v = 0, 1, 2, ..., 31. However, to enhance a specific block by its dominant frequencies, the FFT of the block is multiplied by its magnitude several times. The magnitude of the original FFT = abs (F (u, v)) = |F (u, v)|. Thus, the enhanced block is obtained according to the following equation: ( , )= where

{ ( , ) × | ( , )| }

(2)

(F (u, v)) is given as:

( , )=

1

( , )

{

(

)

(3)

In this method, the pre-processed fingerprint follows the PCA algorithm to obtain the feature vector in the spatial domain. PCA involves a mathematical procedure that transforms several possibly correlated variables into fewer uncorrelated variables called principal components. PCA can be summarized by the following steps: a. The pre-processed fingerprint image is acquired and represented by a 1D vector by concatenating each row (or column) into a long vector. b. The mean M of the training set is calculated and subtracted from the training set: 1

= =

−

is applied to all training sets

c. The eigenvectors and eigenvalues of the training set covariance matrix are calculated. = . . The eigenvectors are made to correspond to the N largest eigenvalues of to form the new basis of the N principal components matrix (V). d. The basis vector matrix (U) is constructed as: =

x = 0, 1, 2 …31 and y = 0, 1, 2 ...31. k is constant (0.45). High k improves the appearance of ridges by filling up their small holes. However, extremely high k may cause the false joining of ridges and thus transform the termination into a bifurcation. Image filtering through FFT techniques is summarized below: 1) The image is converted to suitable dimensions (the number of pixels in the x and y axes must be a factor of 2.). 2) FFT is performed on the image to transform it to a frequency space. 3) A filter mask is created, which performs the tasks set by the user.


·

e. The reduced feature vector is calculated for the training and testing data as: = =

· ·

M must be subtracted from all testing data [13]. The feature vector is eliminated through classification and normalization, thereby enhancing accuracy. K-nearest neighbors (KNN) is the classifier used in this paper. The total success rate (TSR), false rejected rate (FRR), and false accepted rate (FAR) are calculated and compared with those of other algorithms.


1156


III.3. Gabor Filter Gabor filters are widely used because they can remove noise and enhance the ridge and valley structures in fingerprint images. Gabor filters are band-pass filters with both orientation-selective and frequency-selective properties and optimal joint resolution in both the spatial and frequency domains [1]. Using Gabor filter significantly enhances ridge and furrow structures. The proposed Gabor filter has the following general form in the spatial domain: −1 2

( , , , )= = =

(2

+

) (4)

+ −

f is the frequency along the direction  from the x axis, and , are the Gaussian envelopes along the x and y axes, respectively. These Gabor filters are performed in eight orientations with four different  (0 , 45 , 90 , and 135 ) with respect to the x axis. The mean and variance of the pre-processed fingerprint image I are given by the following equations: ( )=

( )=

1

(, )

1

( , )−

(5)

( )

(6)

Let N(i; j) represent the normalized gray level at pixel (i, j). The normalized image is defined as:

(, )=

( , )−

+

( )

(, )> (7) (, )=

( , )−

−

These features are stored in the vector and added to the extracted features based on the gray level cooccurrence matrix (GLCM), as shown in the feature extraction section. Several steps were performed in the proposed algorithm to obtain an efficient fingerprint recognition rate. The following are the steps of the proposed algorithm: 1) The fingerprint images are enhanced. 2) The image is cropped to predetermined size. 3) The Gabor filter is applied to the fingerprint images. 4) Entropy, correlation, contrast, homogeneity and energy are computed. 5) The extracted features are concatenated. 6) The test images are classified by the KNN classifier. 7) TSR, FRR, and FAR are calculated. III.4. DTCWT Two-dimensional multi-scale transforms that represent edges more efficiently than DWT have been designed and implemented in wavelet-related research. DTCWT is an enhancement of DWT, with important additional properties. The main advantage of DTCWT over the original DWT is that the complex wavelets are approximately shift-invariant and have separate subbands for positive and negative orientations. The redundancy and shift invariance of DTCWT imply that DTCWT coefficients are inherently interpolable [14]. This study used DTCWT to decompose an input image into high- and low-frequency sub-bands. Only the high-frequency sub-band images are interpolated and combined with the original interpolated image. Interpolating the input image by α/2 and the highfrequency sub-band images by α and then applying inverse DTCWT (IDTCWT) to both images provide the output image with sharper edges. This result is obtained because interpolating the isolated high-frequency components in the highfrequency sub-band images preserves more highfrequency components.

( )

ℎ M0 and VAR0 are the desired mean and variance, respectively. ( , ) (x, y) be the image component at  for Let each cell. The extracted feature is the standard deviation of each cell according to the following equation: =

(

( , )−

Fig. 3. DTCWT

)

where is the number of pixels in each cell and the mean of the pixel values in .

is


DTCWT has the following properties that improve the efficiency of preserving edges in fingerprint images: 1) Approximation of shift invariance. 2) Good selectivity and directionality in 2D. International Review on Computers and Software, Vol. 9, N. 7

1157


3) Perfect reconstruction using short linear phase filters. 4) Four times redundant for images (4:1). 5) Efficient order, N computation, is 4 times that of simple DWT. III.5. Canny Filter Estimating Edge Directional Features One of the most important features that must be estimated is the local ridge orientation or direction within the fingerprint image. This estimation is crucial to identifying the major core singular characteristics of the fingerprint image. Core singularities include features, such as whorls, deltas, arches, and loops, and their dominant orientation in the image. The steps involved to obtain edge directional features through canny filter are summarized below. First, the images are enhanced, and the gradient is obtained by with a corresponding angle at (m, n) according to the following equations: |+

= (| =

(

∑

cos( ∑

=

1 2

∑ ∑

∑ ∑

(, )

(8)

)

−

Correlation texture measures the linear dependency of gray levels on neighboring pixels. Correlation refers to of the correlation of a pixel to its neighbor over the entire fingerprint image, where i and j are the pixels, K is the row (or column) dimension of the square matrix, and , is the probability of pixel pairs that satisfy the offset. Range = [−1, 1]. 0 indicates no correlation, 1 means perfect correlation, and a non-numeral indicates a constant image:

) ( −

=

Third, the dominant local orientation is calculated from the gradient and coherence according to the following equation: =

(| − |)

)

Quantities and represent the components of the gradient image directions. Second, coherence is determined through the gradient and angle according to the following equation: =

GLCM Haralick [16] described 14 statistical measurements that can be calculated from GLCM to describe image texture. The five features extracted from the fingerprint images were contrast, correlation, energy, homogeneity, and entropy. Contrast measures the local variations in GLCM, where i and j are pixels, K is the row (or column) dimension of the square matrix, and , is the probability of pixel pairs that satisfy the offset. However, equal i and j indicate that ( − ) = 0 and that the fingerprint image is on the diagonal. Weight function ( − ) = 0 represents pixels completely similar to their neighbor. The weights exponentially increase as (i−j) increases. The function returns a measure of the intensity contrast between a pixel and its neighbor over ( the entire image. = [0 ( , 1) − 1) ]. Null contrast indicates a constant image:

2 2

+

)( −

=

()

=

()

) (, )

(9)

2

Let N = 8. Each 8 × 8 window represents one orientation information of the fingerprint image. Fourth, entropy, correlation, contrast, homogeneity, and energy are computed from the gradient and coherence images and stored in the vector. Matching is performed by the KNN classifier. TSR, FRR, and FAR are finally calculated.

=

( −

)

()

=

( −

)

()

III.6. Feature Extraction Features are extracted based on the GLCM of the three proposed methods, namely, Gabor filter, canny filter, and DTCWT. GLCM [15] is a second-order statistical method based on (local) information about gray levels in pixel pairs. This study computed multiple GLCMs for different angles. Thus, five features were extracted for matching.


The energy function measures the homogeneity of an image and sums the squared elements in GLCM. A homogeneous scene contains only a few gray levels and thus yields a GLCM with few but high p(i, j). Thus, the sum of squares is high. is the , probability of pixel pairs that satisfy the offset. Range = [0, 1]. An energy value if 1 indicates a constant image:


1158


=

(, )

(10)

The homogeneity function measures the closeness of the distribution of elements in GLCM to that of the diagonal GLCM. i and j are the pixels, K is the row (or column) dimension of the square matrix, and , is the probability of pixel pairs that satisfy the offset. Range = [0, 1]. Homogeneity equals 1 in a diagonal GLCM: =

(, ) (1 + ( − ))

(11)

Entropy is a statistical measure of randomness that can be used to characterize the texture of the given image. Entropy is a measure of randomness that obtains its highest value when the elements of are all equal:

( )=

[ ( ) − ( )]

(14)

where N is the number of features in the feature vector f, fj(x) represents the jth feature of test sample X, and fj(M) represents the jth feature of the Mth class in the feature library. Test sample X is then classified by the KNN classifier. In the KNN classifier, a test sample is classified by a majority vote of its k neighbors, where k is a typically small positive integer. If K =1, then the sample is assigned the class of its nearest neighbor. However, K should be set as an odd number to avoid tied votes. Therefore, the KNNs in this method are determined, and the test image is classified as the language type of the majority of these KNNs. III.9. Evaluation of Fingerprint Recognition Algorithms

=−

( ) log

( )

(12)

where P (xi) is the probability that the difference between two adjacent pixels is equal to i. logb is the base 2 logarithm. III.7. Feature Vectors Features extracted from Gabor filter, PCA, DTCWT, and canny filter go separately through their own matching modules. The matching score is then normalized within the Euclidian distance to improve accuracy. Let X denote the raw matching scores from a specific matcher and xX. The normalized score of x is denoted by ′ . Min–max normalization This normalization maps the raw matching scores to the interval [0, 1] and retains the original distribution of these scores, except for a scaling factor. Given that max(X) and min(X) are the maximum and minimum values, respectively, of the raw matching scores, the normalized score is calculated as: =

− ( ) ( )− ( )

(13)

III.8. Classification and Recognition In the proposed method KNN is used to classify the fingerprint test samples. The features are extracted from test image X through the proposed feature extraction algorithm. The features are then compared with the corresponding feature values stored in the feature library by using the Euclidean distance given by the following equation: Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

FAR: This rate is the probability that an unauthorized person is incorrectly accepted as an authorized person by using a match count and the product of the number of fingers (NF) with total images per finger: %=

(

)

∗

∗ 100

(15)

FRR: This rate is the probability that the system does not detect an authorized person by using a mismatch count and NF: %=

(

)

∗ 100

(16)

TSR: This is the rate at which a match successfully occurs: %=

IV.

(

)

∗ 100

(17)

Experimental Results IV.1. Database

The experiment is performed on a DB3-A FVC2004 fingerprint database. The fingerprint database of FVC2004 DB3-A has 800 fingerprints of 100 different fingers (eight images per finger). Fingerprint images are numbered 1 to 100, followed by another number (1 to 8), indicating that the image fingerprints contain the first to eighth impressions of a certain finger. This study uses 672 fingers that represent 84 different people. Figure 4 shows an example of the fingerprint images used.


1159


Correlation, energy, entropy, contrast, and homogeneity were computed and stored in the vector. The feature vector obtained from the test image was compared with the feature vector stored in the database. Matching was conducted based on K-means classifier. The next-next-nearest neighbor (3NN) achieves a TSR of 99.9%, a total FRR of 0.1%, and a total FAR of 0.5%. IV.4. Result of DTCWT Fig. 4. Examples of database fingerprints used

IV.2. Image Enhancement with FFT Image enhancement is important in designing fingerprint recognition systems. The image enhanced by FFT can connect falsely broken points on ridges and remove spurious connections between them. The figure below shows the images before and after the application of FFT algorithm to enhance ridges and valleys.

A final high-resolution output image is generated by applying IDTCWT to the interpolated sub-band images and input image. This technique obtains high-resolution images with more edges preserved. Entropy, correlation, contrast, homogeneity, and energy are computed from the output image obtained by DTCWT and stored in the vector. The feature vector of the tested image is compared to the feature vector stored in the database. Matching is conducted by K-means classifier. The 3NN classifier obtains a TSR of 99.8%, a total FRR of 0.2%, and a total FAR of 0.4%.

(a) (a)

(b) Figs. 5. (a) Original image and (b) Image enhanced by FFT (b)

IV.3. Result of Canny Filter

Figs. 6. (a) Gradient and (b) coherence images

This method sets the variance σ to 1.5 and the threshold value (α) to 0.03 for the entire experiment. With these values, the edge is smoothly detected after enhancement by FFT. The result shows a good representation of ridge and valley structures. The ridge and valley structures were unaffected by the threshold because they were significantly enhanced as a pre-processing step. The gradient and coherence images were obtained from this algorithm.

(a)



1160


properties therefore it gives the best TSR and FRR. DTCWT outperforms canny filter in terms of FAR because of, it is approximately shift-invariant and have separate sub-bands for positive and negative orientations. The contribution of this paper results in reducing the complexity and improves the overall accuracy of fingerprint recognition.

Acknowledgements

(b) Figs. 7. Image enhanced by DTCWT

IV.5. Result of Gabor Filter and PCA The images are resized to 256×256 pixels in 256 grayscale levels. The images are enhanced by FFT, which divides the fingerprint image into a set of 32×32 non-overlapping blocks, and then normalized. Gabor filter banks are used in eight orientations (f = 0.1), and the feature vector is first created. However, the normalized Gabor output image is reconstructed. Correlation, contrast, entropy, homogeneity, and energy are computed and concatenated into the created feature vector. Matching is finally conducted based on the 3NN classifier. The 3NN classifier obtains a TSR) of 99.9%, a total FRR of 0.1%, and a total FAR of 0.3%. PCA obtains a TSR of 90%, a total FRR of 10%, and a total FAR of 1.5%. The figure below shows the reconstruction of the normalized output of the Gabor filter. The performances of the proposed techniques are evaluated in terms of TSR, FRR, and FAR. Gabor filter, canny filter, and DTCWT have the best TSR and FRR. However, DTCWT outperforms canny filter in terms of FAR because of its shift-invariance property. Table I below summarize all the results obtained by the four algorithms as discussed before. TABLE I ANALYSIS OF DIFFERENT METHODS WITH VARIOUS PARAMETERS Gabor DTCWT PCA Canny Accuracy Accuracy Accuracy Accuracy (%) (%) (%) (%) TSR 99.9 99.9 99.8 90 FRR 0.1 0.1 0.2 10 FAR 0.5 0.3 0.4 1.5

The authors would like to express their deepest gratitude and appreciation for Malaysia Ministry of High Education (MOHE) for the financial support under the grant ERGS/2013/FKEKK/TK02/UTEM/03/05/E00021, as this financial support allows this work to be completed and published.

References [1] [2]

[3]

[4]

[5]

[6] [7]

[8]

[9]

[10]

[11]

V.

Conclusion

This paper compared four different algorithms, three of which use the same feature extraction method for fingerprint recognition. The uniqueness of the fingerprint and low probability of a false acceptance or false rejection all contribute to the benefits of using fingerprint recognition technology. The image enhancement technique used helps in preserving the edges of fingerprint and improved its quality. Gabor filters has the advantages of its band-pass filter with both orientation-selective and frequency-selective Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

[12]

[13]

[14]

[15]

S. Prabhakar, Maltoni, Davide, D. Maio, Anil K. Jain, Handbook of Fingerprint Recognition (Springer, 2009). J. Canny, A computational approach to edge detection, Pattern Analysis and Machine Intelligence, IEEE Transactions, Vol. 6, pp. 679–698, 1986. Xiao, T., Yin-He, W., Qin-Ruo, W., A face recognition method based on complex network, canny algorithm and image contours, (2013) International Review on Computers and Software (IRECOS), 8 (1), pp. 204-21. P. Singh Sandhu, Juneja, Mamta, Performance evaluation of edge detection techniques for images in spatial domain, Methodology, Vol. 1, n. 5, pp. 614–621, 2009. A. K. Jain, L. Hong, Y. Wan, Fingerprint Image Enhancement: Algorithm and Performance Evaluation, Pattern Analysis and Machine Intelligence, IEEE Transactions, Vol. 20, n. 8, pp. 777– 789, 1998. S.-D. Wang, Lee, Chih-Jen, Fingerprint feature extraction using Gabor filters, Electron. Letters, Vol. 35, n. 4, pp. 288–290, 1999. H. Feng, Wang, Wei, Jianwei Li, Feifei Huang, Design and implementation of Log-Gabor filter in fingerprint image enhancement, Pattern Recognition Letters, Vol. 29, n. 3, pp. 301– 308, 2008. Mtimet, J., Amiri, H., Image classification using statistical learning for automatic archiving system, (2013) International Review on Computers and Software (IRECOS), 8 (6), pp. 1228123. J. Saarinen, Tico, Marius, P. Kuosmanen, Wavelet domain features for fingerprint recognition, Electronics Letters, Vol. 37, n. 1, pp. 21–22, 2001. M. R. Amin, Islam, Md Imdadul, Nasima Begum, Mahbubul Alam, Fingerprint Detection Using Canny Filter and DWT, a New Approach, JIPS, Vol. 6, n. 4, pp. 511–520, 2010. Mudegaonkar, Prajakta M., Ramesh P. Adgaonkar, A Novel Approach to Fingerprint Identification Using Gabor Filter-Bank, International Journal of Network Security, Vol. 2, n. 3, pp. 10–14, 2011. Gottschlich, Carsten, Curved-region-based ridge frequency estimation and curved Gabor filters for fingerprint image enhancement, Image Processing IEEE Transactions, Vol. 21, n. 4, pp. 2220–2227, 2012. Hasan, Hamid M., Waleed A. AL Jouhar, Majid A. Alwan, Face Recognition Using Improved FFT Based Radon by PSO and PCA Techniques, International Journal of Image Processing, Vol. 6, n. 1, pp. 26–37, 2012. Jagadeesh, Pilla, J. Pragatheeswaran, Image resolution enhancement based on edge directed interpolation using dual tree—Complex wavelet transform, Recent Trends in Information Technology (ICRTIT), International Conference, pp. 759–763, 2011. Rafel C. Gonzalez, Richard E.Woods, Digital Image Processing,


1161


(Upper Saddle River, New Jersey 07458 Pearson Education, Inc., 2008, Third Edition, pp.830-836). [16] Haralick, Robert M., Statistical and structural approaches to texture, Proceedings of the IEEE, Vol. 67, n. 5, pp. 786–804, 1979.

Authors’ information 1

Department of Telecommunications Engineering, Faculty of Electronic and Computer Engineering, University Technical Malaysia Melaka, Hang Tuah Jaya, 76100 Durian Tunggal, Melaka, Malaysia. Phone : +605552162 Fax: +605552112 E-mail: [email protected] 2

Department of Industrial Electronics Engineering, Faculty of Electronic and Computer Engineering, University Technical Malaysia Melaka, Hang Tuah Jaya, 76100 Durian Tunggal, Melaka, Malaysia. Phone : +60136286149 E-mails: [email protected] [email protected] 3

Department of Computer Engineering, Faculty of Electronic and Computer Engineering, University Technical Malaysia Melaka, Hang Tuah Jaya, 76100 Durian Tunggal, Melaka, Malaysia. Phone : +601128380518 E-mail: [email protected] Mohammed Saeed Jawad received his BEng (Hons) in Computer Engineering from AlIttihad University (UAE), 2003 and master Degree in computer Sciences from University Sciences of Malaysia, 2006 and his PhD Degree in 2013 from the same university. His current research interests include but not limited to: 2D/3D image processing, Computer networks for Next Generations communication systems, Wireless communications systems, Network Security, Ultra-Wide Band communication systems and applications. Murthad Al-Yoonus was born in Mosul, Iraq on July 1986 obtained his B.Eng in Computer Engineering Techniques, Mosul, Iraq, 2008, and his M.Eng from University Technical Malaysia Melaka (UTeM) on 2013; he is currently working toward PhD in Electrical and Electronic Engineering in University Tun Hussein Onn Malaysia (UTHM). His research interests include image processing, 3D image processing, Embedded System Design. Fares AL-Shargie received the B.Eng. degree in Electronics Majoring in Bio-instrumentation from Multimedia University (MMU), Malaysia, in 2011, the M.Eng. degree in Electronic Engineering from University Teknikal Malaysia Melaka (UTeM), Malaysia, in 2013. He is currently pursuing a Ph.D. degree in Electrical Engineering (Biomedical Engineering) at Universiti Teknologi PETRONAS (UTP), Malaysia. His current research interests include image processing, statistical signal processing, EEG and fNIRS. His research project is for the development of mental stress assessment using multimodal fusion techniques.



1162


A Comparative Analysis on Markov Model (MM) Association Rule Mining (ARM), Association Rule Mining-Statistical Features (ARM-SF), Boosting and Bagging Model (BBM) to Impervious Web Page Prediction Sampath Prakasam1, Amitabh Wahi2, Ramya Duraisamy3 Abstract – The web usage mining techniques are used to analyze the web usage patterns for a web site. To fetch the user access patterns from the user log files. The access patterns are used in the prediction process. Web prediction is a classification model attempts to predict the next set of Web pages that a user may visit based on the knowledge of the previously visited pages. By Predicting user’s behavior can be applied effectively in various critical applications in the internet environment.So that the application has traditional tradeoffs between modeling complexity and prediction accuracy. The association rule mining and classification techniques are used to perform the pattern extraction and prediction process. The technique as Markov model and all Kth Markov model are used in Web prediction. A modified Markov model is proposed to alleviate the issue of scalability in the number of paths. The framework can improve the prediction time without compromising prediction accuracy. The proposed system integrates the boosting and bagging models with the association rule mining technique to improve the prediction accuracy. The statistical features are also discovered and used in the proposed model. Standard benchmark data sets are used to analyze the web page prediction process. The proposed model reduces the number of paths without compromising accuracy. The system improves the accuracy with scalability considerations. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Association Rule Mining, Boosting and Bagging Model, Markov Model

I.

Introduction

Web prediction is a classification problem in which we attempt to predict the next set of Web pages that a user may visit based on the knowledge of the previously visited pages [1]-[9]. Such knowledge of user’s history of navigation within a period of time is referred to as a session. These sessions, which provide the source of data for training, are extracted from the logs of the Web servers. It should contain sequences of pages that users have visited along with the visit date and duration. The Web prediction problem (WPP) can be generalized and applied in many essential industrial applications, as search engines, caching systems, recommendation systems, and wireless applications [5]. Therefore, it is crucial to look for scalable and practical solutions that improve both training and prediction processes. Improving the prediction process can reduce the user’s access times while browsing, and it can ease network traffic by avoiding visiting unnecessary pages.

When a prediction model for a certain Web site is available, the search engine can utilize it to cache the next set of pages that the users might visit. It mitigates the latency problem of viewing Web documents particularly during Internet traffic congestion periods. Personalization as, in which users are categorized based on their interests and tastes. The interests and tastes of users are captured according to the previously visited categorized Web pages. Furthermore, in the wireless domain, mobile Web prediction is used to reduce the number of clicks needed in wireless devices such as PDAs and smart phones, which helps to mitigate problems related to the display size limitations. In Web prediction, we face challenges in both preprocessing and prediction. The technique of Preprocessing challenges includes handling large amount of data that cannot fit in the computer memory. So that it includes of choosing optimum sliding window size, identifying sessions, and seeking/extracting domain knowledge. Long training/prediction time, low prediction accuracy are the challenges faced in the prediction process with the inclusion of memory limitation.


1163

Sampath Prakasam, Amitabh Wahi, Ramya Duraisamy

In our previous work [1], we have computed a new prediction model based on fusing several prediction models, namely, support vector machines (SVMs), artificial neural networks (ANNs), and Markov model. This model fusion, along with domain knowledge exploitation, has enabled us to considerably improve the prediction accuracy. Our previously proposed framework has shown clear improvement in prediction accuracy; however, it suffers from expensive prediction and training overhead. Additionally, some models, such as association rule mining (ARM) and SVM, do not scale well with large data sets. Some models, such as SVM and ANN, do not handle the multiclass problem efficiently because of the large number of labels/classes involved in the WP. In this paper, we extend our previous work [2] to successfully improve prediction accuracy using simpler probabilistic models such as Markov model and ARM. In addition, we aim to reduce the prediction time particularly when there are many prediction models to consult. Our contributions in this paper can be summarized as follows: 1) We propose a new two-tier prediction framework to improve prediction time. Such framework can accommodate various prediction models. The framework is based on creating a special classifier, dubbed example classifier (EC), during training to map each example to one classifier from a pool of classifiers produced during training. In prediction, an example goes through EC first to select a classifier, and after that, the selected classifier is used in prediction. This desired feature enables us to analyze the performance of any single or combination of models and utilize the most effective ones. 2) We present an analysis study for Markov model and all-Kth model in the WPP utilizing different N-grams. Specifically, we show how accuracy is affected when using different N-grams. 3) We propose a new modified Markov model that handles the excess memory requirements in case of large data sets by reducing the number of paths during the training and testing phases. 4) We conduct extensive experiments on three benchmark data sets to study different aspects of the WPP using Markov model, modified Markov model, ARM, and all-Kth Markov model. Finally our analysis and results show that higher order Markov model produces better prediction accuracy. In addition, the results indicate a dramatic improvement in prediction time for our novel proposed two-tier framework. Moreover, the results demonstrate the positive effect of our proposed modified Markov model in reducing the size of the prediction models without compromising the prediction accuracy. Finally, it presents experiments to study the effect of sparsity of pages, training partitioning, and ranking on the prediction accuracy.


II.

Related Work

Prediction models for addressing the WPP can be broadly path-based prediction models. Path-based prediction is based on user’s previous and historic path data, while point-based prediction is based on currently observed actions. Thus the Accuracy of point-based models is low due to the relatively small amount of information that could be extracted from each session to build the prediction model. So that the Researchers have used various prediction models including k-nearest neighbor (kNN), ANNs [7], fuzzy inference [6] SVMs, Bayesian model, Markov model and others. Recommendation systems are one of the early applications of Web prediction. Joachims et al. propose the Web Watcher which is a path-based recommender model based on kNN and reinforcement learning. The combination of previous tours of similar users and reinforcement learning is used in recommendations. Nasraoui et al. [8] propose a Web recommendation system using fuzzy inferences. Clustering is applied to group profiles using hierarchical unsupervised niche clustering. Context-sensitive URL associations are inferred using a fuzzy approximate- reasoning-based engine. Mobasher et al. use the ARM technique in WPP and propose the frequent item set graph to match an active user session with frequent item sets and predict the next page that user is likely to visit. However, ARM suffers from well-known limitations including scalability and efficiency. In the context of adaptive learning, the author Anderson et al. use dynamic links that provide shorter path to reach the final destination. So that Perkowitz and Etzioni utilize adaptive Web sites based on the browsing activities. Su et al. have proposed the N-gram prediction model and applied the all-N-gram prediction model in which several N-grams are built and used in prediction. Levene and Loizou [9] compute the information gain from the navigation trail to construct a Markov chain model to analyze the user navigation pattern through the Web. Pitkow and Pirolli propose longest repeating subsequences to focus on significant surfing patterns to reduce the model complexity. Fu et al. [3] propose the use of N-gram model in mobile Web navigation. They mainly show that, the lower the order of the N-gram, the better the prediction accuracy and performance. Hassan et al. use Bayesian model to focus on certain patterns. It includes short and long sessions, page categories, range of page views, and rank of page categories. Our work is related to recommendation systems and follows N-gram and all N-gram path-based prediction models [4]. However, it is different in three aspects. First, we synergize different prediction models to improve accuracy using simpler probabilistic models, namely, ARM and Markov. Secondly, a two-tier framework to reduce prediction time in case of consulting many classifiers to resolve prediction. Our framework can accommodate different prediction models; hence, we can analyze the performance of any International Review on Computers and Software, Vol. 9, N. 7

1164


single or combination of models, and later, we utilize the most effective ones. Third, try to reduce complexity of models by proposing a modified Markov model in which we reduce complexity by treating sessions as set of pages rather than sequence of pages.

III. Association Rule Mining and Markov Model

Thus the previous actions correspond to the previous pages that have already been visited. In Web prediction technique, the Kth order Markov model is the probability that a user will visit the kth page provided that she has visited the ordered k − 1 pages. For example, in the second-order Markov model, prediction of the next Web page is formulated based only on the two Web pages previously visited.

III.1. N-Gram Representation of Sessions

III.4. All-Kth Markov Model

In Web prediction, the best known representation of the training session is the N-gram. N-gram depicts sequences of page clicks by a population of users surfing a Website. Each component of the N-gram takes specific page ID value that identifies a Web page. Many models further process these N-gram sessions by applying a sliding window to make training examples have the same length.

In all-Kth Markov model, we generate all orders of Markov models and utilize them collectively in prediction. The steps of prediction using all kth model. Note that the function predict (x, mk) is assumed to predict the next page visited of session x using the kth order Markov model mk. If the mk fails, the mk−1 is considered using a new session x’ of length k − 1 where x’ is computed by stripping the first page ID in x. This process repeats until prediction is obtained or prediction fails. For example, given a user session x =P1, P5, P6), prediction of all-Kth model is performed by consulting third-order Markov model. If the prediction using thirdorder Markov model fails, then the second-order Markov model is consulted on the session x’ = x - P1 = (P5, P6). This process repeats until reaching the first-order Markov model.

III.2. Association Rule Mining (ARM) ARM is a data mining technique that has been applied successfully to discover related transactions. In ARM, relationships among item sets are discovered based on their co-occurrence in the transactions. Specifically, ARM focuses on associations among frequent itemsets. In a supermarket store example, ARM helps uncover items purchased together which can be utilized for shelving and ordering processes. In the following, we briefly present how we apply ARM in WPP. In WPP, prediction is conducted according to the association rules that satisfy certain support and confidence as follows. For each rule, R = X → Y, of the implication, ie.X is the user session and Y denotes the target destination page. Prediction is resolved as follows: Note that the cardinality of Y>1, i.e., prediction can resolve to more than one page. Moreover, setting the minimum support plays an important role in deciding a prediction. In order to mitigate the problem of no support for X  Y, we can compute prediction (X’ → Y), where X’ is the item set of the original session after trimming the first page in the session. In this process is very similar to the all-Kth Markov model. So that, unlike in all Kth Markov model, mainly in ARM, we do not generate several models for each separate N-gram. For that we will refer to this process as all Kth ARM model in the following section. Several efficient algorithms have been proposed to generate item sets and to uncover association rules such as AIS algorithm. However, ARM endures efficiency and scalability problems. III.3. Markov Model The basic concept of Markov model is to predict the next action depending on the result of previous actions. In Web prediction technique, the next action corresponds to predicting the next page to be visited.


IV.

Modified Markov Model

In this section, we propose another variation of Markov model by reducing the number of paths in the model so that it can fit in the memory and predict faster. Recall that, in Markov model, we consider lists in building the model, for example, user sessions S1 = (P1, P2) and S2 = (P2, P1) are two different sessions; hence, each session can have different prediction probability. In ARM, S1 and S2 are the same item set on the other hand. The basic idea in the modified Markov model is to consider a set of pages in building the prediction model to reduce its size, so that we can easily find out the next page. For example, we consider all the sessions (P1, P2, P3), (P1, P3, P2), (P2, P1, P3), (P2, P3, P1), (P3, P1, P2), and (P3, P2, P1) as one set P1, P2, P3. Our aim is that a task on the Web can be done using different paths regardless of the ordering that the users choose. In addition, we reduce the size of prediction model by discarding sessions that have repeated pages. All these sessions might result when the user accidentally clicks on a link and hits the back button. The Kth order of modified Markov model computes the probability that a user will visit the kth page given that she has visited the k − 1 pages in any order as a ball example. Note that the last page of the session is assumed to be the final destination and it is separated from the sessions. For example, the prediction of P1, P2 is P3 and P4 with probabilities 5/6 and 1/6, respectively; hence, prediction is resolved to P3.


1165


TABLE I TWO-TIER FRAMEWORK TRAINING PROCESS

Input: M is the set of prediction model of size N: T is a set of training examples Output: A set of trained classifiers and an example Classifier EC. 1) For every classifier model m in M train m on T. 2) For each training example e in T and a classifier model m in M Do if m predicts the target of e correctly then map e to m and record the confidence of m in prediction. 3) For each example e in T, if e is mapped to more than one model then filter the labels so that only one label is kept. 4)Train EC on the training set T’, where T’ is a training set that has all examples in t and each example is mapped to one model in M

V.

Two Tier Prediction Framework

In this section, we present a novel framework for Web navigation prediction. The fundamental idea is to generate different prediction models either by using different classification techniques or by using different training samples. A special prediction model, namely, EC, will be generated and later consulted to assign examples to the most appropriate classifier. Note that each predictor, in the generated bag of prediction models, captures strengths and weaknesses of that model depending on many factors such as the set of training examples, the structure, the flexibility, and the noise resiliency of the prediction technique. Table I presents our proposed two-tier framework. In the step 1, we trained the set of classifiers M on the training set T. Such that a subset of T is sufficient for training when T is large. In step 2, the labeling process is applied in which each example e of the training set is labeled with one of the M classifiers that successfully predicts the outcome of e. Filtering is an example that have more than one mapped label is done in the step 3. In step 4, the EC is trained on the filtered mapped training examples. The different input/output of each stage of the framework. At first, all classifiers are trained on the training set T. The output of the training process is Ntrained classifiers. During the mapping stage, each training example e in T is mapped to one or more classifiers that succeed to predict its target. For example, t1 is mapped to classifier C2, while t3 is mapped to the set of classifiers C1, C2. The mapped training set T’ undergoes a filtering process in which each example is mapped to only one classifier according to the confidence/strength of the classifiers. For example, after filtering stage, t3 in T’ is mapped to C1 rather than C1, C2 because C2 predicts t3 correctly with higher probability, for instance. In case the models have equivalent prediction confidences, one model is selected randomly. Finally, the filtered data set FT is used to train the EC as the final output. In this paper, by applying sliding windows on the training set T, we generate N


orders of Markov models, namely, first, second and Nth order Markov models. So that these prediction models represent a repository that can be used in prediction. Next, we map each training example in T to one or more orders of Markov models. (P1, P3, P5) is mapped to two classifier IDs, namely, C1 (first-order Markov model) and C2 (second-order Markov model) in the training example t3. After that, filtering/pruning process is conducted in which each example is mapped to only one classifier. In our experiments, we choose the classifier that predicts correctly with the highest probability. For example, t3 is finally mapped to C1 because it has higher probability than C2 of predicting t3 correctly. Finally, we generate the EC based on training the filtered data set using one of the prediction techniques. Recollect that EC can use any traditional learning technique, such as ANNs or decision trees, and it is special because it is trained based on training examples mapped to classifier IDs. In prediction, a testing example x submits to two stages of prediction. First, x is fed to EC as input to predict the suitable classifier for x Cx. In the next step, x goes through the predicted classifier Cx to determine the final outcome. Note that, in other multi class prediction models, such as bagging and Ada-boosting, all prediction models are consulted, and a confident scheme is used to pick the outcome of the strongest classifier. This is a timeconsuming process, at the time of when prediction is required online. In our framework, only one classifier is consulted which results in less prediction time and less ambiguity during prediction. It is also important to choose the prediction technique for EC carefully; specifically, it should be multi class classifier and non probabilistic model. The reason for requiring a multi class classifier is intuitive since we need a prediction model that can predict for many labels. On the other hand, the reason behind not being a probabilistic technique is twofold. First, we need to uncover hidden associations/relationships between examples and classifiers. Such relationships cannot be found using probability and counting occurrences of examples. Second, for new experiences/examples that do not exist in the training set, the probabilistic models fail to predict because the probability of such new example is zero.

Fig. 1. Prediction Process in the Two Tier Model


1166


So, models such as ANNs, SVM, and decision tree might be good candidates for EC.

VI.

Web Prediction with Boosting and Bagging Association Rules

The two tier architecture is extended with statistical log features. Boosting association rule mining algorithm is integrated with the prediction system. The prediction system is also improved with bagging technique. The preprocessing techniques are also adopted to select optimal data and sessions. The user web browsing behavior identification is done with user access logs. The system is designed to perform pattern extraction and webpage prediction process. Classifiers are used in the prediction process. The system is divided into five major modules. They are Log optimizer, Markov Analysis, Rule Mining Process, Rule Mining with Statistical Features and ARM with Boosting and Bagging. Log optimizer module is designed to perform access log preprocessing operations. Thus the Web user’s behaviors are identified using the Markov analysis module. The rule mining module is designed for pattern discovery and prediction process. The prediction process is improved under statistical features based rule mining. The boosting and bagging technique Arm is used to improve prediction process.The multicast network in NSGC has time-based cluster structure. Initially Key Generation Centre/Group Controller (KGC/GC) assigns the number of sub-groups (cluster) to be constructed and their respective subscription span values based on which the members are grouped. VI.1. Log Optimizer The web page requests are maintained under the access log files. Redundant page requests are removed from the log files. The page requests are formed into group of sessions. Statistical features are extracted from the log files.

VI.4. Rule Mining with Statistical Features The statistical features are extracted from access logs. The rule mining is applied on feature data values. Mean and median features are used in the pattern analysis. The pattern filtering is carried out with the features. VI.5. ARM with Boosting and Bagging The usage pattern mining process is performed on optimized access sequences. The post mining process is initiated to boost up the rule selection process. The rule grouping process is done with bagging techniques. The system improves the prediction accuracy using refined patterns.

VII. Experimental Result and Discussion The web access pattern mining and prediction scheme is analyzed with different log files. The user access log files are collected from the web servers. The system is tested with various prediction technique as Markov model, association rule mining (ARM) and association rule mining with statistical features (ARMSF) methods. The Apriori algorithm is used for association rule mining process. The prediction accuracy is used as the performance metric to evaluate the quality of the system. The system is designed to successfully improve prediction accuracy using simpler probabilistic models such as Markov model and modified Markov models. The modified Markov model handles the excess memory requirements in case of large data sets by reducing the number of paths during the training and testing phases. Finally the system also reduces the prediction time particularly when there are many prediction models to consult. So that system is tested with different set of log transactions. The analysis shows that the association rule mining method produces more than 5% accuracy levels than the modified Markov model. The association rule mining with statistical features model produces more than 10% prediction accuracy level in association rule mining model.

VI.2. Markov Analysis The Markov model is used to predict the next action depending on the result of previous actions. The Kth Markov model is prepared with request paths. The modified Markov model filters the path based on a threshold. The prediction is carried out with patterns extracted from Markov model. VI.3. Rule Mining Process The rule mining process is applied to extract the patterns. The algorithm used here as Aprior algorithm in the rule mining process. Identifying pattern from the item set collections. Support and confidence ratio are used in the prediction process. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

Fig. 2. Prediction Accuracy Analysis on Web Prediction Techniques


1167


The association rule mining based web prediction process is enhanced with boosting and bagging framework. The boosting and bagging framework improves the web prediction accuracy 15% more than the statistical feature based association rule mining model. TABLE II PREDICTION ACCURACY ANALYSIS ON WEB PREDICTION TECHNIQUES Transactions Markov ARM ARM-SF BBF 100 0.573 0.607 0.652 0.851 200 0.594 0.638 0.698 0.877 300 0.608 0.672 0.745 0.901 400 0.626 0.719 0.796 0.927 500 0.649 0.751 0.848 0.945

References

[3]

[4]

[5]

[6]

[7]

[8]

[9]

2

Associate Professor, Department of IT,Bannari Amman Institute of Technology, Sathyamangalam, Erode. Assistant Professor, Department of IT, Sri Krishna College of Engineering and Technology, Coimbatore, Erode.

Web prediction is applied to predict the next set of Web pages that a user may visit based on the historical knowledge. Prediction model supports the search engine to utilize it to cache the next set of pages that the users might visit. Markov model and association rule mining models are integrated to perform page prediction process. Boosting and bagging rule mining models are applied to improve the prediction accuracy. The system handles web user behavior identification and prediction process for all websites. The number of paths is reduced without compromising accuracy. Accuracy improves with higher orders of all Kth model. Prediction time is minimized without compromising prediction accuracy.

[2]

Research Scholar, Department of CSE, Bannari Amman Institute of Technology, Sathyamangalam, Erode.

3

VIII. Conclusion

[1]

Authors’ information 1

Awad.M,Khan.L and Thuraisingham.B“Predicting WWW surfing using multiple evidence combination”,VLDB J., vol. 17, no. 3, pp. 401–417, May 2008. Awad.M and Khan.L“Web navigation prediction using multiple evidence combination and domain knowledge”,IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 37, no. 6, pp. 1054–1062, Nov. 2007. Fu..Y, Paul.H, and Shetty.N“Improving mobile Web navigation using N-Gram prediction model”, Int. J. Intell.Inf.Technol., vol. 3, no. 2, pp. 51–64, 2007. Hassan.M.T, Junejo.K.N and Karim.A“Learning and predicting key Web navigation patterns using Bayesian models”, in Proc. Int. Conf. Comput. Sci. Appl. II, Seoul, Korea, 2009, pp. 877–887. Mamoun Awad.A and Issa Khalil“Prediction of User’s WebBrowsing Behavior: Application of Markov Model” IEEE Transactions on Systems, Man and Cybernetics—Part b: Cybernetics, vol. 42, no. 4, August 2012. Nasraoui.O and Petenes.C,“Combining Web usage mining and fuzzy inference for Website personalization”, in Proc. WebKDD, 2003, pp. 37–46. Nasraoui.O and Krishnapuram.R“One step evolutionary mining of context sensitive associations and Web navigation patterns”, in Proc. SIAM Int. Conf. Data Mining, Arlington, VA, Apr. 2002, pp. 531–547. Nasraoui.O and Krishnapuram.R“An evolutionary approach to mining robust multi-resolution web profiles and context sensitive URL associations” Int. J. Comput. Intell. Appl., vol. 2, no. 3, pp. 339–348, 2002. Levene.M and Loizou.G“Computing the entropy of user navigation in the Web”,Int. J. Inf. Technol.Decision Making, vol. 2, no. 3, pp. 459–476, 2003.


Prof. Sampath Prakasam, Research Scholar, received his M.E in Computer Science & Engineering and is now pursuing his Ph.D.indataminingat AnnaUniversity of Technology, Coimbatore. Currently, he is working as Associate Professor (Senior Grade) in the department Computer Science and Engineering, Bannari Amman Institute of Tech, Sathyamangalam, Erode, Tamil Nadu, India. He has 18 years of experience in teaching field. He has so far published 3 papers in National Conferences and 5 papers in international journals and he has presented 2 papers in International Conferences held at various reputed engineering colleges. Dr. Amitabh Wahi received the B.Sc. and M.Sc degree from LNM University, Darbhanga, PhD degree from Banaras Hindu University, Varanasi. He is currently working as Professor, Department of Information Technology at Bannari Amman Institute of Technology, Sathyamangalam, Tamil Nadu, India. He is the Life member of ISTE and CSI. His research interests include Neural Networks, Fuzzy Logic and Pattern Recognition. He organized many conferences and seminars in the college. He has published more than 50 papers in refereed journals and conferences. He received funded projects from DRDO and AICTE, New Delhi. Ms. Ramya Duraisamy received her M.E Computer Science and Engineering in Bannari Amman Institute of Technology, Sathy. Currently she is working as assistant professor, Dept of IT, Sri Krishna College of Engineering and Technology, Coimbatore. She had published 1 international conference and 3 National conference and 5 International journals.


1168


A Rough Set Based Classification Model for Grading in Adaptive E-Assessment G. S. Nandakumar1, V. Geetha2, B. Surendiran3, S. Thangasamy1 Abstract – Assessment of students is an integral part of learning process which is related to learning outcomes. The goal of assessment is the estimation of the knowledge that has been acquired by the students via learning. Adaptive assessment is a form of computer-based assessment that adapts to the students’ ability level. Adaptivity is the key functionality of this assessment, in which the questions are selected intelligently to fit the student's level of knowledge. The motivation of this work is to investigate the techniques for the improvement of student assessment. An adaptive grading methodology was developed to effectively assess the performance of students based on time taken to answer the questions and the grade scored and it has been found that this was performing better than predefined grading in terms of classification accuracy. This work focuses on an innovative approach of using of Rough Set Theory to model the students’ performance in the E-Assessment data and to generate the classification rules for knowledge discovery about the students’ performance in both predefined and adaptive grading. The results have shown that adaptive grading performs better than predefined grading methodology. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Adaptive E-Assessment, Degree of Toughness (DT), Adaptive Grading, Rough Sets, Classifier Accuracy, Discriminant Analysis

I.

Introduction

The development of new digital technologies has facilitated the implementation of assessment tools. One of the fastest evolving fields among teaching and learning research is, students’ performance evaluation. Integrating computer based assessments with webbased educational systems is an easiest way of performance evaluation and so they are increasingly adopted by most of the educational institutions [1]-[3]. E-Assessments facilitate the teachers to conduct tests at a larger scale and evaluate the students’ knowledge level. Adaptive testing is a form of computer-based assessment that adapts to the students’ ability level. The main objective of adaptive assessment is to conduct an optimal test for all the students. In conventional tests, the same sets of questions are administered to all the students. Students get frustrated when the difficulty of the questions is higher than their ability level and bored if it is lower than their ability. But in the procedure of adaptive testing, the questions will be generated based on the individual ability of the student. This will result in generating different sets of questions for different students, keeping their enthusiasm to face the test steadily. If the student gets correct answer for a question, the computer will then select the next question from a higher difficulty level; otherwise the next question will be from a lower difficulty level. Students do not waste their time attempting questions which are either very trivial or very difficult to their ability.

Students proficiency in a subject will be assessed with a fewer questions than in conventional tests. A finer discrimination in all ranges of performance is possible in adaptive strategy. This will improve the effectiveness of the assessment. Assessing students’ knowledge mastery from empirical data is an important issue in the design of an adaptive intelligent assessment. Accurate assessment is one of the most critical components of a successful learning process. With large amount of data being generated by large scale online learning environments, there is a challenge in tackling this data. The performance evaluation in most of the assessments is based on the numerical test scores which do not reveal the knowledge of the students. Though it is a valid and reliable measure, it does not address all the factors related to students’ success. The student’s ability is a mix of his mental ability, the time devoted to the item, and the persistence given to finding the right solution. A classification model that infers students’ knowledge by assessing their performance in a specific environment has to be developed. Data mining is the process of discovering interesting knowledge from large amounts of data stored in databases, data warehouses, or other information repositories [4]. Data mining can be used to discover patterns occurring in the databases such as associations, classification models, clustering, sequential patterns etc. Machine learning techniques such as decision tree induction, Bayesian networks, genetic programming are incorporated to provide rapid learning and perfect decision making environment needed for applications


1169

G. S. Nandakumar, V. Geetha, B. Surendiran, S. Thangasamy

like adaptive E-assessment. Classification is a primary data mining task that can be used to extract models to describe important data classes or to predict future data trends. It is one of the most useful techniques in data mining to build classification models from an input data set. Classification technique has been used to help in improving the quality of the higher educational system by evaluating student data to study the main attributes that may affect the student performance in courses [5]. Rough set methodology has gained much interest in the data mining field. The main idea behind rough sets is to approximate sets that cannot be precisely or crisply defined. In this work, rough set based classification is used for analyzing the performance of the students. In adaptive E-Assessment the time taken by the students to answer the questions at each difficulty level may be similar but their final scoring may be different thus leading to inconsistent data. Therefore, it becomes very difficult to classify the performance of the students’ based on the answering time and the grade scored using conventional data mining techniques. Hence, there is a need to incorporate the concept of inconsistency in the process of knowledge assessment in E-assessment. Rough sets and decision trees have better knowledge representation structure in deriving decision meaningful decision rules [6]. The rules establish a relationship between the description of objects represented by attributes and their associated class. These rules can be used for classification for new objects. Rough sets have been used in the field of bioinformatics [7]. Rough sets have been used for biomedical datasets [8] and were able to generate a highly accurate classifier. Rough sets have also been used for handling uncertainty for loan payment prediction [9]. The classification of learner’s learning style in E-learning environments has been discussed [10]. A rule-based rough-set decision system for disease identification from ECG signals has been developed [11]. The characteristic behavior of decision trees and rough set algorithms have been analyzed [12] in industrial applications and it has been concluded that rough set algorithms are suited for applications where precise and detailed rules are needed. Two data mining techniques, decision tree and rough set theory were used to predict students’ performance in object oriented programming course[13]. The prediction accuracy of decision tree algorithm was 63%, while that of rough set was 86%. They have shown that rough set theory classifier is able to produce better results when compared to decision tree classifier. Rough set classifier is more suitable for large attribute sets and it produces different rules with good confidence and support [14]. They have concluded that the rules represented by decision tree may be significantly incorrect even for consistent data which has a large number of attributes. This work describes an adaptive model for conducting E assessment using multiple choice questions for knowledge evaluation in 'C' Language.


The model allows the examinee to start the test with the chosen difficulty level strategy has been formulated. The proposed adaptive E-Assessment strategy has an edge over others for the reason of quantifying the marks for the level of difficulties and also the students’ choice in the selection of the appropriate Degree of Toughness (DT). Various techniques are available for generating dynamic E-Assessment system with different difficulty levels of questions, but performance evaluations are done based on a predefined grading system. The E-Assessment grading depends on the examinee’s problem solving capability, time of the test and their performance. A new adaptive E-Assessment grading methodology had been proposed to effectively assess the examinee performance based on time, answering habits and the test results. In our previous work, an adaptive EAssessment grading (AEAG) model has been proposed to overcome the limitations of the existing grading system [15]. The empirical data was collected and the relationship between the response time and the assessment grading was established using Bayesian classifier [16]. The proposed work is an experimental result of applying rough set theory in evaluating the performance of the students in adaptive E-assessment in both predefined and adaptive grading.

II.

Proposed Adaptive E-Assessment Evaluation System

This section describes the adaptive strategy followed for E-Assessment and the creation of a knowledge base. Application has been developed using PHP software as the front end and MySQL at the back end database server to implement the procedure. II.1.

Knowledge Base Creation

A multiple choice question bank for C-Programming language was created by collecting questions from subject experts. A conventional test (where each of the questions in the question bank has to be answered by all the students) was given to a group of students. Calibration was done with the proportion of the examinees who answered each question correctly to the total population, based on which the questions were initially classified into various levels of DT ranging from 1 to 5 (1 – Novice and 5 – Very Difficult). Each question in the question bank is tagged with a DT. The DT of a question has to be updated periodically, after broad spectrums of students undergo the tests and the question has been asked sufficiently large number of times. A difficult question is assigned a higher weightage than a less difficult question.Deciding the next question’s degree of toughness is based on various factors as shown in (1) below: deciding factor (Qi , DT , result , nDT )

(1)


1170


where: Qi is the current question with ith DT, DT is the degree of toughness, Result is the outcome of the student’s answer for the current question, nDT is the number of questions answered with degree of toughness i , and 1≤ i ≤ 5. The overall process flow is depicted in Fig. 1.

Knowledge Base

Select MCQ by Adaptive E-Assessment Strategy until Stopping Criteria is met

Result Analysis

Grading

Display Results with Grades

Classification of Performance Fig. 1. Adaptive E-Assessment Process flow

II.2.

Case 1: If the candidate answers the first three questions of the kth DT correctly, the system will automatically shift to (k+1)th DT, provided k ≠ 5. When k = 5, the system continues to ask questions from the same level. Case 2: In case the candidate answers all the three questions of the kth DT incorrectly, the system will automatically shift to (k-1) th DT, provided k ≠ 1. It follows from the earlier logic the system continues to display from the 1 st DT irrespective of the number of wrong answers provided. Case 3: This case relates the situation where the examinee answers either one or two questions correctly out of the first three questions from the k th DT. The system exhibits one more question from the same DT. Thus the examinee encounters a total of four questions. A total of three correct answers shifts to (k+1) th DT, provided k ≠ 5; otherwise to (k-1) th DT, provided k ≠ 1. Case 4: In case examinee answers two questions correctly out of the first four questions from the k th DT, one more question from the same DT is given. A total of three correct answers out of five given questions, shifts to (k+1) th DT; otherwise to (k-1) th DT. However shifting to a higher or lower DT does not take place when k=5 or k=1 respectively. The test will get terminated either on the expiry of the time frame or the examinee has attempted questions for the prescribed maximum marks, whichever occurs first. The score and the number of DT – wise questions asked and answered get displayed at the end of the test. II.3.

Procedure

The algorithm for conducting the online assessment using adaptive strategy is given below. The interesting aspect of this model is that it allows the student to initially opt for the DT of the questions soon after he logs into the system of examination. If he opts for the kth DT (k=1, 2, 3, 4, 5) the system will start displaying the questions randomly from the kth for which the candidate answers.

Adaptive E-Assessment Procedure

The adaptive E-assessment was administered to different sets (both computer science and non computer science) of prefinal year Engineering students for a maximum of 15 marks. The assessment dataset is composed of data collected from 275 students and the sample data is shown in Table I. Data pertaining to the time taken for answering each question correctly or wrongly at each DT level, the marks scored at each level, the total mark scored in the entire test and the total time taken to complete the test were collected.

TABLE I ADAPTIVE ASSESSMENT SAMPLE DATA Roll No DT1_CA DT2_CA DT3_CA DT4_CA DT5_CA DT1_WA DT2_WA DT3_WA 13bcs01 18.76 24.50 0.00 0.00 12.35 29.85 44.60 0.00 13bcs02 0.00 19.00 54.33 26.00 0.00 0.00 0.00 56.00 13bcs03 75.67 33.25 108.83 96.67 0.00 78.00 0.00 102.00 13bcs04 20.00 29.29 29.75 53.00 0.00 67.00 13.00 130.88 13bcs06 57.00 31.08 25.50 0.00 0.00 0.00 13.00 34.80 13bcs08 23.56 26.78 37.25 0.00 0.00 40.60 24.86 128.33 13bcs09 10.50 19.00 48.17 188.50 0.00 76.83 30.50 80.00 13bcs10 21.43 27.44 18.00 0.00 25.20 26.00 25.78 32.00 13bcs11 24.89 25.00 38.50 0.00 0.00 33.13 39.63 50.17 13bcs12 26.00 47.67 71.83 67.00 0.00 13.00 47.25 77.80 13bcs13 30.00 50.67 60.00 19.00 0.00 32.00 43.00 19.00 13bcs14 29.28 23.00 0.00 0.00 0.00 45.73 45.57 0.00 13bcs15 0.00 40.78 55.50 0.00 0.00 133.00 73.33 116.60 13bcs17 17.47 55.86 0.00 0.00 0.00 27.79 50.36 0.00 13bcs18 0.00 0.00 60.22 48.00 0.00 0.00 0.00 100.25 DTi_CA - Average time ( in seconds) taken to answer the questions correctly at DT level i ( i = 1,2,…5) DTi_WA - Average time ( in seconds) taken to answer the questions wrongly at DT level i ( i = 1,2,…5)


DT4_WA 0.00 56.67 106.50 91.33 0.00 83.50 43.00 0.00 0.00 144.00 52.17 0.00 79.50 0.00 98.17

DT5_WA 30.57 0.00 0.00 0.00 0.00 0.00 0.00 48.20 0.00 0.00 0.00 0.00 0.00 0.00 0.00

Grade C B O B O A A C A A B C B B A


1171


The marks associated with each difficulty levels also varied. The DT levels are shown in Table II. TABLE II DT LEVELS OF ADAPTIVE E-ASSESSMENT DT level Description 1 Novice 2 Easy 3 Moderate 4 Difficult 5 Very difficult

Measures like the time taken for correctly answering (CA) and wrongly answering (WA) at each difficulty level and the correctness of the answer were collected and stored. Based on the time taken for giving correct and incorrect answers at each level, the average time was calculated. Students were assigned predefined grades as shown in Table III. Predefined grading has a grading schema associated with numerical scores. A grading schema defines the values of letter grades and customized grades. TABLE III PREDEFINED GRADES FOR ADAPTIVE E-ASSESSMENT Grade Test score

O > m1

A B >=m2 and =m3 and xj, find out the ordered pair list using the formula .

=

Step2: Calculate ranking function F (x) outputs a score for each data pair, from which a global ordering of data is constructed for any xi >xj :

.

.

+

, = 1, . . ,

(10)

Input: Training set and testing set as N= xi ,yi , xi ∈ Rn , yi ∈ Rm , i=1,…,N , Kernel function f(x) and hidden neuron Ñ, input weights (wi), biases bi

f(xi )>f(xj )⇔w.xi >w.xj Step3: Find out the weight vector ( ) for all data pairs.w is learnt from support vectors. Subsequently, to minimize the weight or maximizing the margin, it is converted into optimization problem as:

Compute: learning weight factor ′’ Output: ranked alias list.

1 min= w.w+C 2

ξij

Step1: Define the hidden layer node Ñ, randomly assign input weights wi and hidden layer biases bi ,(i= 1,2,… Ñ).

i,j∈|R|

subject to: ∀

xi ,xj :yi Estimated pixel count, and the intensity value at the moment is used as a threshold for initial candidate’s selection. In order to check the area criterion, the pixels in each initial candidate are counted, and the candidates having pixel count more than 125 % or less than 25% of estimated pixel counts are discarded. Then the density criterion is applied on the candidates’ satisfying area criterion. According to the density criterion, the candidates having pixel count less than 40% of the number of pixels occupied by rectangle surrounding that candidates are discarded. Finally, the entropy of vascular direction for the centriod of the remaining candidates is calculated. The entropy of vascular is calculated using (3): (3)

Figs. 3. Vessel Extraction process: (a) RGB fundus image, (DRIVE 12), (b) Green channel image (c) Gray scale image of green band, (d) CLAHE’d image, (e) Vessel enhancement by Gabor filtering, (f) Vessel extraction using thresholding

III.3. OD Localization Algorithm Step 1: Threshold estimation for candidate selection. Step 2: Initial candidate’s selection using thresholding. Step 3: Select candidate regions which satisfying area criterion; If no candidate selected, decrease threshold and go to step 2. Step 4: Select candidate regions which satisfying density criterion; If no candidate selected, decrease threshold and go to step 2.


where n – number of directions. – Probability of occurrence of vessels in direction i. To calculate the entropy, a histogram of vessel directions needs to be determined in a window centered at that point. The window size of 151 × 75 was set for the images of the DRIVE and STARE databases and it can be easily adjusted for other databases. The candidate having maximum entropy and the value of entropy is greater than .5 is identified as OD region. If no such candidate is identified, then the threshold is decreased and the OD localization process is repeated with new threshold value. The obtained results of various stages of OD


1286

J. Benadict Raja, C. G. Ravichandran

localization on DRIVE_23 and STARE_44 images are shown in Figs. 4 and Figs. 5 respectively. DRIVE_23 is a non-pathological retinal image, but it is affected by nonuniform illumination. STARE_44 is a pathological image, it contains bright lesions.

Each image has been JPEG compressed. The database contains 33 normal and 7 pathological images with exudates, hemorrhages and pigment epithelium changes. The images were acquired using a Canon CR5 nonmydriatic camera with a 45◦ field of view (FOV). Each image was captured using 8-bits per color plane at 768 by 584 pixels. The STARE database was composed with the intention to create a difficult database. The database contains 31 images of healthy retinas and 50 of pathological retinas. The pathological images exhibit a wide variety of lesions and other symptoms of diseases. The images were acquired using a TopCon TRV-50 fundus camera at 35◦ FOV, and subsequently digitized at 605 × 700 pixels in resolution, 24 bits per pixel. IV.2. Performance Measures Success rate (accuracy) and computation time are the most commonly used measures for evaluating OD localization methods. Once the detected OD center is within the circumference of the OD in the reference standard, it is then considered to be a successful detection

Figs. 4. OD localization (image 23 of DRIVE): (a) RGB fundus image, (b) Initial candidates, (c) Candidates satisfying area criterion, (d) Candidates satisfying density criterion, (e) Vessel extracted image, (f) Blood vessels on final candidates, (g) The candidate with maximum and optimum entropy value (OD region), (h) OD localized gray scale image (black dot represent the OD position)

IV.3. Results on DRIVE and STARE Databases The method was implemented on 3.06 GHz core i3 processor with 3.17 GB RAM using MATLAB 7.6. The proposed method achieved a success rate of 100% (i.e., the OD was detected correctly in 40 out of the 40 images contained in the DRIVE dataset). Additionally, the OD was detected correctly in the pathological images of STARE database. The method is fully automatic and localizes the OD of one DRIVE image within 2.2s without optimization of its Matlab code (See Table I). The computation time for all the 40 images of DRIVE dataset is reported in Table I. In order to obtain an optimal vessel structure the vessel extraction algorithm is repeated for several times; therefore the vessel extraction time is differed for healthy and un-healthy retinas. Similarly, the OD localization time is differed for images based on the illumination of the image. The average computation time for processing a DRIVE image is only 2.2 seconds.

Figs. 5. OD localization (image 44 of STARE): (a) RGB fundus image, (b) Initial candidates, (c) Candidates satisfying area criterion, (d) Candidates satisfying density criterion, (e) Vessel extracted image (f) Blood vessels on final candidates, (g) The candidate with maximum and optimum entropy value (OD region), (h) OD localized gray scale image (black dot represent the OD position)

IV.

IV.4. Discussions

Results and Discussions IV.1. Material

The performance of the proposed method is evaluated on publicly available DRIVE (Digital Retinal Images for Vessel Extraction) and STARE (STructural Analysis of REtina) databases. The DRIVE database photographs were obtained from a diabetic retinopathy screening program in Netherlands. The screening population consisted of 453 subjects between 31 and 86 years of age.


In Figures 6 and 7, the results of the proposed method for some retinal images of DRIVE and STARE datasets are shown. In presence of abnormality in the eye, using the vascular information is more effective. Pathological regions, exudates, and optic disc are bright regions in the retina images. Therefore, methods such as brightness-based or template matching or methods which are based on the segmentation results of blood vessels fail to localize the center of optic disc in presence of pathological regions and exudates in retina image.


1287


Figs. 6. Obtained results on DRIVE images (a) RGB fundus image (b) Vessel extracted image(c) Final candidate (d) OD localized gray scale image(black dot represent the OD position)



1288


Figs. 7. Obtained results on STARE images (a) RGB fundus image (b) Vessel extracted image(c) Final candidate (d) OD localized gray scale image (black dot represent the OD position)



1289


(a) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

TABLE I COMPUTATION TIME ON DRIVE IMAGES (A) IMAGE NO (B) VESSEL EXTRACTION TIME IN SECONDS (C) OD LOCALIZATION TIME IN SECONDS (D) COMPUTATION TIME IN SECONDS (b) (c) (d) (a) (b) (c) 3.30 0.09 3.39 21 1.38 0.03 3.81 0.06 3.88 22 2.56 0.05 1.33 0.05 1.38 23 2.58 0.19 1.34 0.05 1.39 24 2.55 0.05 1.36 0.03 1.39 25 1.36 0.03 1.36 0.13 1.48 26 1.34 0.03 2.64 0.03 2.67 27 2.53 0.03 1.36 0.05 1.41 28 2.56 0.05 1.33 0.03 1.36 29 2.56 0.05 1.36 0.03 1.39 30 1.55 0.05 2.56 0.05 2.61 31 1.34 0.05 2.55 0.03 2.58 32 1.38 0.03 2.52 0.03 2.55 33 1.34 0.03 2.58 0.08 2.66 34 5.06 0.03 2.61 0.05 2.66 35 2.52 0.05 2.58 0.03 2.61 36 2.55 0.05 1.31 0.05 1.36 37 2.53 0.03 2.59 0.03 2.63 38 2.53 0.03 2.58 0.03 2.61 39 2.55 0.05 1.34 0.03 1.38 40 1.31 0.03 Average Processing Time 2.16 0.05

Figs. 6 and 7 show the result of the proposed method on normal retina image and retina images with pathological regions and exudates. Despite the existence of dark hemorrhages or bright exudates and pathological regions, the results of the proposed method are satisfactory and it shows the effectiveness of the proposed method for localizing the center of optic disc. In Figs. 8, some retinal images with incorrectly detected optic disc center are shown. In the first row of Figs. 8, there is not any vessel in vicinity of optic disc and the characteristic of optic disc like brightness and high number of vessels in vicinity of optic disc cannot be seen; therefore, our proposed method failed to localize the optic disc center. For the retinal image in the second row of Figure 8, optic disc is in the corner of image and there is really no vessel in optic disc. Therefore, our proposed method failed to localize the optic disc center. Table II compares the performance of our method with the performance of other solutions reported in the literature for DRIVE dataset.

(d) 1.41 2.61 2.77 2.59 1.39 1.38 2.56 2.61 2.61 1.59 1.39 1.41 1.38 5.09 2.56 2.59 2.56 2.56 2.59 1.34 2.21

TABLE II PERFORMANCE COMPARISION OF DIFFERENT TECHNIQUE ON THE DRIVE DATABASE Technique Success Rate in % Compu-tation time in seconds Configuration of the PC Sinthanayothin et al. [9] 60 NA NA Walter and Klein[10] 77.5 NA NA Hsiao et al. [11] 93 NA NA Godse and Bormane[13] 100 NA NA Lalonde et al. [14] 80.6 NA NA Dehghani et al. [16] 100 27.6 Intel Core 2 Duo, 2.67 GHZ, 3.24 GB Sopharak et al. [20] 98.61 NA NA Zhu et al. [23] 90.4 NA NA Lu and Lim[24] 97.5 270* NA Youssif et al.[25] 100 210 Intel Core 2 Duo, 1.7 GHZ, 512 MB Foracchia et al.[26] 100 120* Intel Core 2 Duo, 2.0 GHZ, 512 MB Rangayyan et al. [28] 100 2294 Intel Core 2 Duo, 2.5 GHZ, 1.96 GB Mendonca et al. [29] 100 8 NA Ying et al. [31] 97.5 NA NA Welfer et al. [32] 100 7.89** Intel core 2 Quad,2.4 GHZ, 4.0 GB Park et al. [33] 90.3 4 Intel 1.0 GHZ, 1.0 GB Ravishankar et al. [35] 86.1 NA NA Proposed 100 2.2 Intel core i3, 3.06 GHz and 3.17 GB RAM * computation time on STARE image, ** computation time on ImageRet image

Figs 8. Retinal images with incorrectly detected optic disc center: (a) RGB fundus image, (b) Vessel extracted image, (c) Final candidate, (d) OD localized gray scale image (black dot represent the OD position)



1290


The results of other methods are obtained from their original papers. The most recent methods such us Godse and Bormane [15], Dehghani et al. [18], Youssif et al. [28], Foracchia et al. [29], Mendonca et al. [32], Welfer et al. [35], and Rangayyan et al. [31] have obtained 100% success rate on DRIVE database. The proposed work also has obtained 100% success rate on DRIVE dataset with very low computation time.

V.

Conclusion

We have proposed a fast method for detecting optic disc, in retinal images, by combining intensity and vascular information in this paper. Intensity-based algorithms are fast, simple, and reasonably robust only for normal and healthy images. Therefore we combine both intensity and vascular information for OD localization. We have evaluated the proposed algorithm on publicly available DRIVE and STARE databases. A significant reduction in computation time and the attainment of a useful OD localization even in pathological or poor quality images are the most relevant improvements of this work. In addition to giving competitive performance, the proposed approach does not have any parameters to be tuned depending on datasets. The OD location obtained with this automatic approach will be used as an initial step for the segmentation OD area.

References [1]

K.G. Alberti, P.Z. Zimmet, Definition, diagnosis and classification of diabetes mellitus and its complications, part 1: diagnosis and classification of diabetes mellitus provisional report of a WHO consultation. Diabet. Med. , vol. 15 no. 7,1998, pp. 539–553. [2] R.N. Frank, Diabetic retinopathy. Prog. Retin. Eye Res. , vol. 14 no. 2,1995, pp. 361–392. [3] O. Faust, U.R. Acharya, E.Y.K. Ng, K.H. Ng, J.S. Suri, Algorithms for the automated detection of diabetic retinpathy using digital fundus images: A review, J Med Syst, vol. 36 no. 1,2012, pp. 145-157. [4] A. Hoover, M. Gold Baum , Locating the Optic Nerve in a Retinal Image Usingthe Fuzzy Convergence of the Blood Vessels, IEEE Transactions on Medical Imaging, vol. 22 no. 8,2003,pp. 951-958. [5] J. Lowell, A. Hunter, D. Steel, A. Basu, R. Ryder, E. Fletcher, L. Kennedy, Optic nerve head segmentation, IEEE Transactions on Medical Imaging, vol. 23 no. 2,2004,pp. 256-264. [6] DRIVE: Digital Retinal Images for Vessel Extraction. URL:/http://www.isi.uu.nl/Research/Databases/DRIVE/S [7] M.R.K. Mookiah, U.R. Acharya, C.K. Chua, .M. Lim, E.Y.K. Ng, A. Lude, Computer-aided diagnosis of diabetic retinopathy: A review, Computers in Biology and Medicine, vol. 43 no. 1,2013, pp. 2136-2155. [8] M. S. Haleem, L. Han, J. Hemert, B. Li ,Automatic extraction of retinal features from colour retinal images for glaucoma diagnosis: A review, Computerized medical imaging and graphics, vol. 37 no. 1,2013, pp. 581-596. [9] C. Sinthanayothin, J.F. Boyce, H.L. Cook, T. H. Williamson, Automated localization of the optic disc, fovea, and retinal blood vessels from digital colour fundus images, Br J Opthalmal, vol. 83 no. 8,1999,pp. 902-910. [10] T. Walter, J.C. Klein, Segmentation of color fundus images of the human retina: detection of the optic disc and the vascular tree using morphological techniques, in: Proc. 2nd Int. Symp. Med. Data Anal., 2001, pp. 282–287.


[11] H.K Hsiao, C.C. Liu, C.Y. Yu, S.W. Kuo, S.S. Yu, A novel optic disc detection scheme on retinal images, Expert Systems with Applications, vol. 39 no. 1,2012,pp. 10600–10606. [12] F.T. Haar, Automatic localization of the optic disc in digital clolour images of the human retina, M.S thesis, Utrecht University, Utrecht, The Netherlands, 2005. [13] D.A. Godse, D.S. Bormane, Automated Localization of Optic Disc in Retinal Images, International Journal of Advanced Computer Science and Applications, vol. 4 no. 2,2013, pp. 65-71. [14] M. Lalonde, M. Beaulieu, L. Gagnon, Fast and robust optic disc detection using pyramidal decomposition and Hausdorff-based template matching, IEEE Trans. Med. Imaging, vol. 20 no. 11,2001, pp. 1193–1200. [15] A. Osareh, M. Mirmehdi, B. Thomas, R. Markham, Automated identification of Diabetic Retinal Exudates in digital colour images, Br J Opthalmal, vol. 87 no. 10,2003,pp. 1220-1223. [16] A. Dehghani, H.A. Moghaddam, M.S. Moin, Optic disc localization in retinal images using histogram matching, EURASIP Journal on Image and Video Processing,2012 [17] S.A. Ramakanth, R.V. Babu, Approximate nearest neighbor field based optic disk detection, Computerized medical imaging and graphics, vol. 38 no. 1,2014,pp. 49-56. [18] H. Yu, E.S. Barriga, C. Agurto, S. Echegaray, M.S. Pattichis, W. Bauman, P. Soliz, Fast localization and segmentation of optic disk in retinal images using directional matched filtering and level sets, IEEE Transactions on information technology in biomedicine, vol. 16 no.4,2012, pp. 644-657. [19] H. Li, O.Chutatape, Automated feature extraction in color retinal images by a model based approach, IEEE Transactions on Biomedical Engineering, vol. 5 no. 2,2004,pp. 246-254. [20] A. Sopharak, B. Uyyanovara, S. Barmanb, T.H. Williamson, Automatic detection of diabetic retinopathy exudates from nondilated retinal images using mathematical morphology methods, Computerized medical imaging and graphics, vol. 32 no. 1,2008, pp. 720-727. [21] S. Lu, Accurate and efficient optic disc detection and segmentation by a circular transformation, IEEE Transactions on medical imaging, vol. 30 no. 2,2011, pp. 2126-2133. [22] A.D. Fleming, K.A. Goatman, S. Philip, J.A. Olson, P.F. Sharp, Automatic detection of retinal anatomy to assist diabetic retinopathy screening, Phys. Med. Biol. , vol. 52 no. 1,2007, pp. 331-. [23] X. Zhu, R.M. Rangayyan, A.L. Ells, Detection of the optic nerve head in fundus images of the retina using the hough transform for circles. J. Digit. Imag., vol. 23 no. 3,2010, pp. 332–341. [24] S. Lu, J.H. Lim, Automatic optic disc detection from retinal images by a line operator, IEEE Transactions on Biomedical Engineering, vol. 58 no. 1,2011,pp. 88-94. [25] A. Youssif, A.Z. Ghalwash, A. Ghoneim, Optic disc detection from normalized digital fundus images by means of a vessels direction matched filter, IEEE Trans. Med. Imag. vol. 27 no. 1,2008,pp. 11-18. [26] M. Foracchia, E. Grisan, A. Ruggeri, Detection of optic disc in retinal images by means of a geometrical model of vessel structure, IEEE Trans. Med. Imaging , vol. 23 no. 10,2004, pp. 1189–1195. [27] A.E. Mahfouz, A.S. Fahmy, Fast Localization of the optic disc using projection of image features, IEEE Transactions on Image Processing, vol. 19 no. 12,2010,pp. 3285-3289. [28] R.M. Rangayyan, X. Zhu, F.J. Ayres, A.L. Ells, Detection of the optic nerve head in fundus images of the retina with Gabor filters and phase portrait analysis. J. Digit. Imag., vol. 23 no. 4,2010, pp. 438–453. [29] A.M. Mendonca¸ A. Sousaa, L. Mendonca, A. Campilho , Automatic localization of the optic disc by combining vascular and intensity information, Computerized Medical Imaging and Graphics , vol. 37 no. 1,2013,pp. 409-417. [30] M. Niemeijer, M.D. Abràmoff, B.V. Ginneken, Fast detection of the optic disc and fovea in color fundus photographs, Medical Image Analysis, vol. 13 no. 1,2009,pp. 859–870. [31] H. Ying, M. Zhang , J-C. Liu, Fractal-based automatic localization and segmentation of optic disc in retinal images. In: Proceedings of the 29th Annual International Conference of the IEEE EMBS. 2007, pp. 4139–4141.


1291


[32] D. Welfer, J. Scharcanski, C.M. Kitamura, M.M.D. Pizzol, L.W.B. Ludwig, D.R. Marinho, Segmentation of the optic disk in color eye fundus images using an adaptive morphological approach, Computers in biology and medicine, vol. 40 no. 1,2010, pp. 124-137. [33] M. Park, J.S. Jin, S. Luo, Locating the optic disc in retinal images, in: Proceedings of the International Conference on Computer Graphics, Imaging and Visualisation, IEEE, Sydney, Australia, 2006, pp. 14 –145. [34] K.W. Tobin, E. Chaum, V.P. Govindasamy, T.P. Karnowski, Detection of anatomic structures in human retinal imagery, IEEE Transactions on Medical Imaging , vol. 26 no. 12,2007, pp. 17291739. [35] S. Ravishankar, A. Jain, A. Mittal, Automated feature extraction for early detection of diabetic retinopathy in fundus images, in: IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 210–217. [36] C.G.Ravichandran, J.Benadict Raja, A fast enhancement/thresholding based blood vessel segmentation for retinal image using contrast limited adaptive histogram equalization, J.Med. Imaging Health Inf., 4,2014, pp. 567-575. [37] Chauhan, R., Manocha, P., Chandwani, G., Retinal vessel extraction and vessel path prediction by active contouring, (2014) International Review on Computers and Software (IRECOS), 9 (3), pp. 450-45. [38] Karthikeyan, S., Rengarajan, N., Hybrid feature analysis for assessment of glaucoma using RNFL defects, (2014) International Review on Computers and Software (IRECOS), 9 (1), pp. 178-18.

Authors’ information 1,*

Department of Computer Science and Engineering, PSNA College of Engineering and Technology, Dindigul, T. N, India. E-mail: [email protected] 2

Excel Engineering College, Komarapalayam, T. N., India. E-mail: [email protected] J. Benadict Raja received the M.E. degree in Computer Science and Engineering from Anna University, Chennai,India in 2008. Since 2011 he has been pursuing the Ph.D. degree at Anna University, Chennai, India. His current research interests include Medical Image Processing, Parallel programming and Parallel Architecture.

C. G. Ravichandran received the Ph.D degree in Medical Image Processing from Anna University, Chennai,India in 2009.Currently he is a Principal of RVS College of Engineering and Technology Dindigul,Tamilnadu,India. His research interests include image segmentation, medical image processing, network architecture, distributed systems and web services.



1292


Optimized Network Selection Using Aggregate Utility Function in Heterogeneous Wireless Networks C. Amali, Dhanasree Jayaprakash, B. Ramachandran Abstract – To provide global connectivity with high speed and high quality at any place and any time is now becoming a reality due to the integration and co-ordination of different Radio Access Technologies(RAT) such as Worldwide Interoperability for Microwave Access (WiMAX), Universal mobile Telecommunication Systems (UMTS), Wireless Local Area network (WLAN) and Long Term Evolution (LTE).Such a diversity of networks offers different choices in terms of bandwidth, security, cost, latency and the coverage area for the mobile user. For such a heterogeneous wireless environment, it is important to design an intelligent handover algorithm which selects the optimal target network. The existing works do not consider the interdependence between the criteria and the use of application QoS in weight calculation during network evaluation process. To address this issue, an appropriate multicriteria network selection algorithm based on multiplicative weighted utility function is proposed to provide complete solution for seamless connectivity in heterogeneous environment based on network conditions, application QoS, Mobile Terminal (MT) battery level and user preferences. MATLAB based simulations are conducted to highlight the effectiveness of proposed scheme and simulation results confirm that the proposed scheme selects the best suitable network for both Real-Time (RT) and Non- Real Time (NRT) applications while achieving optimization between QoS, cost and energy consumption. The Cobb-Douglas based user satisfaction degree is also estimated to verify whether the selected network maximizes the end user satisfaction for the offered service. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Heterogeneous Networks, Network Evaluation, Multicriteria Network Selection, Multiplicative Weighted Utility Function, Quality of Service

Nomenclature U(x) ui(xi) i=1,…,n xl, xm and xu ζ wi µ ε udelay(x) udatarate(x) uenergy(x) ucost(x)

Global utility function Elementary utility function of parameter ‘i' Number of parameters considered lower, threshold and upper values of parameter for particular application User sensitivity Weight values assigned Sensitivity to QoS Sensitivity to cost Utility function of delay Utility function of data rate Utility function of energy Utility function of cost

I.

Introduction

Future wireless communication systems expect a user to access services independent of its location in order to realize the concept of “Always Best Connected” (ABC) [1]. The wireless technologies in a heterogeneous wireless network usually differ in terms of their offered bandwidths, operating frequencies, cost, coverage areas and latencies.

Due to the heterogeneity and the diversity of access networks, various user applications with different QoS requirements pose new challenges on the multi – interface MT in designing optimal network selection algorithm for guaranteeing seamless QoS support to the users. This is where, the need for a well- organized Vertical Handover (VHO) decision algorithm for heterogeneous wireless technologies become evident. VHO is the seamless transfer of an ongoing user session between different heterogeneous networks. The Vertical Handover Decision (VHD) process determines when and where to hand over in a heterogeneous environment when the user is on the move. The strength of 4G system relies on integrating the existing and newly developed wireless technologies instead of putting efforts into developing new radio interfaces and technologies to provide seamless connectivity and better service quality for mobile users. Thus, users will need to take the benefit from the diversity and also can seamlessly switch between different wireless systems to improve the QoS for mobile users. Most of the studies on VHO employ Received Signal Strength (RSS) as the basic handover decision indicator.


1293

C. Amali, Dhanasree Jayaprakash, B. Ramachandran

Considering RSS alone is not applicable for VHD in heterogeneous environment since each network uses different threshold values to determine the quality of the link. The criteria involved in VHO decisions are very important to take decisions for switching to the target network from both the application requirements and MT capabilities. Among the different approaches addressing the multicriteria decision problem provided in the literature, a utility theory based approach is considered since it determines the ability of the network to satisfy the QoS requirements of a particular service. The criteria involved in the network selection process decide the performance of the proposed approach. In this paper, the drawbacks of the existing aggregate utility models are discussed, and then a new multiplicative weighted utility function is proposed. The key features of our proposed approach are the following. i) The use of data rate and delay requirements of service to quantify the quality of particular access network. ii) The use of energy consumption metric for different class of service calculated according to the data rate requirements. iii) The use of sigmoidal function to model the elasticity of applications. iv) Providing optimization between QoS, energy consumption and cost. The rest of this paper is structured as follows. The section 2 summarizes the review of existing related works. The sequence of operations to be carried out in the analysis of network selection process is described in section 3. Section 4 presents the numerical results to evaluate the effectiveness of the proposed scheme. Finally the paper is concluded with possible future extension.

II.

Related Works

Recently, various network selection algorithms based on multiple criteria have been developed to improve the VHD in heterogeneous wireless networks. Ding Zhang et al. [2] proposed access selection algorithm in which Fuzzy Analytic Hierarchy Process (FAHP) is used to find the weights of attributes and Gray Relational Analysis (GRA) determines the final ranking of networks by considering the parameters such as RSS, bandwidth, delay, network load, packet loss rate and security. In [3], Genetic Algorithm (GA) is used to reduce the in consistency of FAHP in VHD process. The performance of GA is analyzed and compared with different orders of pair wise comparison matrix of FAHP. In [4] the handover decision is made based on the comparison of cost functions of different access networks. The cost function is estimated based on network characteristics and user preferences including bandwidth, connection delay, RSS and service cost. Determining the most suitable weights for the criteria is the main problem in Multi Attribute Decision Making (MADM) algorithms.


A. Singhrova et al. [5] developed a Neuro fuzzy based vertical handover decision algorithm which considers RSS, velocity, bandwidth, number of users, battery level and coverage. Adaptive Neuro Fuzzy Interference System (ANFIS) based system improves the throughput and QoS by reducing ping-pong effect compared to fuzzy technique. Faisal Kaleen et al. [6] presented the implementation of fuzzy preprocessing module in which the handover necessity is performed to trigger the handover process in the proper time while achieving the end user’s satisfaction by providing uninterrupted QoS to the users. In [7], a new intelligent Vertical Handover scheme is presented that utilizes Fuzzy Logic based Linguistic Variables to estimate the necessity of handover and also to determine a new point of attachment in order to fulfill the end users requirements. As each traffic has different set of QoS requirements, separate Fuzzy Logic Controllers (FLC) are used for each traffic to improve the overall performance of proposed system. In [8], a hybrid multimedia delivery solution named as Adapt or Handover is presented which balances an energy-quality-cost trade-off by employing a combined adaptive multimedia delivery mechanism. Ioannis Chamodrakas et al. [9] proposed Energy efficient network selection method which incorporates the use of utility function to model the diverse QoS elasticity of different application and also estimates the energy consumption for real time and non-real time application. In [10], a Modified Weight Function based Network Selection Algorithm (MWF-NSA) that considers user preference, network condition and application profile has been proposed to evaluate the utility functions of the networks. In this approach, null utility effect is not considered which may cause wrong network selection. Quoc-Thinh Nguyen-Vuong et al. [11] approached multicriteria utility function based network selection method that uses sigmoidal function to determine the elementary utility of each attribute. These elementary utilities are then aggregated to find the global utility of each network by considering the interdependence among the criteria. Lusheng Wang et al. [12] provided the various mathematical models and theories such as utility theory, MADM, Fuzzy logic and Game theory in order to achieve the throughput optimization and low complexity in network selection problems. In [13], Cobb-Douglas utility function is introduced to measure the subscriber’s satisfaction to different networks, which considers both user preference and QoS criterion. Giupponi et al. [14] proposed a Joint Radio Resource Management (JRRM) algorithm which guarantees user acceptance probability for a given service and also increases the network operator’s revenue. Jun Tu et al. [15] proposed media independent based VHD developed by IEEE802.21 standard to support the integration of Wi-Fi and WiMAX networks. This approach has improved the performance of networks by decreasing the call blocking rate and handover delay.


1294


In [16], QoS aware vertical handover decision making algorithm is proposed to decide the best suitable network for VHO in 4G wireless networks based on RSS, coverage area, data rate, velocity and latency. After an extensive literature review, it is found that the utility function has been used to select the optimum network in order to satisfy the QoS requirements of application. For example, if a single attribute (either data rate or delay) is used for network selection, then the selected network may not satisfy both the data rate and delay requirements of a particular service. Thus, there is a need to consider the multiple characteristics of network to calculate the utility value in the network selection mechanism. Any utility function can be expressed mathematically as a function of attributes and user preferences. User preference reflects the importance of each metric in the selection process according to the requirements of running application (e.g., Voice over IP, Video Streaming applications) and the type of user. Different utility forms are used in the literature in order to normalize the different attributes with different units in the network selection strategies. Service adaptive Multiplicative Weighted Utility function is proposed which exploits the sigmoid function to model the elasticity of applications and also to provide optimization between QoS, energy consumption and cost. The degree of user sensitivity to QoS parameters for a particular service should be taken into account in the analysis of utility model for network selection mechanism. Thus, it is necessary to consider user’s timely preference with respect to the QoS parameters in addition to the application demands to model the utility function adaptively. From the above points, it is clear that S-shaped sigmoid function is appropriate to satisfy all the conditions of network selection problem. The main contribution of our work is eliminating the null utility effect during the aggregation of elementary utility functions and maximizing the user satisfaction degree by selecting the best suitable network.

III. A Framework of the Proposed System The main issue in heterogeneous wireless network is to provide uninterrupted service to the users in order to satisfy the ABC concept. In additive utility function, interdependence between criteria was not considered in the network selection process. For example, high data rate requirements, frequent handovers and unnecessary interface selection can also consume considerable amount of energy of the multi interface MT. Thus, it is necessary to incorporate the energy consumption during network selection process. If the MT battery level is less than 15% of full battery, user preference is automatically changed to energy sensitive mode. In the proposed scheme, MT is connected with the serving network as long as possible to reduce the power consumption.


For example, when the MT battery level is very low, it is very difficult to perceive good QoS even though the network supports high data rate. Also, desirable increase in QoS may result in undesirable increase in the cost of service. Thus, the focus of this paper is to provide optimization between QoS, cost and energy consumption while selecting the optimal network. The proposed system consists of three modules: 1.Network Discovery and Information Collection module 2. Network Evaluation module 3.Network Selection module as shown in Fig. 1.

Network Discovery and Information Collection Module

Network Evaluation Module

Network Selection Module

Fig. 1. System Model

III.1. Network Discovery and Information Collection Module Whenever handover is initiated, multi interface MT discovers the available networks in the coverage area. It collects the information for the discovered networks based on the network conditions, mobility characteristics, QoS profile and also type of user. The decision metric involved in the network selection process decides the effectiveness of the proposed scheme. The parameters invoked in order to provide complete solution for seamless connectivity in heterogeneous wireless networks is shown in Table I. TABLE I PARAMETERS CONSIDERED Types of parameters Attributes considered Network Data rate, latency, link quality, cost and power consumption Terminal Velocity, battery power Service QoS level User User profile

III.2. Network Evaluation and Network Selection Modules Fig. 2 shows the operations involved in the network evaluation and selection modules. After collecting the information, available networks are evaluated based on weight adaption and multiplicative utility function calculation. III.2.1. Weight Adaption The characteristics of different traffic must be taken in to account in the design of network selection algorithm in order to provide service continuity with guaranteed QoS to the users in the integrated networks. Therefore an appropriate weight value should be assigned to each metric to account for its importance in providing QoS requirements to particular application.


1295


Eq. (1) reflects the condition that aggregate utility U(x) should increase when the elementary utility u(x) increases. Eq. (2) shows the null utility effect while aggregating the elementary utility functions. Eq. (3) says that if all criteria considered satisfy the user requirements and application profile, global utility function should be high. The proposed weighted utility functions consider the application requirements in terms of QoS, estimated energy consumption and cost of service offered by network:

Discover the available networks

Obtain the network parameters, user preferences, mobile terminal battery level

Adapt weight values based on user preferences

Calculate the multiplicative utility function based on sigmoidal function for serving and detected networks

Arrange the utility functions of the networks in descending order to rank them

Select the network with maximum utility function Fig. 2. Network Evaluation and selection process

Different users may have different sensitivity to different parameters such as QoS level, cost of service and energy consumption. It can be considered in the network selection process by including the weight values or by selecting an appropriate utility function to capture user preferences. User preferences are categorized based on user sensitivity to QoS , cost and MT battery level. If the user is more sensitive to the cost of the service, a higher weight is given for cost than other criterion. If the user requires more quality oriented services, quality is given a higher priority. For energy sensitive devices, higher importance will be given to energy consumption of the network. Thus, the objective is to provide optimization between quality, cost and energy in the automatic network selection mechanism. In the proposed scheme, QoS is determined in terms of data rate and delay requirements of the ongoing service. But, the weight values are assigned in the utility calculation to satisfy the minimum QoS requirements of service irrespective of the type of user. III.2.2. Multiplicative Weighted Utility Function To show how the interdependence between the criteria is reflected in the multi criteria utility function, the following conditions must be satisfied:

U  x  ui

lim U  x   0

0

 i  1,....,n

(1) (2)

ui  xi 0

lim U  x   1

(3)

u1 ,....un 1


0      x  xl   x x    m l    x  xl   1    xm  xl   u  x      xu  x     1   xu  xm     1   xu  x    xu  xm   1 

x  xl

xl  x  xm (4)

xm  x  xu

x  xu

where ζ determines the user sensitivity to the variation of network characteristics. It should be high for inelastic RT applications to show higher user sensitivity, but small for elastic NRT applications. The elementary utilities of QoS, cost and energy consumption are calculated using sigmoidal function with lower and upper limits is given by Eq. (4). For each application, there is a minimum and maximum requirement for each criterion. For example, the upper (xu) and lower values (xl) of data rate for voice application are 64kbps and 32kbps respectively. i. Quality utility Elementary utility value of QoS is calculated by combining the utility values of data rate and delay. To apply the sigmoidal function for utility calculation, the values of lower limit (xl), upper limit (xu) and threshold value (xm) should be known for each metric for a particular application. They are assigned according to the Table II [9]. Upward and downward criteria should be identified for each network during network evaluation, because users usually prefer high value for upward criteria and low value for downward criteria. This may result in the low value of downward utility which in turn affects the overall utility of each network. To compensate this effect, effective value of downward criteria is obtained as follows:

udelay  x   1  u1delay  x 

(5)

where u1delay(x) is the utility value obtained using sigmoidal function given by Eq. (4).


1296


Application Voice Video Conferencing Video Streaming Web Browsing

TABLE II QOS REQUIREMENT OF APPLICATIONS Data ratel Data rateu Data ratem Delayl (kbps) (kbps) (kbps) (ms) 32 64 50 75 512 5000 3000 75 128 1000 500 2000 128 1000 500 250

ii. Cost utility Elementary utility value for the cost of service is calculated for each network using sigmoidal function. The values of xl, xm and xu for the cost are calculated according to the data rate requirements of the application and is different for the three networks as shown in the Table III [11]. Since the cost utility is downward criteria, the effective value is calculated as:

ucos t  x   1  u1cos t  x 

Networks UMTS WiMAX WLAN

TABLE III COST CALCULATION costl costu (units/kbps) (units/kbps) 20 60 1 50 1 40

Delayu (ms) 150 150 4000 500

Delaym (ms) 100 100 3000 350

n

U  x 

  ui  xi  

wi

(8)

i 1

where:





w1 w2  u    delay  x   * udatarate  x   *  U  x    w3 w4  u   x * u x       energy cost     

(9)

(6)

costm (units/kbps) 35 25 20

iii. Energy utility The energy utility value should be high for the network with low energy consumption. The effective value should be expressed as:

uenergy  x   1  u1energy  x 

(7)

If E is the actual energy consumption of the network, then xl, xm and xu values of energy consumption can be determined as follows: xl=E/32, xm=E/50 and xu=E/64, where 32, 50 and 64kbps are the xl, xm and xu values of data rate for the voice application. u1energy(x) is the utility value obtained using sigmoidal function. Table IV shows the energy consumption calculation for the three networks. TABLE IV ENERGY CONSUMPTION CALCULATION energyl energyu energym Networks (J/kbps) (J/kbps) (J/kbps) UMTS 0.018 0.037 0.024 WiMAX 0.054 0.700 0.109 WLAN 0.070 0.140 0.090 TABLE V WEIGHT VALUES FOR THREE TYPES OF USERS User preference w1 w2 w3 w4 QoS 0.4 0.2 0.2 0.2 Cost 0.3 0.1 0.1 0.5 Energy 0.3 0.1 0.5 0.1

where w1, w2, w3 and w4 are the weight values assigned according to the user preferences and QoS profile. Table V shows the weight adaption for voice application according to the type of users. After the evaluation of networks, the available networks are ranked in descending order based on the value of global utility function. The network with the maximum utility function is selected as the target network.

IV.

Performance Evaluation

In this section, the performance of the proposed scheme is analyzed numerically using MATLAB 7.5. Numerical simulation is conducted to demonstrate the efficiency of multiplicative aggregate utility function for automatic network selection in heterogeneous networks. IV.1. Comparison Between Additive and Multiplicative Aggregate Utility Function A general performance analysis of additive and multiplicative aggregate utility functions is performed from which it is known that how the multiplicative utility function is more efficient than additive utility function. The simulation parameters are described in Table VI. Fig. 3 and Fig. 4 show the utility values for additive and multiplicative utility functions as a function of QoS and Cost. The utility values for some particular values of QoS and Cost are shown in Table VII in order to highlight the efficiency of multiplicative utility function. TABLE VI SIMULATION PARAMETERS Attributes

iv. Global utility function The overall utility function is calculated for each network using multiplicative utility function given by:


Delay (d) Data rate (dr) Cost (c)

Preferences

xl

xu

xm

w1=0.3 w2=0.3 w3=0.4

5 10 0

70 90 80

30 40 40


1297


Cost (c) (units/kbps) 1 40 80 100

TABLE VII SIMULATION RESULTS Utility values QoS level Multiplicative (Q) Additive Utility Utility 1 0.7002 0 30 0.3767 0.4445 70 0.2681 0.2632 100 0.2770 0.1917

The network with (c=40, QoS=30) satisfies both the cost and QoS requirements of user and hence it is selected as the target network. Similar is the case with networks (c=80, QoS=70) and (c=100, QoS=100). The interdependence between the different criteria is considered in the case of multiplicative utility function. It is inferred that if any one of the elementary utility is zero then, the overall utility becomes zero. Fig. 3. Variation of Additive utility with cost and QoS

IV.2. Network Selection for Various Applications

Fig. 3 and Table VII show that a network with (c=1, QoS=1) has a higher utility than network with (c=40, QoS=30) and can thus lead to a wrong network selection. Even though the first network offers low cost, it does not fulfill the QoS requirements of user. But the second network satisfies both the QoS as well as cost requirements, but it is not selected using additive utility function and hence leads to wrong selection of network. Similar is the case for the networks with (c=80, QoS=70) and (c=100, QoS=100). The network with (c=80, QoS=70) suits the user requirements for both QoS and cost than the other network (c=100, QoS=100) which satisfies only QoS requirements. But the former is having a smaller utility value than the latter and hence qualified network is not selected. Fig. 4 shows the variation of multiplicative aggregate utility as a function cost and QoS. It is seen that the network with (c=1, QoS= 1) could not satisfy the user’s QoS requirements and hence it is not selected.

In the numerical simulation, a heterogeneous environment with three networks UMTS, WiMAX and WLAN is considered. The application QoS and network parameters and their values for the three networks considered [11] are listed in Table II and Table VIII respectively. TABLE VIII NETWORK PARAMETERS Attributes UMTS WiMAX Cost (units/kbps) 20-60 1-50 Energy consumption in 1.2 3.5 active state (J) Data rate (Mbps) 1 10 Delay (ms) 10 50

WLAN 1-40 4.5 5 130

Fig. 5. Network selection for voice application

The main objective of this simulation is to verify how accurately the multiplicative utility function selects the best candidate access network according to QoS profile and user preference. Voice application is considered for the analysis of multiplicative utility function for QoS sensitive users. Since delay is the most important attribute of the application considered, a higher preference weight is assigned for it. Fig. 5 shows the result of simulation. It is inferred that the utility value is also high for UMTS and WiMAX networks. UMTS is better preferred since it has less delay and high coverage than WiMAX.

Fig. 4. Variation of Multiplicative utility with cost and QoS



1298


The minimum delay requirement for voice application is 75ms as given in Table II. Hence the networks which satisfy this requirement are UMTS and WiMAX networks. It is inferred that the utility value is also high for UMTS. To highlight the proficiency of the proposed utility function, the three networks are simulated for various RT and NRT applications and the simulation results for QoS and energy sensitive users are shown in Fig. 6 and Fig. 7 respectively. RT applications include Voice and Video Conferencing and NRT applications are Video Streaming and Web browsing. For QoS sensitive users (Fig. 6), it is observed that UMTS has a higher utility value for voice application considering the delay requirements. Likewise for the video conferencing application, both UMTS and WiMAX satisfy the delay requirements of the application. But, when we consider the data rate constraints, WiMAX is more desirable than UMTS. As the interdependency between the different criteria is considered in computing multiplicative aggregate utility function, it selects WiMAX as the access network. For both video streaming and web browsing applications, all the three networks satisfy both delay and data rate specifications. Therefore, cost plays an important role in differentiating the available networks. Thus, WLAN is selected for NRT applications. For energy sensitive users (Fig. 7), it is observed that the MT selects the UMTS interface for low voice application as it satisfies the delay requirements and also it consumes less power compared to other networks.

Thus the utility value of UMTS is higher than WLAN and WiMAX for voice application. For high data rate applications such as video conferencing, UMTS network cannot afford the data rate constraints where as WLAN and WiMAX can satisfy this requirement. WiMAX has a higher utility value than WLAN since it satisfies both data rate and energy consumption constraints. For NRT applications, WiMAX is better preferred as it can better support the QoS and energy consumption requirements for energy sensitive user. IV.3. User Satisfaction Degree A concept of Cobb-Douglas [11] based user satisfaction degree is introduced to analyze whether a user is convinced with the network selected using the multiplicative utility function. User satisfaction degree implies how much the user is satisfied with the service obtained from a particular network and is measured in terms of Utility (U) and cost (c). It is expressed as:



A U ,c   1  exp  KU  u  c 





(10)

where µ>0 and ϵ>0 control the sensitivity to utility and cost respectively and K is a positive constant representing the satisfaction reference value. Here, utility corresponds to a Global utility which is computed using the multiplicative utility function considering the attributes: energy consumption, data rate and delay. A separate utility (u(c)) is calculated for cost using the single criterion utility function.

Fig. 6. Network selected for QoS sensitive user

Fig. 8. User Satisfaction Degree for the selected network

Fig. 8 shows the user acceptance degree for various applications when a particular network is selected for a particular application. The values of μ, ϵ and K used in the simulation are 2, 0.2 and 0.5 respectively. It is perceived that user acceptance for UMTS is high for voice, video streaming and web browsing applications. Fig. 7. Network selected for Energy sensitive user



1299


User satisfaction for the video conferencing application is poor and hence it is not selected. Since WLAN is suitable only for NRT applications, its user acceptance is high whereas it is very low for RT applications. As WiMAX is suitable for all the four applications, its user satisfaction degree is high. Thus this concept of user satisfaction can be considered as a feedback from user after the network selection to ensure whether the user is satisfied with the sevice offered by the network.

V.

References [2]

[3]

[4]

[5]

[7]

[8]

[9]

Conclusion

This paper proposes a multiplicative weighted utility function to decide the optimal network based on network conditions, user preference, cost, QoS and energy consumption. Hence, the users will always be connected to the best suitable access network to enjoy seamless connectivity and ubiquitous service access in the heterogeneous environment. Multiplicative utility function is compared with additive utility function to show that the null utility effect while aggregating the elementary utilities has been eliminated using multiplicative utility function during the network evaluation process. The asymmetry between cost of service and QoS obtained from the network is addressed in the calculation of utility values. The proposed algorithm provides the optimization between QoS, cost and energy consumption while selecting the optimal network. The effectiveness of the proposed utility function is verified using Cobb Douglas based user satisfaction degree for the service offered by the selected network. But, the proposed approach does not address the impact of QoS and cost of service on the dynamics of network selection and the network level performance. For future research, the proposed algorithm will be integrated with residual residence time estimation particularly in an overlapping region of available networks to analyze the dynamics of network selection for high speed users using number of handover failures and unnecessary handovers.

[1]

[6]

E. Gustaffson, A. Jonsson, Always Best Connected, IEEE Wireless Communication, Vol.10, no. 1, pp. 49-55, 2003. Ding Zhang, Yu Zhang, Na Lv, Yibo He, An Access Selection based on GRA Integrated with FAHP and Entropy Weight in Hybrid Wireless Environment, in Proc. of 7th International Conference on Application of Information and Communication Technologies,” Baku, pp. 1-5, 2013. Chien- Hua Wang, Sheng - Hsing Liu, Chin- Tzong Pang, Using Genetic Algorithm Improve the Consistency of Fuzzy Analytic Hierarchy Process, IEEE Consumer Communications and Networking Conference, Las Vegas, USA, pp. 485-489, 2013. S.Y. Lee, K. Sriram, K. Kim, Vertical Handoff Algorithm for Providing Optimized Performance in Heterogeneous Wireless Networks, IEEE Trans. on Vehicular Technology, Vol. 58, no. 2, pp. 865-881, 2009. A. Singhrova, N. Prakash, Vertical Handoff Decision Algorithm for Improved Quality of Service in Heterogeneous Wireless Networks, IET Communications, Vol. 6, no. 2, pp. 211-223, 2012.


[10]

[11]

[12]

[13]

[14]

[15]

F. Kaleem, A. Mehbodniya, K. K. Yen, F. Adachi, A Fuzzy Preprocessing Module for Optimizing the Access Network Selection in Wireless Networks, Advances in Fuzzy Systems, ISSN: 16877101, 1687711X, pp. 1-9, 2013. C. Amali, Bibin Mathew, B. Ramachandran, Intelligent Network Selection using Fuzzy Logic for 4G Wireless Networks, International Journal of Electronics and Communication Engineering and Technology, Vol. 2, no. 2, pp. 451-461, 2013. R. Trestian, O. Ormond, Gabriel-Miro Muntean, “Energy– Quality–Cost Tradeoff in a Multimedia-Based Heterogeneous Wireless Network Environment, IEEE Trans. On Broadcasting, Vol. 59, no. 2, 2013. I. Chamodrakas, D. Martakos, A Utility-based Fuzzy TOPSIS method for Energy Efficient Network Selection in Heterogeneous Wireless Networks, Applied Soft Computing, Vol. 12, no. 7, pp. 1929-1938, 2011. C. Amali, B. Ramachandran, Modified Weight Function Based Network Selection Algorithm for 4G Wireless Networks, in Proc. ACM International Conference on Advances in Computing, Communication and Informatics, Chennai, India, pp. 292-299, 2012. Quoc-Thinh Nguyen-Vuong, Yacine Ghamri-Doudane, Nazim Agoulmine, On Utility Models for Access Network Selection in Wireless Heterogeneous Networks, in Proc. IEEE Networks Operations and Management Symposium, Salvador, Bahia, pp. 144-151, 2008. L. Wang, G. S. Kuo, Mathematical Modeling for Network Selection in Heterogeneous Wireless Networks – A Tutorial, IEEE Communications Surveys & Tutorials, Vol. 15, no. 1, 2013. Sun Lei, Wang Jianquan, Lv Zhaobiao, An Adaptive Network Selection Scheme in 4G Composite Radio Environments, in Proc. 8th International Conference on Wireless Communications, Networking and Mobile Computing (WiCOM), Shanghai, China, pp. 1-4, 2012. L. Giupponi, R. Agusti, J. Perez Romero, O. Sallent, Towards Balancing User Satisfaction and Operator Revenue in Beyond 3G Cognitive networks, in Proc. 15th IST Mobile Wireless Summit, Myconos, Greece, pp. 1–5, 2006. Tu, J., Li, D., Zhang, Y., Tong, W., Performance analysis of vertical handoff in WIFI and WIMAX heterogeneous networks, (2012) International Review on Computers and Software (IRECOS), 7 (4), pp. 1866-1871... Nirmal Raj, T., Suresh, R.M., QoS aware vertical handoff decision for UMTS-WiMAX networks, (2013) International Review on Computers and Software (IRECOS), 8 (12), pp. 2812282.

Authors’ information Amali Chinnappan received Bachelor’s degree in Electronics and Communication Engineering from the Govt. College of Technology, Coimbatore, India and M.E. degree in Applied Electronics from College of Engineering, Guindy, India, in 2007. Her research interest includes vertical handover in heterogeneous wireless networks, resource management and mobility management in wireless networks. She is currently working towards Ph.D., degree with SRM University, Chennai, Tamil Nadu, India. She has 12 years of teaching experience in Engineering Colleges. She is working as an Assistant Professor in Valliammai Engineering College. She is a member of ISTE and ISC. Dhanasree Jayaprakash received Bachelor’s degree in Electronics and Communication Engineering from College of Engineering Trikaripur, Kerala, India in 2007 and M.Tech in Communication Systems from SRM University, Chennai, India in 2014. Her current research interest includes wireless networks, vertical handovers and network selection strategies.


1300


Ramachandran Balasubramanian received Bachelor’s degree in Electronics and Communication Engineering from Thiagarajar College of Engineering, Madurai in 1990 and Master’s degree in Satellite Communications from National Institute of Technology, Trichy in 1992. He obtained his Ph.D., degree in the area of Wireless Mobile Networks from Anna University Chennai in 2009. He authored a textbook on Digital Signal Processing. His teaching and research interests include Digital Communication, Wireless Networks, UWB Antenna Design, Network Security, and Mobile Computing. He has published 37 research papers in national and international conferences and journals. He was awarded IETE-SK Mitra Memorial Best Research Paper Award in 2009.Dr. B. Ramachandran is a Professor in SRM University, Chennai. He is a member of ISTE, and fellow of IE(I), and IETE.



1301


Rainfall Intensity Classification Method Based on Textural and Spectral Parameters from MSG-SEVIRI Y. Mohia1, S. Ameur1, M. Lazri1, J. M. Brucker2 Abstract – In this paper we propose the rainfall intensity classification algorithm .The study is carried out over north of Algeria. The developed rain intensities classification technique (RICT) is based on the artificial neural multilayer perceptron network (MLP). The (MLP) model is created with three layers (input, hidden, and output) that consist of 15 input neurons, which as ten spectral features that were calculated from MSG (Meteosat Second Generation satellite) thermal infrared brilliance temperature (BT) and brilliance temperature difference (BTD) and as five textural features, and 6 output neurons in the output layer that represent the 6 rain intensities classes: very high, moderate to high, moderate, light to moderate, light and no rain. The precipitation process areas and the rainfall intensity subareas classified by the proposed technique are validated against groundbased radar data. The rainfall rates used in the training set are derived from Setif radar measurements (Algeria). The results obtained after applying this method for the north of Algeria zone show the neural network performs very well and indicate an encouraging performance of the new algorithm concerning rain area classification using MSG SEVIRI. We found that the introduction of textural parameters as additional information works in improving the classification of different rainfall intensities pixels in the MSG–SEVIRI imagery compared to the techniques based only on spectral information. Copyright © 2014 Praise Worthy Prize S.r.l. All rights reserved.

Keywords: Artificial Neural Network, Convective and Stratiform Cloud, Image Classification, Radar Data, Rainfall Intensities, Satellite Image, Spectral and Textural Features

I.

Introduction

Image Classification can be defined as the process of partitioning an image into different regions, each region being homogenous as regards the three fundamental pattern elements to wit spectral, textural and contextual features. More information can be extracted from the satellite image and can be used in different climatologically applications such as the detection of clouds [1], [2], the classification of clouds [3], [4], [5], [6] and precipitation estimation [7], [8], [9], [10], [11]. The precipitations are considered as an important factor in the water cycle, whose role is very important in human life. Many satellite rainfall retrieval methods have been developed based on geosynchronous satellite individual or combined visible and infrared channels. Geostationary weather MSG satellite system with his high spectral resolution of the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board [12], [13] with spatial and temporal resolution, offer the potential for area-wide rainfall retrievals. It also have the possibility to offers a mean of extracting the microphysical and dynamic structure of precipitating clouds allowing for an enhanced discrimination between convective and stratiform rain areas and thus contributing to the improvement of the satellite rainfall estimation [8].

Most of the rainfall retrieval techniques based on MSG data use a relationship between cloud top temperature (CTT) measured in an infrared (IR) channel and the rainfall probability and intensity [14], [7]. Using the results of Nauss and Kokhanovsky [15], [16], a new technique for the detection of precipitating cloud areas in the mid-latitudes applicable to the (MSG) satellite is introduced by Thies et al. [17], [18]. This technique relies on information about the cloud water path (CWP) and the cloud phase (CP) in the upper cloud parts. Precipitation regions can be classified into convective and stratiform areas [19]. The convective and stratiform rain components are the result of distinct cloud dynamical processes that influence the atmospheric circulation [20]. Convective areas are associated with high rainfall intensities because the highly convective clouds that can be easily identified in the infrared and/or water vapour channels [21], [22], are characterized by a large vertical extension with a cold cloud top rising high into the troposphere. Stratiform areas are associated with low rainfall intensities which covers a large rainfall region [23]. The accuracy of satellite rainfall estimation obtained with many schemes using different techniques or parameterizations for convective and stratiform clouds [24], [4], [25], [7], [26] can be improved by rainfall classification methods [27].


1302

Y. Mohia, S. Ameur, M. Lazri, J. M. Brucker

Many methods have been developed in order to assigning rain rates to cloud pixels [28], [29], [26], [7], [30], [23]. The accuracy of these schemes strongly depends on capacity of this classification method to adequately distinguish between convective and stratiform rain clouds. The majority of these methods use the spectral features for distinguishing cloud types using the information on the cloud radiance in different spectral bands [31]. The spectral features due to their physical importance have proved effective and simple in cloud classification [31]. Classification techniques based on this features in a single infrared channel are most applicable for deep cold convective clouds and thus work best in the Tropics [21], [22] and adding other channels is essential in order to apply these techniques in mid-latitudes. A new technique for rainfall process separation as a part of a satellite-based rainfall retrieval scheme in the mid-latitudes using multispectral satellite data is developed by Thies et al. [32]. Another scheme that classifies convective and stratiform precipitation areas based on the high infrared spectral resolution of the MSG–SEVIRI is devlopped by Feidas and Giannakos [33]. The spectral similarities of certain cloud features are major drawbacks of these techniques. In order to overcome the difficulty of ameliorate these techniques; the contribution of other features is needed. The most important parameter to be used is textural features that refer to the spatial distribution of gray levels in a pixel array of an image. The adding of this information helps to improve as well rainfall retrieval performance [34], [10], [3] and contributes to ameliorate the rainfall classification algorithms [35], [36], [31]. The textural features of cloud are often distinct and usually less sensitive to the effects of atmospheric attenuation or detector noise [37]. The objective of our study is therefore to investigate the potential of using spectral and textural features in the MSG–SEVIRI data for developing a technique capable of classifying rainfall intensities into six classes over North of Algeria. Motivated by the need to use multispectral information to better rainfall intensities classification, we used five textural features combined with ten spectral parameters from MSG satellite. For this purpose a statistical approach (artificial neural network, ANN) based on the correlation of spectral and textural parameters with different rainfall intensities is used. These two types of parameters (spectral and textural) represent the fifteen input MLP algorithm and the precipitation data from radar of Setif (Algeria) are used as the six output of this model in terms of a rainfall intensity classification. Two stages are included in the MLP model, the first is the training stage, the second consists of the application stage. The calibration and validation are carried out by comparing spectral and textural parameters with six classes’ rainfall intensities radar observations. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

After training process, the fifteen parameters of the pixels in the test dataset are put into the network to calculate the rain flag of each pixel. Finally, five statistical standard verification scores are applied in order to appraisal the results of classification given by the proposed technique.

II.

Study Area

The study area selected is covering the northern of Algeria. The bounding coordinates are from 34° north to 37° north latitude and from 2° west to 9° east longitude as shown in Fig. 1. Algeria is located on the South shore of the Mediterranean region; it is bordered on the East by Tunisia and Libya, on the South by Niger and Mali, South-West by Mauritania and Western Sahara and West by Morocco. The rainy season extends from October to March, with maximum rainfall occurring during November-December. In the north, the climate is Mediterranean transit, marked by seasonal oscillations. The average annual rainfall is estimated at about 600 mm. The minimum rainfall is recorded in the southern regions. It is about 50 mm while the maximum is observed in the Djurdjura massif located in Kabylia and the massif of Edough located a little farther east, where it exceeds 1500 mm.

Fig. 1. A study area and the position of the weather radar of Setif. The red circle shows the radar domain with a radius of 250 km verage response time per number of sites

III. Data Sets For this study, two types of data are required: MSG SEVIRI data together with corresponding ground-based radar. The SEVIRI data has been received at the Marburg Satellite Station (MSS) [38], [39]. The radar data are provided by the ground-based C band radar network of Setif (Algeria) of the National Office of Meteorology of Algeria. The scan interval for both data sets is 15 minutes. III.1. MSG Data The EUMETSAT Meteosat Second Generation (MSG) geostationary satellite with a Spinning Enhanced Visible and InfraRed Imager (SEVIRI) on board is positioned at 0° longitude and 0° latitude, approximately 36 thousand km above the Gulf of Guinea and was


1303


launched on the 29th of August 2002 but data has been available free to the academic and scientific communities since January 2004. The MSG SEVIRI operates with 12 spectral channels that provide measurements at the sub satellite point every 15 minutes whose 11 channels measurements have a resolution of 3x3 km2 and one canal (a High Resolution Visible (HRV) channel) measurements have a resolution of 1×1 km2 [40],[41]. Data is then processed and wavelet compressed, then uplinked via the EUMETCast service – a new C-band satellite reception facility to collect data from SEVIRI – to the commercial telecommunication geostationary satellites from which it can be disseminated to these communities. III.2. Radar Data The radar of Setif is one of the seven radars of the Algerian Meteorological Network. It is associated with SANAGA (Digital Acquisition System for the Analysis of African Grains), system of acquisition and digitization of images. The radar is installed near to the town of Setif (Algeria) and located at latitude 36 ° 11 'N, longitude 5 ° 25' E with an altitude of 1033 m. The displacement in azimuth is between 0 to 360 degrees in continuous and the movement in inclination is between -20 ° to 90 °. Its polarization is linear and horizontal. The technical characteristics of used radar are: Wavelength (5.5cm), Peak power (250kw), Repetition frequency (250Hz) and Pulse duration (4μs). It records an image of size 512x512 pixels every fifteen minutes. Each pixel coded on four bits, it has a resolution of one km 2. The representative physical parameter of the radar reflectivity factor is noted:





Z  mm6 m 3 . The relationship used to convert the reflectivity factor Z into precipitation intensity R (mm / h) is the following equation:

Z  300.R1.5

(1)

For the spatial comparison the radar data with an original spatial resolution of 4 by 4 km were projected to the viewing geometry of SEVIRI with a spatial resolution of 3 by 3 km. The radar data are used to validate the proposed scheme and are divided into training and validation data set. The training data set is used for the development of the technique and consists of 2109 scenes of precipitation events from November 2006 to March 2007. The validation data set is considered for the appraisal of the proposed technique and consists of 1936 scenes of precipitation events from November 2009 to March 2010.

IV.

Methodology

IV.1. Spectral Parameter The spectral parameters used in our study are derived from MSG satellite described in precedent section. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

For this, ten parameters are retained: one brilliance temperature infrared channel (TIR10.8), six brilliances temperature difference between two infrared channels (ΔTIR10.8-IR12.1, ΔTIR8.7-IR10.8, ΔTIR3.9-IR10.8, ΔTIR13.4-IR10.8, ΔTIR8.7-IR12.1, ΔTIR9.7-IR13.4),one brilliance temperature difference between infrared channel and water vapour channel (ΔTIR3.9-WV7.3), one brilliance temperature difference between water vapour channel and infrared channel (ΔTWV6.2-IR10.8) and one brilliance temperature difference between two water vapour channels (ΔTWV6.2WV7.3). 1) BT at the 10.8-μm channel (TIR10.8): It allows the detection of medium and high-level clouds with a 10.8µm brightness temperature lower than the clear sky surface brightness temperature. It also allows an indication of the vertical extent of the cloud because, in general, brightness temperature of the system depends on the cloud-top height [42], [43], [6], [9], [32]. 2) BTD between the 10.8-μm channel and the 12.1-μm channel (ΔTIR10.8-IR12.1): it helps to discern cirrus from snow and ice and is positive at low optical thicknesses. It allows the detection of thin cirrus clouds and cloud edges characterized by higher 10.8µm –12.0µm values than cloud-free surfaces. It can be used over all surfaces and for any solar illumination. This brightness temperature difference is a good indicator of the cloud optical thickness is very effective in discriminating optically thick cumuliform clouds from optically thin cirrus clouds [42], [44], [45]. Optically thick cumulus type cloud shows the smaller BTD due to their black-body characteristics, while optically thin cirrus cloud shows the larger BTD due to the differential absorption characteristics of ice crystals between the two channels [46]. It is expected that optically thick and deep convective clouds are associated with rain [45]. Even though the split window technique is very effective in detecting and removing optically thin cirrus clouds with no precipitation, it sometimes incorrectly assigns optically thick clouds like cumulonimbus in place of optically thin clouds [46]. 3) BTD between the 8.7-μm channel and the 10.8-μm channel (ΔTIR8.7-IR10.8): It allows detecting thin cirrus clouds over all surfaces for any solar illumination. It is based on the fact that high semi-transparent clouds are characterized by relatively high 8.7µm–10.8µm difference compared with surface values. The information about the cloud phase can be obtained using this BTD [32].The imaginary (absorption) component of the index of refraction, which is a direct indicator of absorption/emission strength, differs for ice and water at these two wavelengths. More specifically, the difference in water particle absorption is small between the two wavelengths, but very large for ice particles. Radiative transfer simulations show that for ice clouds, Τ8.7-Τ10.8 tends to be positive in sign, whereas for low-level water clouds, Τ8.7-Τ10.8 tends to be small negative.


1304


4) BTD between the 3.9-μm channel and the 10.8-μm channel (ΔTIR3.9-IR10.8): The 3.9µm channel is used when the 1.6µm observations are unavailable. During the day, it includes solar reflection and thermal emission and the solar reflection part is roughly approximated by the 3.9µm -10.8µm brightness temperature difference. It allows the detection of high semi-transparent clouds or sub-pixel cold clouds. It is based on the fact that the contribution of the relatively warm grounds to the brightness temperature is higher at 3.9µm than at 10.8µm, due to a lower ice cloud transmittance, and to the high nonlinearity of the Planck function at 3.9µm. The brightness temperature difference 3.9µm -10.8µm is a function of the cloud height, thickness (for cirrus) and cloudiness (for subpixel clouds). 5) BTD between the 13.4-μm channel and the 10.8-μm channel (ΔTIR13.4-IR10.8): it allows to detect cumulus cloud growth development and to estimate the cloud top height, in MSG images [47]. The large negative BTD (T13.4–T10.8) values are corresponding to the low-level clouds because the absorption from CO2 above them reduces the temperature at 13.4.The small negative values of this BTD are corresponding to the growing cumulus clouds since their cloud tops are above most of the CO2 layer and ice crystal absorption is similar for both wavelengths [48]. 6) BTD between the 8.7-μm channel and the 12.1-μm channel (ΔTIR8.7-IR12.1): it can be used to give information about clouds’ optical thickness [2]. The positive values of this BTD correspond to the high cloud optical thickness due to the scattering processes and the dependence on particle size which is stronger in the 8.7-μm channel compared to the 12.1-μm channel. For large cloud particle size and large optical thickness this values are high positives.In contrast, the negative values of this BTD correspond for low-cloud optical thickness because water vapour absorption in the 8.7-μm channel is higher than in 12.1-μm channel [49]. 7) BTD between the 9.7-μm channel and the 13.4-μm channel (ΔTIR9.7-IR13.4): it gives information of cloud top height. The 9.7 μm channel is dominated by O3 absorption and the 13.4 μm channel is dominated CO2 absorption. For low-level clouds the BTD (ΔT9.7–13.4) produces negative difference values because 9.7-μm channel temperatures are colder than temperatures at the 13.4-μm channel due to the absorption from O3 above them[50].For high clouds over 12 km the brightness temperature at 9.7 μm show significantly larger values than brightness temperature at 13.4 μm due to the warming by stratospheric ozone [50]. 8) BTD between the 6.4-μm channel and the 7.3-μm channel (ΔTWV6.2-WV7.3): This two water vapour channels are dominated by atmospheric water vapor absorption. The BTD (ΔTWV6.2-WV7.3) allows detecting the cloud height and giving information about the early detection of convective activity. The large


negative temperature differences are corresponding to low clouds. The very high clouds that can reach the stratosphere tend to give positive temperature differences. 9) BTD between the 6.2-μm channel and the 10.8-μm channel (ΔTWV6.2-IR10.8): this parameter distinguishes between high-level and low-level/mid-level clouds [51]. The temperatures of low-level clouds at the 6.2μm channel are lower than their actual cloud top temperatures. But at the 12.0-μm window channel their cloud-top temperatures are representative of actual cloud-top temperature since the atmosphere is transparent to this wavelength. As a result, for lowlevel clouds, the BTD (∆TWV6.2-IR1.8) tends to be very negative in sign. For the upper level thick clouds , temperatures at the 6.2-μm channel close to their actual cloud-top temperatures and BTD (∆TWV6.2IR10.8) gives very small negative values. When water vapor is present in the stratosphere above the cloud top positive differences may occur, which is a sign of convective cloud tops [1],[52] as opposed to mere cirrus clouds. 10) BTD between the 3.9-μm channel and the 7.3-μm channel (ΔTIR3.9-WV7.3): it should have generally similar characteristics as BTD (ΔT3.9-10.8). But, BTD (ΔT3.9-7.3) should be higher than (ΔT3.9-10.8) due of the diminishing effect of the water vapour absorption and emission in mid- to low tropospheric levels on the brightness temperature (BT) in the 7.3 μm channel [13]. Therefore, The BTD (ΔT3.9-7.3) can provide additional information about the CWP. The lower difference values correspond for optically thin clouds. In contrast, the large particles together with a high optical thickness give the medium to the high difference values. IV.2. Textural Parameters We can find several types of texture descriptors that are used to obtain information about the texture of the image. The most effective, commonly used are: First Order Statistics (FOS) [53], [54], Grey Level Cooccurrence Matrix(GLCM), [55], Gray Level Difference Vector (GLDV) ,local histograms parameters [56], local extrema parameters [57], curvilinear integral parameters [57], Neighborhood Gray Tone Difference Matrix (NGTDM)[58], [59] Statistical Feature Matrix (SFM) [60], Laws Texture Energy Measures (TEM) [61], [62], Fractal Dimension Texture Analysis (FDTA) [63], [64] and Fourier Power Spectrum (FPS)[65]. In our study, we chose the Gray level co-occurrence matrix (GLCM) that has been proven to be a very powerful tool for texture image segmentation [40], [41]. This matrix is used to extract second-order statistical texture features. Fourteen parameters describing textures, as defined by Harralick et al. [57] can be extracted from the co-occurrence matrix. Among these, we used the most popular features based on recommendations found in the literature [66] and the nine most frequently used in


1305


remote sensing [67],[68]: mean, variance, correlation, contrast, angular second moment, entropy, homogeneity, directivity and uniformity (also called energy). Then, we select the less correlated textural parameters on each other in order to reduce the number of the textural measures to be used in the rainfall intensity classification technique. In order to determine the least correlated parameters, the coefficient of correlation between parameters is used. The initial parameters are: 1) Angular second moment (ASM): Increases with regularity of the texture. ASM is high when image has very good homogeneity or when pixels are very similar. It is the sum of squares of entries in the GLCM: Ng 1 Ng 1

  x2ij

ASM 

i 0

(2)

7) Correlation (Cor): measures of grey tone linear dependencies of grey levels of neighboring pixels in the image: Ng 1 Ng 1

Cor 

 i 0

i  x   j   y   x y

j 0

8) Variance: is a measure of the dispersion of the values around the mean. It is similar to entropy. Variance calculated using i or j gives the same result, since the GLCM is symmetrical: Ng 1

 x2 

2

 i  x  M x i 

ASM

(3)

3) Entropy (Ent): Entropy is the opposite of energy; it gives the disorderliness of an image and Increases with irregularity of the texture: Ng 1 Ng 1

  xij log xij i 0

(4)

j 0

4) Mean (Mea): gives the average grey level with respect to the central position: Ng 1 Ng 1

Mea 

  i  j 

2

xij

(6)

j 0

6) Homogeneity (Hom): also known as "Inverse Difference Moment", assesses image homogeneousness and for smaller difference between grey values it takes on larger values. Related to contrast of the texture and it is high when local gray level is uniform: Ng 1 Ng 1

 i 0

j 0

(10)

j 0

9) Standard Deviation (Sd): reflects the degree of distribution of the grey level values and the copiousness of the data in the image. There is no particular advantage to using Standard Deviation over Variance, other than a different range of values:

Sd x   x , Sd y   y

(11)

10) Directivity (Dir): It allows summing the diagonal elements of the co-occurrence matrix. It gives a value that is more important when the texture has a privileged position in the direction of the translation direction: Ng 1

Dir 

  i,i,R 

(12)

For a more detailed description of the textural features, readers may refer to the study of Chen et al. [5].

Ng 1 Ng 1

Hom 

2

 i   y  M y  j 

(5)

j 0

5) Contrast (Cont): Also known as “Sum of squares variance”, measures the variance in grayscale levels across the image. It is related to contrast of the texture:

i 0

 y2 

i 0

  ixi, j i 0

Cont 

(9)

i 0 Ng 1

Ent 

(8)

j 0

where i, j are the spatial coordinates of the function p (i, j), Ng is gray tone. The square root of the ASM is sometimes used as a texture measure, and is called Energy. 2) Energy (Ene): also called ‘Uniformity’ is the square root of the ASM:

Ene 

xij

xij

2 1  i  j    

(7)


IV.3. Parameters Selection As we noted above, the parameters used in the classification process must be discriminating and their number must be limited if possible. The selection of parameters is a very difficult procedure for classification. It is very important to select the most relevant and therefore representative in order to achieve good classification parameters. There are several methods and tools for parameter selection [63]. Those are the simplest, studying redundancy or correlation of texture attributes. Thus we chose the method of selection attributes based on correlation coefficients and used by Rosenberger [69]. This method aims to evaluate the redundancy of information between attributes to keep only the least redundant.


1306


The linear correlation coefficient reflects the redundancy between attributes. We calculate the correlation cor (x, y) between two attributes x and y using the following equation:

  ai  ai   a j  a j 





cor ai ;a j 

i, j 2

  ai  ai    a j  a j  i

2

(13)

j

We get a symmetric correlation matrix that has elements in the interval [-1, 1]. This matrix is used to select the parameters that have a qualifying mutual redundancy. The purpose of the study is to identify the attributes providing no additional information for classification. To do this, we will consider only the absolute value of the correlation coefficient to quantify the redundancy between attributes. Indeed, we get the same result when the classification is done using two attributes (a1, a2) and (a1,-a2). Attribute (a1) is redundant relative to the attribute (a2) if the absolute value of the correlation between these two attributes is close to 1. Therefore, two additional attributes are relevant or if the absolute value of their correlation factor is negligible. The threshold value S, which corresponds to the maximum tolerated redundancy for two attributes, here is fixed at S = 0.95 as proposed by Rosenberger. Then, of the two attributes (ai) and (aj) are highly correlated, we must remove the most correlated to the other according to the following criteria: If

 cor  ak ;a j    cor  ak ;a j  k

,then

DBZ. The Rainfall rate estimated is greater than 12 mm/h. 2) The moderate to high precipitation intensities: this class corresponds to the precipitation zone which the corresponding radar reflectivity is between 42.0 DBZ and 46 DBZ. The rainfall rate estimated varies from 7.5 mm/h to 12mm/h. 3) The moderate precipitation intensities: this class corresponds to the precipitation zone which the corresponding radar reflectivity is between 34.0 DBZ and 38 DBZ. The rainfall rate estimated varies from 2.8mm/h to 4.6mm/h. 4) The light to moderate precipitation intensities: this class corresponds to the precipitation zone which the corresponding radar reflectivity is between 26.0 DBZ and 30 DBZ. The rainfall rate estimated varies from 1.0mm/h to 1.7mm/h. 5) The light precipitation intensities: this class corresponds to the precipitation zone which the corresponding radar reflectivity is between 12.0 DBZ and 22 DBZ. The rainfall rate estimated varies from 0.2mm/h to 0.6mm/h. 6) The no precipitant class: this class corresponds to the precipitation zone which the corresponding radar reflectivity is lower than 12.0 DBZ. The Rainfall rate estimated is least than 12 mm/h. Table I gives an overview of the five subareas of differing precipitation processes and intensities with the corresponding radar reflectivity and the characteristic rainfall intensity. TABLE I THE SIX CLASSES OF RAINFALL INTENSITIES ACCORDING TO RAINFALL RATE AND RADAR REFLECTIVITY. ODE SPECIFICOFFLINE PARAMETERS

the

k

attribute (ai) is removed, otherwise (aj) is discarded. Therefore, there will only be an eligible attributes with redundancy set by the threshold value. After applying this criterion, the remaining attributes for classification are: Contrast, Correlation, Entropy, homogeneity and Angular Second Moment (ASM). These are the remaining attributes that will be used as input parameters to the second input of the model to classify the original image into six zones.

Class 1 2 3 4 5 6

IV.4. Precipitation Classification into Six Classes The radar reflectivity maps can be used to diagnose areas of stratiform and convective precipitations [70], [71], [72]. Convective precipitations are characterized by high reflectivity and a clearer spatial intensity gradient [71]. Several authors propose thresholds to identify convective clouds [71], [73]. Based on the radar data, the developed technique allows classifying the MSG image area into the six following subareas of different rainfall intensities: 1) The very high precipitation intensities: this class corresponds to the precipitation zone which the corresponding radar reflectivity is greater than 46.0


Rainfall intensities very high precipitation moderate to high precipitation Moderate precipitation light to moderate precipitation Light precipitation No precipitation

Rainfall rate (mm/h)

Radar reflectivity (dBZ)

>12

>46.0

7.5 to 12.0

42 to 46

2.8 to 4.6

34 to 38

1.0 to 1.7

26 to 30

0.2 to 0.6

12 to 22