The state of play in machine / environment interactions Aitkenhead, M.J., McDonald, A.J.S. Department of Plant & Soil Science, University of Aberdeen, Aberdeen AB24 3UU, Scotland, UK Tel +44 (0)1224 272700 Email:
[email protected] Abstract Due to the breadth of the subject, it is no longer possible to provide a review of all of the work being carried out in the field of Artificial Intelligence. However, a more localised review of research taking place in the overlap between engineering, AI and psychology can be meaningfully performed. We show here that while there have been marked successes in the past few years, there is an identifiable set of ‘classic’ problems that remain to be solved, and which largely direct the work ongoing in this area. This review aims to discuss the directions being taken at the current time, in particular the developing and maturing possibilities provided by neural networks and evolutionary computation, and by the use of our knowledge of the mind in developing artificial agents capable of mimicking our abilities to interact with the environment. Keywords Artificial intelligence, autonomous systems, complex systems, evolutionary computation, neural networks, robotics
1. Introduction 'The environment that we live in' is a phrase that has different meanings for different people. Whichever response is given to the question 'where are you from?' the answer will always encompass some system that is complex, full of patterns and events which the individual has to react to in order to survive. The very fact that we are capable of observing our environment and ourselves is compelling evidence that the Universe is structured, that it contains physical patterns and chains of cause and effect. Biological brains have evolved to take advantage of this fact of information compressibility, allowing them to learn and recognise patterns. Human intelligence is a highly developed application of this principle, including as it does complex communication of information between individuals, and thus the spreading of these patterns from one mind to another. The Scientific Method is a way of trying to learn about our environment that has grown from this basic intelligence. We have developed a system where not only do we carry out observations of our environment in an attempt to learn about it, but we form hypotheses that, once proven or disproved, are incorporated into our knowledge and then built upon. Smith (1980) discussed the nature of discovery from a scientific standpoint, and showed that the relationship between science, technology and everyday activity is a complex one in which actions cannot be attributed to one specific category or the other, but are instead involved with all three. Many people have a great depth of understanding for their selected environment. These are the people that study one small part of the overall Big Picture in great detail and as a result become very good at predicting the behaviour of the Universe under a very narrow set of conditions. Others have greater breadth of knowledge, allowing them to form opinions about a wider-ranging set of circumstances. However, everyone is restricted to a small surface area on this mental map, whether it be wide and shallow, or narrow and deep. This concept of a mental map is discussed, along with others ways of knowledge representation, in Bench-Capon (1990). Knowledge representation was defined by Bench-Capon as 'A set of syntactic and semantic conventions that makes it possible to describe things'. The advantage that our own human evolution has given us, with the more recent development of the Scientific Method, is that once a particular area of this map has been explored, the records are there for everyone who can read to look at it. Artificial Intelligence may allow us to uncover the blank areas of the knowledge map faster and more reliably, through the development of automated learning and pattern-recognition processes. The use of computers, with 1
their speed and storage capacities and their ability to share error-free information at the digital level, to mimic the properties of biological intelligence has great potential if it can be made to work. The problem is twofold: first, understand the workings of one of the most complex devices known to science, the human brain; second, express this understanding in a computerbased form which works in a functionally identical manner to the original. In many ways, the attempts that have been made so far in the subject of Artificial Intelligence resemble not so much scientific endeavour as a combination of science, engineering and artistry: • Science, because we are building upon known information and at the same time exploring new areas on the knowledge map • Engineering, in the use of known methods to solve a structural problem • Art, for the many forms, interpretations and expressions that each individual brings to the subject 1.1. Definition In order to capably interact with the environment within which it finds itself, an intelligent organism must have both sensory and motor abilities, and some way of processing the sensory input. Implementing a system that does fulfil all three requirements requires a fourth concept, that of integrating each of the components into a functioning whole. There are two important concepts that guide researchers in the area of artificial intelligence implementation: the method by which learning takes place; and the existence of some set of goals that the implementation must satisfy. These two ideas are firmly entwined, and the existence of one without the other within a system implies either a finished tool lacking in developmental potential or knowledge without application. The majority of research in this area to date has comprised the search for a learning method that will allow the system to satisfy specific, preconceived goals. Only in a smaller number of cases has investigation been made into a more general method that will allow a system to learn about and adapt to any environment, and in most of these cases any set of life-goals has been noticeable by its absence. The concept of integration as mentioned above involves finding input, processing and output paradigms that can each be described using the same terms. This requirement for complementary designs implies that the system’s senses allow it to know how it is influencing or interacting with its environment. A crude example of a system that would not satisfy this requirement would be to have a robot capable of movement but which detected its environment using only a sense of hearing. Certainly it would be theoretically capable of learning, over time, how its actions led to different sounds around it, but it would be far more efficient to provide the robot with a different sense, such as vision, which would allow a more direct awareness of its actions. From the standpoint of this review, the term ‘machine/environment interactions’ is used to mean any situation where a device is used in such a way that its behaviour and internal representation is affected by the context of its surroundings, and the surroundings themselves are affected in turn by the behaviour of the device. More specifically, examination is made of those situations in which complex environments exist, requiring some or all of the aspects of biological intelligence – including pattern recognition, behavioural adaptation or planning. An important emphasis is placed on the abilities of the device to act without human intervention, or to be able to provide a statement saying that such-and-such an action is required for such-and-such a situation. The range of concepts and physical systems used within this definition is broad, covering control systems, robots in the classical sense and knowledge representation tools. 1.2 Potential benefits AI is not simply a means to an end, a series of steps towards a future where machines carry out the work while humans laze around. There is also the question of our drive to understand of the Universe. Intelligent machinery that is capable of learning about its environment can potentially provide us with a method of augmenting our own senses and minds. Here again we encounter the dichotomy that exists in science: research for application, or research for no reason other than to satisfy our curiosity. While it can be argued that this curiosity is simply another way for us to improve our chances of survival, it is also an aspect of human nature without which life would be much duller. This review, as well as providing a discussion of current methodologies and thinking in the scientific and engineering community, will cover the potential benefits of improved machine/environment interactions in terms of both scientific and practical benefits to humanity. 2
Specific cases and general topics will be included, with examples given of those problems that are currently the greatest challenges to be overcome. 1.3 Organisation The aim of this review is to provide a broad description of the present state of machine/environment interactions. The material is divided into six sections of which this Introduction is the first. The second section covers the history and background theory behind artificial intelligence as pertaining to interactive systems. The third section describes those areas of more general interest to human activities. In the fourth section, discussion is made of the relevant research that are being carried out in this area, and which define the subject by their existence. This section also highlights the concept of biological plausibility, the idea that intelligent machines capable of interacting with their environment can be designed along lines that mimic the structure or function of biological organisms. Section five builds upon section four by describing several specific challenges to AI research. These challenges combine certain aspects of biological intelligence that are of special interest to those working in the fields of robotics and artificial intelligence. The final section gives a brief conclusion of the points covered and discusses possible future developments.
2 The history of machine/environment interactions 2.1 Psychological models of mind Although implementations of AI were not possible until the middle of the twentieth century, studies of the peripheral and central nervous systems (PNS & CNS) of humans and animals allowed researchers to create conceptual models of the mind, its purpose and workings. These models provided an important foothold for the concept that if a system can be described completely, then an algorithm can be written that will perfectly simulate its behaviour. Leahey (1980) provides a good account of the earliest periods of psychology, within the philosophy of the Greeks. At this point all scientific thought was an effort to understand the Universe, and it is at this early stage (approx. 500 B.C.) that we see the first psychological models, in which human acts were attributed to internal reason rather than to the acts of gods. Aristotle (384-322 B.C.) in particular attributed to living things a soul that functions within the body to produce actions and thoughts. Descartes (1596-1650) was one of the first to try to understand the manner in which the brain and the rest of the body are connected, and how the brain can affect and be affected by actions of the body. Bolles (1993) gave a good discussion of this and other examples throughout history of the invocation of some 'spirit' that is in some way necessary and separate from the body in order for intelligent behaviour to be observed. Johannes Müller (1801-1858) is an important figure in the history of mechanistic psychology. A professor of physiology, he had a deeper understanding of the workings of the human body than the more usual philosopher-psychologists whose theories had dominated to the early nineteenth century. Müller is important not only because of his conjecture that neural and nervous activity is electrical in form, but because of four of his students (Emil du Bois-Reymond, Hermann von Helmhotz, Carl Ludwig and Ernst Brücke) who became powerful figures in psychology. In the last two hundred years, the work of these researchers and others has provided more and more evidence for the mechanistic nature of the brain and reducing the need for an inexplicable soul. For an in-depth coverage of Western psychology's most influential minds throughout the last 19th and 20th centuries, see Nordby & Hall (1974). The current level of understanding of the mind can be split into two categories: smallscale, dealing with individual components and how they function; and large-scale, concerned with the overall structure and design of the nervous system and how these influence intelligence. Small-scale knowledge of the nervous system covers a wide range of disciplines such as biochemistry, neurophysiology and even quantum mechanics. The chemical, biological and physical properties of synapses and nodes have been intensively studied, and these investigations allow conceptual models of these components to be improved. Studies of chemical behaviour of components including those of Matzel & Gandhi (2000), and of neuronal and synaptic dynamics (Rosenberg et al., 1998), have highlighted the importance of the relationship between design and function down to the lowest structural levels. Much work has been done specifically on the regulation of synaptic efficacy, for example Davis (1995) and Friedlander et al. (1996). This is an important topic as synaptic efficacy is the 3
dominant method of controlling signals between neurons. Markram & Tsodyks (1996) showed that the direction of information to specific regions of the brain could be achieved by altering the 'gain' of synapses, expressed in the level of neurotransmitter release. This mechanism could allow for the effects of Long-Term Potentiation (LTP), an important aspect of memory. A potential mechanism for this 'gain' alteration in synapses is through the recent discovery of biochemical two-way communication between neurons and astrocytes, another major cell population of the brain, by Carmignoto (2000). Wolff et al. (1995) and Neubig & Destexhe (2000) have showed that structural as well as biochemical alterations take place in the brain as a result of learning. The results of Wolff et al. (1995) are particularly interesting as they show that, particularly during early CNS development and possibly later in life, synapses may be remodelled or replaced several times. This realisation may have important implications for AI researchers attempting to design models of mind. Large-scale knowledge of the nervous system is usually expressed in psychological terms. Duch (1996) discussed developments in the field of psychophysics, in which the manner through which sensory stimuli cause computation to occur within the mind is studied. This computational physics approach treats the function of the brain at the symbolic level, looking for patterns that might allow us to decode the programming language of the brain. Others study the effect of anatomy on the function of the brain, such as Daya & Chauvet (1999) who argued that the geometry or location of individual components are important to the stability of the neural networks patterns of activity, or Rusakov et al. (1995) who studied the population dynamics of synapses during training events. Treves et al. (1997) showed that the speed of certain selective responses such as visual pattern recognition can only be achieved with feedforward pathways rather than the slower recurrent feedback loops. This is an important example of how larger-scale structure affects the brain’s performance. Another is the existence of more than one pathway involved in the vestibuloocular reflex (VOR) which functions to maintain the eye pointed in one direction when the head is moving, as modelled by Quinn et al. (1998). An important problem in such areas is in finding terms, either mathematical or linguistic, to describe the processes being observed. Freeman (1996) described studies of functional connectivity in the visual cortex of cats, but only gave descriptions of what was taking place in localised regions containing small numbers of cells. Kavanau (1997) discussed generalised procedures that have evolved within the brain to maintain memories and behavioural patterns, such as sleep. These behaviours include the entire CNS yet in many cases their actual mechanisms are poorly understood. What is known about the general function of the brain is that it is hierarchical in structure and that the information it contains is often distributed across a large number of neurons, rather than being located in one particular area. Beyond this basic comprehension, there are many different models demonstrating functions observable in the brain. Examples include Hirsch (1997) and Chialvo & Bak (1999). Each of these models may work perfectly well, but the problem remains that our evolutionary development has resulted in one highly complex model that accomplishes all of the observed functions. The models that we do have are not sophisticated enough to predict everything that takes place in the mind. Integration of biological and electronic signalling components is the highest expression yet of our understanding of the nervous system. This requires an understanding of not only the biological, chemical and physical structures of our nervous system, but also of the signalling processes involved, for example the relationship between spike rate in auditory nerve fibres and sound pressure, as investigated by Schoonhoven et al. (1997). In essence, we have to translate the language with which our bodies communicate with themselves. Baev (1997) discussed a model if biological nerve signalling that in many ways mirrors the function of the internet, with increasing levels of abstraction as one moves from the bit-by-bit transfer at the lowest level (the physical layer) up to the presentation of concepts in the application layer. Albert (1999) presented a model of an early evolutionary stage of the nervous system, addressing the problem of how the first form of adaptive behaviour came about. The conclusion reached was that within a network of epithelial-like cells, changes that improve the behaviour of the system are those that make the spreading of the electric potential across the group more efficient. This implies that the action potential was evolved as a method of transferring electric charge across the cell group, and that the concept was later made more complex by the introduction of synaptic efficacy. Okamoto et al. (1999) developed an artificial neuron capable of mimicking the behaviour of a biochemical one, an important step towards the construction of artificial neural networks that can be grafted onto natural ones. This concept has also been explored by Heiduschka & Thanos (1998) with their discussion of successfully implanted bioelectronic interfaces such as artificial 4
cochlea and peripherally implantable nerve stimulators. At the moment, such devices are still being experimentally developed to improve their functional similarity to natural systems, but the level and rate of success leads one to conclude that the barrier between humans and machines is not as tall as it once was. 2.2 Early history of AI implementations McCorduck (1979) gave a lucid account of the earliest attempts at artificial intelligence, including the animated statues of Daedalus; the legend of Joseph Golem of Rabbi Judah ben Loew; and the mechanical duck, which 'ate, drank and beat its wings', of Jacques de Vaucanson. Prior to these efforts, Aristotle (5th century B.C.) had provided the first steps to artificial intelligence with his syllogistic logic, a system of reasoning with information. A logical framework within which facts may be contained and manipulated is a vital component of any intelligent system, and there are strong links remaining to this day between the fields of logic and AI. Another important figure in this stage is that of Ramon Lull, about whom Glymour et al. (1998) wrote: “Many of the fundamental ideas in artificial intelligence have an ancient heritage. Some of the most fundamental, surely, are that thinking is a computational process, that computational processes involve combining symbols, that computation can be made mechanical, and that the mathematics of computation involves combinatorics. All of these ideas have their origin, so far as we know, in the work of an eccentric 13th century Spanish genius, Ramon Lull (1232-1316). Lull's sources were partly mystical, but the interesting part of his thought drew from - or against - an analytic tradition in logic and combinatorics.” 2.3 Technological trends The development of technological and theoretical advances in AI is strongly driven by the need for practical, commercial applications. Initial predictions given based on the early successes of AI have largely failed to appear, and it is instructive to ask why this is the case. While materials science and technological developments since the birth of AI in the 1950’s have been dramatic, it can be argued that the main obstacle to advancement has been more to do with limitations in our models of how intelligence actually works. This has resulted in an expansion in the number of applications of low-level intelligence, but the notable absence of anything HAL-like. Another plausible restriction on human- or even animal-level intelligence is that of computational requirements. Estimates given for a computer to be able to match humans in terms of raw processing speed vary widely, but limits are commonly placed between 10-1000 THz. Even when we achieve the necessary computational power however, this will not be the same as saying that machines have the same capabilities for learning and interacting as humans. A great expansion in our understanding of the workings of the brain is required before that can be achieved. While a discussion of psychological models of mind and may seem irrelevant to the topic of current knowledge in machine/environment interactions, it is instructive to see how our understanding of how intelligence works has been expressed in models. This process is continuously evolving, with new paradigms of mind resulting in new implementations of AI. Our current stage of development reveals that we know more about some areas of how the mind works than about others, and shows that work is required towards an understanding of how the smallest components of the mind are organised so as to provide the emergent properties of intelligence.
3 Areas of potential benefit 3.1 Military It has been said that the two oldest professions, the sex industry and the military, are the first to benefit from any technological advances. While there is no evidence - yet - of artificially intelligent sex toys, there is plenty of evidence that the implications of AI have not been ignored by military minds. Automating many of the processes which are currently carried out by humans has the potential to provide faster, more reliable and safer operations in combat situations, and to enable a reduction in personnel numbers in non-combat, support tasks. Ibnkahla (2000) discussed the use of neural networks in digital communications applications, and found that they are potentially very useful in channel identification, fault 5
diagnosis and other areas. For the military, whose reliance upon the ability to communicate is vital, improved reliability and security of a communications network is an obvious goal. Cadutal (1998) also emphasised the use of AI technologies in creating and maintaining an Information Grid for use in deploying military force, pointing out that ‘smart’ systems can operate much faster and reliably than humans when simple tasks are being carried out. Pattern recognition in its many forms hold potentially great benefits for military applications. Kamm et al., (1997) discussed the importance of real-time voice command recognition for computer control in the near future, while Pasquariello et al. (1998) described an AI system of target recognition from radar information that could be used to automate fire control systems. Sumpter & Bulpitt (2000) discussed a method of object tracking using neural networks that is also applicable to this area, while O’Malley et al. (2002) discussed human tracking using video camera technology, for example in airports where there are areas that are off-limits to unauthorised personnel. Talukder & Casasent (2001) demonstrated a neural network capable of recognising facial poses and individual faces, using only one example of each original face to develop the necessary information. Other similar systems could no doubt be used to recognise and pinpoint specific objects, such as aircraft, ground troops or tanks. Zardecki (1995) and Lin & Wang (1998) gave more specific military-oriented examples. Zardecki discusses the uses of fuzzy controllers in efforts to safeguard against discrepancies in nuclear material inventories. Lin & Wang proposed an adaptive controller for bank-to-turn missiles that relies upon a self-organising system, and found that the AI-based system had superior tracking and convergence performance to non-AI systems. Illi (1996) gave a detailed discussion of how an expert system could be used to change the U. S. Army's present system of dealing with equipment maintenance through reaction to problems to a more proactive system of predictive problem solving. This system would cut maintenance time and reduce the number of catastrophes encountered when a vital subsystem fails at a critical moment. Faller & Schreck (1996) discussed the implications that neural networks can have on automatic control systems within aircraft to augment the flying ability of pilots and to monitor potentially thousands of sensors in parallel, allowing for aircraft systems with enhanced safety and performance capabilities. 3.2 Scientific Biological sciences have begun to take advantage of AI in order to model complex systems in which the governing equations are partially or not at all known. Flood (1998) discussed the practicality of using neural networks in situations governed by poorly understood partial differential equations, and shows that highly accurate results can be obtained by using this method to predict processes. Alberdi & Sleeman (1997) presented a system for adjusting taxonomy systems to deal with fresh data, which could prove to be the forerunner of a learning expert system, one which mimics the learning and hypothesising abilities of a human researcher. Ladunga (2000) described a method of predicting expressed proteins from genomic information, while Walley & Fontama (1998) used a similar system to predict the biological sensitivity of a river from the environmental characteristics of the site. These are good examples of how AI methods can be used to model and allow us to understand large, complex data sets rapidly. AI systems are very applicable to the problems encountered by many field scientists, such as Tipping et al. (1999) for whom the large volume of experimental data was found to be difficult to organise and explain using anything other than general, qualitative terms. They are also applicable to scientists attempting to model a system in which many subjects are combined (e.g. geology, physics, chemistry and mathematics for modelling attributes of a cometary body in Benkhoff and Boice, 1996). Zhai et al. (2006) and Levine et al. (1996) both used neural networks to classify soils using the backpropagation algorithm and found that the accuracy of results was equal to or greater than that obtained from other methods. Other applications of AI to soil science include soil hydrology, e.g. Schaap & Leij (1998). Minasny et al. (1999) provided a comparison of various AI and statistical methods in estimating the hydraulic properties of soil. Weiss & Baret (1999) found that when modelling biological systems using AI methods, it is necessary to ensure that the variables used are good indicators of the actual dynamics of the system. If the variables are ambiguous or irrelevant then the system will perform no better than other statistical methods. This point is indicative of the need, as stated earlier, for an AI system to be able to interact with its environment in an optimal fashion, using ‘senses’ that are attuned to the environment in which it exists. Time series prediction is a subject very relevant to weather forecasting, something that not only scientists are influenced by. The nature of climate dynamics is chaotic, with many
6
empirical rules and high sensitivity to initial conditions. Zhang et al., (1997), and De Oliveira et al. (2000) used AI (notably neural network) methods in prediction of weather-related subjects. Other applications of artificial intelligence to biological sciences cover more general areas. Lek & Guégan (1999) discussed AI as a tool in ecological modelling, finding several useful applications using different types of neural network (NN) algorithm. Fedorenko et al. (1999) also showed how NNs, in conjunction with the recordings of natural processes can be used to reveal previously undetected dynamics and relationships. Physical sciences have also benefited from connectionist AI methods. The triaxial compression behaviour of sand and gravel has been studied by Penumadu & Zhao (1999), who used NNs to develop a nonlinear stress-strain model. Numerous other models exist describing specific aspects of the physical world, of which the above is merely one example. The coupling of databases with AI systems has been shown to be a useful method of developing expert systems, although work remains to be done in the area of translation of information from databases into forms that a learning system will accept. Yang (1997) discussed the main methods of integration that have been developed and compared a novel one with existing methods. De Jong & Rip (1997) also investigated the potential of AI as a tool for scientific discovery, concluding with a set of guidelines for developers to follow that was intended to optimise the work of both scientific researchers and users of AI as a discovery tool. Horneck (1996) discussed the problems of artificial ecosystems in biologically hostile situations, focussing on the Moon. The complexities of such a system are such that an AI control system would likely be required in order to achieve and maintain an environment capable of supporting life. Another application of AI in space exploration is as a flight control system, coordinating schedules and handling sudden alterations to the situation. Nakasuka & Tanabe (1996) investigated the practicality of using an AI system architecture in control of space operations, and found that such a system worked well in conjunction with human operators. 3.3 Commercial & Industrial In industry, any processes that can be reliably automated will have the effect of both reducing labour costs and speeding operations. With this in mind, AI-based systems have become more common over the last decade in their use as control systems for industrial plants. Morimoto et al. (2000) discussed a pattern recognition technique in fruit shape classification, a topic that could be extended into many more factory quality-control applications. They found that a combination of neural network, fractal theory and fruit image database could provide them with a reliable automated fruit shape classification system. Timmermans & Hulzebosch (1996) used a similar system for pot plants identification that gave >99.9% accuracy between two types of pot plant. Ray & Hsu (1998) discussed two methods of reliably classifying image sequences into recognised situations, while Sumpter & Bulpitt (2000) suggested methods of system behaviour prediction. Wong et al. (1998) described a hidden Markov model capable of detecting abnormalities in system operation that could be used as a monitoring system for industrial plants, once again highlighting the need for a synthesis between observation and interaction in any functioning AI system. In a detailed paper, Di Sciascio et al. (2002) discussed the problems of retrieving objects from within complex images, without relying on low-level characteristics such as colour distribution or texture. With faces, low-level characteristics may not be enough to allow correct recognition. Colour and lighting are variable, and image quality and texture may also differ between images. For high-level characteristic recognition, a hierarchical method of feature extraction that builds upwards from simple structures to more complex ones is required. The fact that so many different methods are being developed in one subject area prompts the suggestion that a synthesis of methods combined in a results-driven approach may work better than a single method-driven search for a solution to the image recognition problem. Paraskevas et al. (1999) described an expert system with the capability of learning from mistakes while supervising the entire control system of a wastewater treatment plant. Tsaih et al., (1998) used a hybrid of neural networks and rule-based systems to develop an AI approach to stock-market predictions. During a six-year testing period from 1988 to 1993, their Integrated Futures Trading System (IFTS) outperformed standard investment strategies in trading. These are important examples of AI being used to model a system in which the governing relationships are poorly known. Another example is that of Chen et al. (1998) who developed a neural network capable of predicting the chemical properties of an unknown compound. This approach has major potential applications in materials research, where a chemical or material with specific properties is often sought through exhaustive and expensive trial-and-error experimentation. Other examples of AI methods dramatically reducing R&D costs is that of Wong et al. (2000) who used NNs to 7
predict hydrocarbon reservoir capacities, or Yeh (1997) who designed a system capable of automatically controlling the cutting speed in a tunnelling system under different conditions. Agatonovic-Kustrin & Beresford (2000) discussed the implications for AI in R&D for commercial chemical companies. They concluded that the potential applications cover a wide range of topics including data analysis, computer-user interface and the prediction of chemical properties. In this field as in many others in the commercial world, AI is proving to have many applications. The main obstacle at the moment appears to be that each specific application is being developed independently, with a myriad of AI implementations that all share common features but that are all designed specifically for one task. The development of a more generalised methodology would perhaps prevent many people from having to cover the same ground in basic R&D work. Bench-Capon (1990) argued however, that it is the narrow field of expertise available within expert systems that has resulted in their relative commercial success compared to more generalised methods. It is much easier to build a system designed to learn and advise only in small subset of possible situations, and the results in practical terms are more useful and reliable. The relevance here to this review lies in the way different approaches are used depending on the breadth and depth of knowledge required to solve a particular problem, and also depending upon the goals of the person trying to solve that problem. Academic research has an eye on the future potential of a method as well as its present utility, while corporate, military and industrial users are too constrained by present events to have one eye on what may happen further down the road.
4 Relevant fields of research 4.1 Expert systems An expert system is a type of AI in which generalised problems are not addressed. Instead, a particular field of interest is focussed on in the hope that depth of ability in a narrow set of circumstances will outweigh the inability to deal with a wide range of situations. A common method of problem solving with expert systems is through a sequence of options in which the system narrows the problem successively to a point where a particular known solution can be applied (Bonnet et al., 1988). In this manner, expert systems can be thought of as a method of problem classification, with successful classifications resulting in a problem-category that is sufficiently narrow that all situations within that category can be considered identical and therefore identically treatable. However, this ability to identify known problems is not a perfect option. New problems are continually cropping up, and a system that only knows how to solve the known problems is soon redundant. Smith (1980) provided a good case for the argument that the most important scientific discoveries take place when two disciplines are brought together. Inventiveness is a capacity that expert systems are not designed with, and the same can be said for many other symbolic approaches to AI, as discussed by Bellman (1969). The creation of new concepts is something that symbolic approaches rarely tackle, although they can be efficient in other ways. De la Rosa et al. (1999) used a neural network with an expert system to develop new rules that dramatically improved the expert system’s performance, once again showing that a combination of several methods may be more useful than a single technique. Bench-Capon (1990) discussed how the implementation of expert system techniques has resulted in commercially viable working systems, as opposed to the more abstract results obtained through generalised AI research. Pigford & Baur (1990) discussed the pros and cons of expert systems in terms of business applications, and find that these systems can be very advantageous in terms of reliability, cost and explanation. However, they also find that the main problem lies in extracting the expert knowledge in the first place, when many of the heuristics that the experts use may not be known to them in terms that they can describe, or when the rules may be extremely vague or subjective in nature. Bonnet et al. (1988) explored the development of an expert system from first concept to industrial application. They revealed that most expert systems founder at a relatively early stage due to one of two main reasons: (1) lack of enthusiasm by those in charge of the developers, who are often working on the project without the knowledge of those higher up; (2) lack of success, which is better expressed as a failure to live up to the inflated promises made by the developers. However, in recent years there have been many examples of small to moderately complex expert systems applied in such a manner that their presence, although arguably very helpful, goes largely without comment. Examples of this hidden technology are web-based search engines, industrial plant control systems and monitoring and control in modern cars (Darwiche & Provan, 1997). 8
There is a strong argument for the fact that expert systems have been largely successful where their developers have not been too ambitious, and that these successes have been gradually accumulating in usefulness to the point that many tasks would be impossible without them. 4.2 Neural networks Artificial Neural Networks (ANNs) are a deceptively simple yet powerful concept. The basic idea behind training a network is that when two nodes are connected, changes to the connecting synapse depend upon the level of activation of the nodes. This gives the network the ability to use past experience to predict how likely it is that the occurrence of one event (activation of the sending node) will take place in conjunction another event (activation of the receiving node). Using specific relationships for the dynamics of such a system, simple nervous systems can be constructed. Much of the current work in this area is focussed on developing these specific models of network dynamics in order to produce complex and useful emergent behaviours. In terms of interaction with complex environments, neural networks hold the potential to provide a method of developing autonomous, complex and hierarchical behavioural patterns. The model from which the study of ANNs has evolved and developed was first introduced by McCulloch & Pitts (1943), with their 'Threshold Logic Unit', in which the outputs from each node were of the all-or-nothing kind (0 or 1), depending upon whether the received signal was higher than some threshold. Since that time, many people have introduced models of neural networks that are more complex in terms of structure or component dynamics, but the basic concept remains, that two events occurring repeatedly together are related. Neural network study can be divided roughly into two groups. There are those that attempt to use our knowledge of the brain’s structure in order to produce biologically plausible NN models, for example Bugmann (1997). In addition, there are those who attempt to come up with novel network dynamics models in attempts to develop working models, at the expense of biological plausibility, e.g. Liu (2000). This includes those researchers who use neural networks for purposes that are completely different from intelligence as it is commonly viewed. Examples of this class are analog signal filtering (Mehr & Sculley, 1996) and chemical modelling (Fernández & Caballero, 2006). Although initially this second group developed dramatically more successful systems, the biologically plausible methods have been gaining in ability as our knowledge of brain structure and function improve and we are able to implement new concepts (Changeux & Dehaene, 2000). In recent years, a third group has been emerging, one that is taking advantage of technological progress by building neural networks directly in silico, avoiding the traditional reliance upon the serial processors of conventional computers. This third group often uses a combination of biological realism and novel concepts. For example, Agranat et al. (1996) developed a silicon-based synapse to be used as the basis for large-scale ANNs, bridging the gap between modelled and real neural networks. The synapse behaved exactly how theoretical models predicted it should, although the models were very different to the behaviour of a biological synapse. The concept of attractors is one that is vitally important to neural networks. Ezhov & Vvedensky (1996), among others have described a method of generating attractors within a neural network. Networks may contain several attractors that can be activated individually from one another, allowing the network to distinguish between different input node activation patterns. The utility of attractors is revealed by the fact that if we have two initial activation configurations that are sufficiently close to one another, and each is fed into an attractor network, then over a period of time each will fall into the same attractor. This fact has great relevance for temporal (Kühn & Cruse, 2005) and spatial (Tsodyks, 2005) pattern recognition, and has been used to develop theories of memory (Ng & Feng, 2001). Many people have studied the dynamics of attractors and their relevance to memory, for example Erichsen & Theumann (1995) and Kinzel (1999). Another important concept within the subject of neural networks that has a parallel with mathematical and physical sciences is that of energy. A rough way of describing the concept is this: if the configuration of synaptic efficacies in a network of N nodes is {a1, a2, a3….aN}, and the configuration required for it to respond correctly to inputs is {b1, b2, b3….bN}, then the system can be given a mathematically-derived energy of:
E = ∑ | a n − bn |
(eqn. 1)
N
Using this concept has allowed researchers to derive methods of reducing the energy, and thus the errors, of the system (Brouwer, 1995). Various extensions of this concept have been 9
investigated, including Lyapunov Functions (François & Zaharie, 1999) and stored pattern capacity (Lin et al., 1998). In terms of applications, neural networks are very useful for identifying nonlinear patterns in systems. Gupta & Sinha (1999) discussed this property and the advantage that it gives over other methods. Particular interest has been paid to image recognition systems, for example Cheng et al. (1997) and Chella et al. (1997). Much work has been done on analysis of the portion of the brain that carries out face recognition and related functions (Dailey & Cottrell, 1999). Franco & Treves (2001) demonstrated a neural network based facial expression recognition system. Their method, which gave higher accuracy than many other methods, used a self-organising NN capable of learning autonomously. Kirchberg et al. (2002) used the Hausdorff method for detection of frontal face images. This method uses a predefined edge model of the human face to detect candidates in images, and was adjusted using genetic algorithm methods and existing facial images to modify a simple ellipse model. The use of genetic algorithms is particularly useful in situations such as facial recognition, where an exact definition of the system being studied is not available. This lack of a template is also discussed by Liu & Wang (2001) who argued the importance of feature extraction and perceptually meaningful representations in relation to classification problems, including facial recognition. NNs are commonly applied to voice and language related problems, for example McNamara et al. (1998) and Albesano et al. (2000). Speech recognition requires the ability to deal with temporal sequences, for which a variety of methods have been applied. Attractor neural networks are particularly popular, for example Pulasinghe et al. (2004) and Gicquel et al. (1998). Other methods include the Hidden Markov Model, a statistical method in which both a transition probability matrix and a feature observation matrix are derived from observations for each action, and comparison is made between the observed action and the candidate actions, in order to find the best fit. This method has been applied by Choi & Rhee (1999) and Salmella et al. (1999), amongst others. Hidden Markov Models are often combined with neural networks to produce adaptive time-sequence recognition systems. Traditionally, the skills that humans are capable of are the ones researched most commonly, both because of the drive to increase automation in the workplace and because it allows us an easy method of evaluating the system performance against an existing benchmark (ourselves). Bouslama (1999) compared three NN methods of printed Arabic character recognition and found that the most effective were those that carried out substructure recognition, otherwise known as feature extraction, which is an important characteristic of human pattern recognition. Yun & Bahng (2000) discussed a method of substructure detection and stress the importance of this factor in reducing the descriptive length of the system under discussion. This concept goes back to the discussion in the Introduction of the use of intelligence to compress the information that is available to them from their environment. Stoecker et al. (1996) described a system of substructure recognition or scene segmentation using correlations between neuronal activity in a network where neuron activity was purposely desynchronised. This is an interesting and potentially useful concept, as frequency modulation in neural activity is possibly another way in which information is stored in the brain. Another function that NNs can perform is that of behaviour adaptation. With a suitable method of adjusting synaptic values following responses to sensory input, an NN can be trained as a control system, e.g. Starrenburg et al. (1996), Samejima & Omori (1999). In many such examples, the NN is integrated into the control system of a robot, allowing an expression of the behavioural patterns developed. A diametrically opposite approach has been taken by Bieszczad & Pagurek (1998), who described a method by which NNs could rapidly search their state space to solve problems. This method made use of certain aspects of NN dynamics without requiring or providing a true neural network system with adaptive behaviour. From this and other work, it is obvious that no true consensus has yet emerged concerning the best approach to take with neural networks when attempting to solve even a specific problem. When trying to tackle the requirements of a more general situation, the disagreement is even more obvious. 4.3 Artificial life The field of Artificial Life includes many different topics, but can be described relatively simply: it is the study of methods which allow us to simulate aspects of biological life. As such, it can be considered inclusive of all artificial intelligence research, and also of robotics. This broad categorisation means that the field of Artificial Life tends to be used as an umbrella term, rather than as a topic that anyone studies in itself. However, it is a useful definition to use as it includes all aspects of work involving simulation of biological and environmental processes.
10
Much effort has gone into expressing our understanding of biological systems through models. The supposition is that if we understand a process sufficiently well then we should be able to model it effectively, as explained by Horiuchi & Koch (1999) in their modelling of neurobiological systems. Langton (1996) discussed this area of work in terms of the various scales at which the modelling of biological systems may occur. One of the most successful efforts in these terms is that of Schleiter et al. (1999) who developed an ecosystem model operating at several hierarchical levels successfully. The importance of this lies in the fact that biological intelligence tends also to comprehend its environment in a hierarchical manner, using different assumptions and rule sets at different levels of scale and complexity. Cellular automata (CAs) are an important component of Artificial Life, as the concept of many small components interacting to produce emergent properties is shared both by CAs and biology. Olson & Sequeira (1995) in a review of ecological modelling methods found that CA systems gave highly effective emergent property predictions. Aitkenhead et al. (1999) used CA methods to model moisture retention curves in soils and found that, at scales larger than that of the individual CA cells, the method gave effective predictions of emergent properties of soils including hysteresis. Rietman (1994) gave several examples of systems that have been successfully modelled using CA methods, including NNs and genetic evolution. In a converse of this, Giles et al. (1995) discussed a method of training neural networks so that they mapped between states finite state machines, a generalisation of CAs. Other CA examples include ecosystems modelled at the organism level, traffic control systems and Turing's Universal Computer (Brooks & Maes, 1996). Pentland & Liu (1999) used a combination of AI and CA models to predict human movement when driving. The use of animats (artificial animals) is an effective way of developing models of animal behaviour, singly or in groups, in real life or in a virtual system. Animats can exist as mobile robots, or as simulations within a computer program. Ilg & Berns (1995) applied a hierarchical NN to the coordinated movement of a six-legged walking robot, while Krebs & Bossel (1997) investigated several theories of behaviour in populations of animats. The animat concept is central to Artificial Life, as they exist as expressions of biological organisms in simulated or modelled form. As theoretical test beds or demonstrations of observed behaviour, they allow us to investigate the interactions between organisms and the environments they exist in. One of the most important relationships between AI and Artificial Life is that in both, simple concepts are used to develop large, complex systems. Agre & Horswill (1997) gave a description of a method of depicting how objects interact with their environment, using static and dynamic symbolic relationships. Such a system would be useful as a framework for integrating different artificial life models, and could lead to the development of new ways of expression of artificial intelligence. 4.4 Robotics The field of robotics is one that can be considered in its own right, and which overlaps with AI rather than being a subset of the topic. Robots are not always designed to mimic intelligent behaviour, but rather can be custom-built for a specific purpose. There are two main areas in which robots are used; for research purposes and in industrial applications. These two areas strongly reflect the presence or absence of AI in their design, for reasons that are themselves reflective of how people consider the field of robotics to be important. When used in industry, robots tend to be employed in simple tasks that are too dangerous, monotonous or physically difficult for a human to be able to accomplish. One of the most recognisable applications is that of a car production line, where robots perform repeated tasks such as spray-painting, welding or manoeuvring body parts. Industrial applications often use AI control systems that do not rely on learning through behaviour adaptation, but instead have experts systems or fuzzy control mechanisms. The reasons for this are twofold: first, we do not yet have the ability to develop artificially intelligent robots to the point where they can be trusted to learn and act without anything going wrong, and secondly because the use of robots in an industrial application is generally for the purpose of replacing some human operator who carried out a specific task that was too dangerous or otherwise physically demanding. We have yet to ask robots to think for us, only to replace our muscles. Depending upon the complexity of the robot behaviour, the control system may be no more complex than a set sequence of operations (a controlled system), or it may be organically integrated with the sensory and motor systems and contain an adaptive system (an uncontrolled system, Brafman & Tennenholz, 1996). In recent years there has been an explosion of research into the integration of AI system brains with robot bodies. The adaptive control system principles 11
used have covered a wide range, including neural network architectures (Jerbic et al., 1999), symbolic representations (Kilmer, 1997) and imitation learning (Schaal, 1999). In many adaptive control cases, some form of reinforcement function is used to optimise the robot's behaviour in its phase-space (Johannet & Sarda, 1999). Other options include a random walk through the phasespace searching for advantageous couplings between input and output states (Steels, 1997), which takes longer but is guaranteed to give optimal results eventually. Fuzzy control systems such as those developed by Maeda et al. (1995) and Castellano et al. (1997) directly map the sensory input to control commands without using any internal representation of the data. Output commands are given weightings depending upon the comparison with pre-set patterns within the set of input data, allowing the robot to interact with an environment that it does not have complete understanding of. This solves the problem of dealing with a complex environment but does not allow for adaptation or optimisation of behaviour. An evolutionary approach that alters the behaviour of a robot in order to make it more effective in its environment is likely to produce a system better capable of generalising, but which performs less well in highly constrained situations than a system specifically designed for these situations. Harvey et al. (1997) and Nolfi (1997) used evolutionary methods, which have an advantage over the fuzzy system in that over time, the robot’s behaviour changes to fit the situation at hand. However, it is time that can prove to be an obstacle in this case, with many generations of required to optimise the system. An additional consideration is that the use of an evolutionary approach does not constitute true learning by the system. Whether this is meaningfully different from the adjustment of internal knowledge in terms of environmental interaction is a matter for some discussion. Neural networks have been intensively applied to robot control, particularly autonomic. Akbarzadeh-T et al. (2000) have shown that neural networks in combination with pre-defined actions have allowed behaviour adaptation to take place. Mobile goal-oriented robotic behaviour research has been particularly popular, with Gaussier et al. (2000) and Zhao & Collins (2005) presenting systems capable of moving to specific points or following particular routes. One of the main goals in this area is vehicle guidance. Broggi & Bertè (1995) gave a good discussion of the problems associated with vehicle guidance automation, of which the main one appears to be dealing effectively with information in real-time. Daxwanger & Schmidt (1996) succeeded in reproducing the behaviour of a skilled human driver in docking situations using a neural architecture, while Buessler & Urban (1998) obtained success in using a modular neural network design. An important requirement of successful vehicle guidance is that of generating smooth transitional movements. Seung (1998) described analog position control achieved using a continuous attractor in which any activation sequence eventually collapsed to a single point in the activation phase space. Damper et al. (2000) used a combination of learnt and hard-wired approach in an animat control system to allow smooth movement. The combination of adaptive and hard-wired systems allowed complex behaviour to develop from simple, predefined hardware actions. An important aspect of vehicle navigation is planning ahead. Artale & Franconi (1998), for example, presented a symbolic logic system for planning & representing temporal action sequences. An important consideration here is that it is arguably the case that the ability to plan ahead is something that humans do as a result of their ability to both communicate through language and memorise environmental features, while nervous systems of the simplest animals exist solely in the here and now. In attempting to implement aspects of higher-order intelligence using lower-order intelligence design factors, the situation becomes muddied. The main sensory method used by autonomous mobile systems is that of vision. For navigation, visual examination of the environment is almost a prerequisite, especially for truly autonomous systems as discussed in Von Wichert (1999), in which the importance of a selforganising image recognition system is emphasised. Most of the work in vision-based navigation to date depends upon landmark recognition, in which recognition of specific objects within the environment allows determination of position (Von Wichert, 1998; Lin et al., 1998). Other approaches include recognition of particular feature types, such as lane-detection (Baluja and Pomerleau, 1997), and the use of a search for the most distinctive features in a scene (Weng & Chen, 1998). A combination of feature recognition and situation awareness is often important (Gaussier et al., 1997). Surprisingly, colour is rarely used in feature recognition due to complications of shading, lighting etc., but it can be used to discriminate between components in a scene (Buluswar & Draper, 1998). Image recognition is seen as not only important for robots, but also as a method of human-computer interaction superior to those that are currently available (Cipolla & Pentland, 1998). McCafferty (1990) presented an account of the current knowledge in statistical methods of 12
image analysis, and showed that there are several aspects of vision important to human object recognition that use imaging methods different to the standard pixel intensity concept. Edge and line detection, texture and differentiation of patches of colour are all important aspects of human vision. In terms of design, many successful intelligent robots rely on a hierarchical structure of behaviour. Saksida et al. (1997) and Kaiser & Dillmann (1997) both emphasised the low-level skill integration into more complex actions, while Thrun & Mitchell (1995) argued that the transfer of knowledge from one task to another would allow a robot to become more capable at dealing with new tasks over time. Mahajan and Figueroa (1997) discussed an adaptation of this concept in which knowledge and learning skills were embedded at several levels within a robot control system, so that the knowledge and behavioural patterns used at any time were related to the context of the situation. Certainly, there is a lot to be said for retaining knowledge that may not be applicable at the present time but that might prove useful later. This argument emphasises the need for a long-term memory component within the system, one that can allow recall of past events in order to compare them with present circumstances. Recent work suggests a drive towards integration of several different AI methods within a single control system design, allowing a robot to accomplish a variety of navigation and other tasks. Burgard et al. (1999) discussed such a multi-methodology system used in a museum tour guide robot capable of communicating with people, manoeuvring and planning. This approach of solving different problems using different methods is probably of more practical use than any of the approaches that rely on a single method, although one has to wonder if it contributes anything to the advancement of the field as a whole. The use of several different paradigms at once to solve a problem or interact with the environment cannot be as effective in the long term as using a single more generalised paradigm that allows better interaction with the environment because it is more grounded in reality. 4.5 Evolution The use of evolutionary strategies in the development of artificial intelligence has been applied at many levels and in many ways. The common goal in each application however has been to maximise the fitness of the system according to some predefined criteria, through the mutation or some other alteration of the system over many generations. Within the system, control variables are identified and treated as the 'genes' of the system, and adjusted in a manner related to biological evolutionary development to produce an optimal system. Evolutionary methods are particularly suited to complex systems. Morimoto & Hashimoto (2000), for example, adopted an evolutionary strategy in their approach to industrial plant production system optimisation. Sette et al. (1998) used a combination of genetic algorithms and neural networks for a similar purpose, and found that optimisation of the neural network architectural control variables through genetic algorithms resulted in an improved control strategy. The advantages of using evolutionary methods over other techniques of system optimisation are relatively obvious and easy to implement when the fitness of the system is easy to determine, such as the growth rate of a plant (Morimoto et al., 1996) or comparing the predictions of a model against reality (Noever et al., 1996). However, when the fitness is calculated subjectively it is easy to end up with a system that does not actually give meaningful or useful results. Zhang et al. (1997) discussed the evolution of symbolic relationships for modelling complex systems. They found that the fitness definition was important to the evolution of an optimal system, and that when complex systems were being dealt with the fitness should be partially dependent upon the complexity of the model. The reason for this is that in many cases, the model may grow extremely large and complex with only slight improvement in actual ability. With increased complexity comes lower likelihood of altering the model in an effective manner and therefore rejecting the accepted model in favour of a superior one. Beer (1997) gave a good discussion of the general field of genetic algorithms from a control system perspective, with great emphasis placed on the importance of considering the dynamics of the system being controlled. The argument expressed was that if one fully understood the system dynamics then one would have a better appreciation of the fitness of the system, and thus would be better able to evolve an optimal control paradigm. Smith & Cribbs (1997) also took this view, with their discussion of the evolution of autonomous system behaviour. The difference here is that the fitness was defined in terms of some behavioural error indicator, or punishment/reward factor. The determination of this factor is given in terms of the actions taking place. If a mobile system bumps into something, for example, it is punished. If it moves without bumping then it is rewarded.
13
Problems are encountered in situations that are so complex that reward or punishment attributes have to be given in terms that humans can understand, such as sex, hunger or pain. The problem then becomes one of expressing these concepts meaningfully to the learning system. Floreano & Mondada (1998) discussed this in terms of robot control systems and concluded that different environments can result in different control strategies. Moriarty et al. (1999) also covered this problem when dealing with adaptive behaviour. All of the above research dealing with complex systems relies heavily upon a userdefined measure of fitness. This seems to be the main failing in the use of evolutionary methods of system optimisation, as it places constraints on the system’s ability to adapt further than the user’s imagination. The evolutionary paradigm is highly effective at optimising a solution to a known problem, but each individual problem has its own set of control variables. Therefore the evolutionary method is not one that can be applied to generalised learning in artificial intelligence without being combined with another paradigm that supplies the new concepts. To summarise, the umbrella term of evolutionary computation covers a variety of methods developed relatively recently by which optimisation is achieved in computer systems. Various aspects of biological evolution are used, in a variety of different way, to improve the performance of computer processes. While there are many ways in which these aspects are implemented, there are some common themes that can be used to generalise about the techniques used: • The methods do not aim to produce solutions that are already defined; rather, as in biology, the level of fitness achieved by a system regardless of how it works is the measure of evolutionary success. • Random changes to the genotype of the system are caused through mutation or crossover; these changes cause alterations to the behaviour of the phenotype (with some methods, the genotype and the phenotype are actually the same). • The ‘fitness’ of offspring is determined dependant upon the situation in hand, with fitter offspring being more likely to produce offspring of their own and so propagate the successful genotype. The suite of techniques used to evolve an optimal solution to a particular problem has been proven very useful in situations where the problem (the system’s ‘environment’) is complex or where the optimal solution is difficult to predict. In situations where the solution is unknown either from environmental complexity or a difficulty in definition, it is almost always the case that as long as the phenotype can be in some way expressed as a genotype capable of being subjected to evolutionary pressures, at least some improvement in system performance can be found. There is a conflict apparent between the use of an evolving neural network system and its connection with a machine body. Whereas a biological organism’s physical structure evolves in concert with the nervous system, it is impossible to ‘evolve’ the physical structure of a robot. This implies that a half-way house in which the nervous system evolves while the physical structure remains the same, or changes only in large increments through the addition of a new component. 4.6 Biological realism When arguing the case for biological plausibility in AI, there is an immediate temptation to say ‘And why not? At least we know it works!’ While this is a perfectly good reason, there are others. It has been argued that intelligence as we understand it can only be achieved with a massively parallel connectionist system such as the brain. Certainly rule-based systems that start with a top-down approach have problems in generalising, and require constant special-case adjustment. Big rules have little rules, the better to describe them, while little rules have smaller rules and so ad infinitum. McCarthy & Hayes (1969) proposed dividing the problem of AI into two parts, epistemological and heuristic. Epistemology studies which and what kinds of facts are available to the researcher given certain observational opportunities, how these facts can be represented and what rules allow conclusions to be drawn from these facts. The heuristic part of AI examines how to search for a certain response and how to recognise patterns. In the past thirty or more years, this division has underlain the development of logical and symbolic approaches to AI. Yet the fact that the division can be made at all is caused by the dependency on the concepts of facts, rules and symbols, and their representation within an AI system. When approached from this standpoint, the field of AI becomes a minefield of conceptual problems, many of which have to do with the representation of ‘common sense’. One of the most important and general of these obstacles is the Frame Problem, an example of which would be: if you are stirring a pan of soup on the cooker with your right hand and the liquid begins to boil over, how do you know to use your left hand to turn down the heat? 14
From a biologically plausible, connectionist standpoint, the Frame Problem and others like it are not so much resolved as they never get a chance to become an issue. Neural networks do not work with concepts, or if they do then these concepts are at the ‘atomic’ level where one node interacts with another through the transmission of a single signal. Behaviour synonymous with intelligence from this point of view is emergent, built up from many simple interactions that act independently of one another on the local level, but which come together on larger scales to produce complex, intuitive and commonsensical activity. 4.6.1 The Importance of Sensory Feedback In order for an organism, either natural or artificial, to interact successfully with its environment, it has to learn which patterns of behaviour are appropriate in certain situations. For this to happen, the organism must have some comprehension of what it is doing. It is necessary, therefore for those parts of the organism that define its behaviour (e.g. arms, legs) to also have some sensory abilities, in order to communicate back to the central control system what is happening to them. Nechyba & Xu (1994) emphasised the importance of humans’ ability to coordinate their activities through sensory feedback, and also investigated the potential of neural networks in identifying and controlling systems, and through this their potential to identify human control strategies that use sensory feedback. They argued that sensory feedback was a highly important component of processes that occur rapidly and which are highly variable. 4.6.2 Modularity Within a complex system, there are often components that perform different operations from one another. It makes sense therefore to have a control system that is structured so that signals sent to one subsystem are not also sent to all others. Within the central nervous system, the division of the brain into modules in this way prevents disruptive interference from other parts of the system, and allows more complex tasks to be carried out through their subdivision into ever smaller tasks that are carried out in parallel to one another. While it is easy enough to state that modularity is a benefit to the AI practitioner, there still remain problems in implementation. How can we arrive at a description of the system’s modularity that optimises performance? In a hierarchical system with several layers of complexity, which subsystems should be put together, and which should be isolated from one another? It is often difficult to work out even a simple system’s optimal topology, while in systems sufficiently large and complex to show ‘interesting’ behaviour, altering the design becomes no more than guesswork. A method is required that allows the system’s topology to be both described and efficiently manipulated in a search for the ‘fittest’ system, leading us back to evolution. 4.6.3 Adaptation Having a method of describing a learning system that can be subjected to evolutionary pressures allows the system to improve its performance, or ‘fitness’ within its environment. In the case of AI, this fitness is defined as a measure of the ability to learn. The success of using evolutionary methods is dependant of course, on a definition of the fitness being available and implementable. Fitness in terms of a goal being accomplished or not is not a useful concept in this case. Complex systems that either work perfectly or do not work at all are not susceptible to the charms of evolutionary programming, as there are always so many more permutations of the system that will not work than there are that will. A smooth gradient of fitness within the environment is required, thus negating the concept of a ‘goal’ that is either accomplished or is not. Karl Popper’s concept of Evolutionary Epistemology, which describes knowledge acquisition as a process of conjectures and refutations, is an important concept running in parallel with that of evolutionary computation. Theory development as a process of trial and error, while messier and less pure than the ideals of the Scientific Method, is nevertheless an important underlying process in daily scientific life. 4.6.4 Connectionism Neural networks can be viewed from the stance that each node is a concept, and each connection is a rule. These concepts and rules are so simple that in terms of real-world meaning their simplicity prevents them from contributing very much. However, when observed from 15
several steps up the hierarchical scale, the emergent behaviour of a large neural network admits sophisticated learning and behaviour development. One of the most important characteristics of NNs is that these atomic concepts and rules can change over time, dependant upon their activity. This flexibility within the system eliminates problems of symbolic, logic-based systems that attempt to trap and define the complex, fluid behaviour of biological organisms. However, it is one thing to have a system understand its environment, but another thing entirely to interpret that understanding. A solution to this problem would appear to be to train a neural network in some topic and then translate the knowledge embedded within the network into symbolic notation. In situations where it is important to not only have an accurate model but also to explain the model to others, such as land use systems as discussed in Rodrigue (1997), or any other complex environmental model that could affect peoples livelihood, e.g. Mackay & Robinson (2000) this would be a great benefit. Schenker & Agarwal (1997) developed a neural network system based on system states that could be used to develop expert system-like models of systems. However, the reliance upon pre-existing system states removed some of the apparent promise of this method. The syntactical self-organising map (Grigore, 1997) also has these same pros and cons. Kozma (1997) described a NN system in which the network structure itself changed over time, resulting in a system with deep (structure-dependent) and shallow (synapse efficacy dependent) knowledge about its environment. This approach allows fundamental relationships between features in the environment to be noted through the presence or absence of a connection between them, and simplifies extraction of information from the network. Pedrycz (1991) described a method in which first a fuzzy rule set was developed, then was translated into a neural network. The problem with going the other way, from neural network to fuzzy rule set or expert system is that even when concepts have been found, labelled and placed in a framework they are not necessarily recognisable. García-Pedrajas (2006) described changes in NN structure during training and evolution as being dependent upon both training data and the semi-random evolutionary approach applied. The concepts and relationships between these concepts may well have relevance and meaning in the real world, but each neural network is written differently, with its own syntax and semantics, as a result of sensitivity to initial setup conditions and the sensory inputs that the network has processed since its initiation. 4.6.5 Autonomy In the past, useful AI methods have largely been applied using reactive model representations, which is to say that the model of a particular system has been developed from observation and then put into practice using an AI representation. This has been most visible in the field of expert systems, where human knowledge is transferred into a computer that then gives advice on a specific topic. In many cases where a behavioural control system is used and the system training is automated, this training is carried out in a controlled simulation of the environment before being used in the real world. AI systems which are proactive, i.e. which learn directly from their environment, have largely been used in an effort to develop evidence of intelligent behaviour in machines, although Cohn et al. (1996) argued that interaction with the environment provides a powerful method of obtaining new information and learning how the environment operates. French (2000) discussed how far people have come in their quest to pass the Turing Test, a goal that has strongly influenced efforts in the field of AI for the last 50 years. The conclusions given by French are that while the initial hopes of success have proved unfounded, the Turing Test still remains an important landmark and goal of our abilities to enable machines to behave independently. Dean (1998) argued that animats could provide a method of developing models about environments, if one could extract from them the information that they have learnt. As mentioned above in reference to the evolution of effective behavioural patterns, neural networks have been used many times to develop automated controllers or models of specific systems, but in each case the developed network has been inscrutable, giving results that may very well be accurate but that are given without explanation. In many cases, the problem is compounded by the fact that the practicalities of AI system design are elusive, leading to difficulties not only in interpreting the stored information but also in ensuring that the information is stored well in the first place. Husmeier (2000) discussed the fact that the success of a neural network model is usually dependent upon a trial-and-error system of design development, showing that the concept is still imperfectly understood. Nikravesh et al. (1997) published guidelines for the structure of a dynamic neural network control system, based upon efforts to optimise the stability of the network. Even this however did not constitute an understanding of how overall neural network properties (such as a measure of learning ability) 16
depend upon the structural control variables. Sensitivity and uncertainty analysis carried out by neural networks is another area in which the separation of model uncertainty from network error is important but difficult (Ricotti & Zio, 1999).
5 Frontier problems In many fields of research, there are ‘classic’ problems that often define the current work being carried out. In mathematics for example, there exists a set of unsolved problems, many of which carry large prizes for those able to solve them. Examples include the search for a proof of the Goldbach conjecture, that every even number is the sum of two primes, or the search for a solution of the P versus NP question, in which it is asked if it is possible to check an answer quickly in a computational sense, but impossible to find the answer to that question quickly. In much the same way, the current research in the field combining artificial intelligence and robotics has thrown up, if not well-defined classic problems, then at least topics around which research has crystallised and intensified. Some of these topics and challenges are discussed here from the standpoint of environmental interaction. 5.1 Chess Due to high-profile events, chess-playing programs have become synonymous with AI in the public view. The most famous example of these events was in 1997, in which reigning World Chess Champion Gary Kasparov was defeated by IBM’s Deep Blue. The method used involved a brute force approach, with the computer performing a search of all moves possible from its current position, and looking multiple moves ahead. This arguably has less to do with intelligence than with raw computing power, although it demonstrated an approach to parallel processing that could be used in certain areas of AI. More recently in 2003, Kasparov played a newer commercially available system named Deep Junior, developed by Shay Bushinsky and Amir Ban. This successor to Deep Blue relies less on exhaustive search and more on game strategy, and was able to tie in six games. While a chess board can be considered an environment of sorts, it is certainly a narrowly restricted one. Many aspects of AI research including vision and robotics are ignored completely in this situation, while planning procedures dominate. It is also fair to say that a computer does not comprehend chess in the same way as a human. A more realistic challenge would be to have a human play against a robotic opponent, one that was capable of visually scrutinising the board, of moving pieces, and also of attempting to defeat an opponent. This would involve a marriage of several aspects of AI research, and would arguably provide us with a first step towards true mimicry of specifically human behaviours. Despite these objections however, chess-playing is still a useful example to use here as it demonstrates one of the most popularised ways in which human and computer interactions take place. 5.2 Football Alongside the FIFA 2006 World Cup in Germany was the Robot Football World Cup competition. A large number of teams participated, each designed by a group from a different country. The rules of robot football are roughly similar to those applied to the ‘beautiful game’, with eleven autonomous players in each team. The AI challenges existing to participating designers include navigation, football control (dribbling), real-time information processing, team coordination and ‘goal’ satisfaction. Each of these challenges presents a serious side to the contest, with multiple prospective real-world applications for technology that provides a working solution. A stated aim of the international RoboCup organisation is to ‘By the year 2050, develop a team of fully autonomous humanoid robots that can win against the human world soccer champions’. In order for this aim to be achieved, a great many obstacles will have to be overcome. Tyler & Czarnecki (1999) discuss many of these obstacles and present a neural network-based vision control system for a robot footballer. As is obvious from other research efforts ongoing, the benefits of achieving even partially in this area will undoubtedly lead to advances in machineenvironment interactions in a wide number of different fields. 5.3 Agriculture Concerns have been raised in recent years about the level of chemicals, in both herbicide and fertiliser form, that are added to crops to improve yields. In 1996, the amount of chemical 17
herbicide used in the UK alone was 23,000 tonnes, at a cost of approximately £400 million (Marchant, 1996). One possible way of reducing this pressure on the environment and human health from chemical residues is to improve our ‘spot application’ ability of herbicide, through automated visual discrimination between crops and weed plants. In recent years, much research has been carried out in utilising computer vision in agricultural inspection. Emphasis is necessary on the ability to deal with real-time, high speed processing of colour images, as described by Brosnan & Sun (2004). Various methods have been applied to automated plant recognition, and plant morphology has been recognised as a useful method of discriminating between species, since leaf patterns and overall shape provide unique ‘fingerprints’ of individual species. Morphology measurements can be carried out relatively quickly in comparison to other image analysis techniques, and quite simple methods can be shown to provide good results (Aitkenhead et al., 2003). Von Wichert (1998) discussed the fact that in a natural environment, it is impossible to define every visual parameter described for autonomous navigation to be possible. This implies that the system must be able to learn for itself which aspects of its environment are important. Vision-based systems based on self-organising neural networks have already been shown to work (Gaussier et al., 2000), and have provided a route through which undefined visual patterns may be learned. This work has potential not only for plant discrimination, but for many other aspects and applications of machine vision. 5.4 Natural language processing The abilities of humans to communicate through speech and writing are often taken for granted, and fail to be appreciated for the sophistication and complexity of the processes involved. This is possibly partly because we are capable of carrying out these processes with little or no conscious effort. However, as has been pointed out previously in this work, just because such a thing is common and easy for humans to perform does not mean that the problem of automating it is one that can be solved easily. It has been clear for some years that one of the operations that the human brain has specifically evolved, and which it is indirectly shaped by the requirements of, is speech processing, although the specific purpose of language in human society are still being argued over. The common view is that the ability to communicate with one another gives an obvious boost to our chances of survival through the sharing of information about food or danger. Dunbar (1996) however, postulated that language developed less as a method of communicating useful information and more as a substitute for grooming and other social activities. Pattern recognition is not the only issue involved with the processing of natural languages. Individual words can be recognised by current systems with high accuracy, but this is of little use if the meaning behind the words is not understood. In order to communicate meaningfully with a human, an intelligent machine will also require background knowledge, or common sense as it is normally termed. This is by far the most challenging aspect of this area, and encompasses areas of research that go beyond language. In truth, if an AI system were capable of conversing meaningfully with humans then it would be solving most if not all of the major problems of artificial intelligence research. Natural language processing therefore is in many ways the holy grail of AI researchers.
6 Conclusions To date, the interaction of artificially intelligent systems with complex environments has been restricted mostly to the realm of scientific research, with a few tried-and-tested applications for specific, narrow sets of situations. There is a sense in the artificial intelligence community that much work is being done on solving individual problems, but that the methods developed in finding these solutions are not being integrated, or investigated in terms of their comparative merits. A gap is therefore developing between the growing knowledge base of the topic and the application of this knowledge base into working systems capable of generalisation, or of autonomous behaviour. One of the fields of AI which originally promised so much but which has dwindled in effectiveness over recent years is that of neural networks, the very paradigm which has the most potential to provide generalised learning and behavioural skills. The suite of evolutionary methods for optimising neural network performance has been proven many times over. However, to date nearly every application using these methods has been developed using some set of higher-level, control variables as a genotype for the network. This makes the system dependent upon the user’s choice of these control variables, which may limit the 18
future success of the method. Instead of control variable evolution, a method with more latitude in describing the neural network topology and dynamics is needed. One obvious method of solving this problem would be to find a more fundamental way of describing the system, moving towards a model reflective of biological DNA that involves gene expression and cell differentiation. Already some researchers are looking in this direction in the hope of modelling the growth and development of biological organisms, but there is still some way to go. In terms of environmental interaction, the basic problems of navigation and obstacle avoidance have been solved in a manner that allows robots to manoeuvre in a simple environment, but little more than this has been accomplished. The real-world environment still requires generalisation and recognition abilities far beyond what we can currently develop. However, there is much potential in research being carried out into the low-level structure and dynamics of biological learning systems, with their massively parallel implementation of small, rapid and relatively simple components. Minsky’s Society of Mind puts this concept well, although the situation is complicated by the fact that these simple components may not be as simple as we once thought. While virtual environments are a cost-effective and useful way of interacting with simple AI systems, the difficulties of simulating a complex environment imply moving the development of more sophisticated systems into the real world. This brings in the engineering aspects of robotics as discussed earlier, along with the problems of designing a suitable environment for the animat to interact with and develop within. As these systems progress up the scale of complexity and intelligence, moving from insect analogues towards the world of speech and abstract concepts occupied by humans, their learning will require more and more interaction with human trainers. After all, if a human infant with its enormous learning potential is kept in isolated conditions without interactive stimulus, it will never learn to communicate with the rest of us.
7 References Agatonovic-Kustrin S, Beresford R (2000) Basic concepts of artificial neural network (ANN) modelling and its applications in pharmaceutical research. Journal of Pharmaceutical and Biomedical Analysis 22: 717-727 Agranat AJ, Schwartsglass O, Shappir J (1996) The charge controlled analog synapse. Solid-State Electronics 39: 1435-1439 Agre P, Horswill I (1997) Lifeworld Analysis. Journal of Artificial Intelligence Research 6: 111145 Aitkenhead MJ, Dalgetty IA, Mullins CE, McDonald AJS, Strachan NJC (2003) Weed and crop discrimination using image analysis and artificial intelligence methods. Computers and Electronics in Agriculture 39:157-171 Aitkenhead MJ, Foster AR, FitzPatrick EA, Townend J (1999) Modelling water release and absorption in soils using cellular automata. Journal of Hydrology 220:104-112 Akbarzadeh-T M-R, Kumbla K, Tunstel E, Jamshidi M (2000) Soft computing for autonomous robotic systems. Computers and Electrical Engineering 26:5-32 Alberdi E, Sleeman DH (1997) ReTAX: a step in the automation of taxonomic revision. Artificial Intelligence 91:257-279 Albert J (1999) Computational modeling of an early evolutionary stage of the nervous system. BioSystems 54:77-90 Albesano D, Gemello R, Mana F (2000) Hybrid HMM-NN modeling of stationary-transitional units for continuous speech recognition. Information Sciences 123:3-11 Artale A, Franconi E (1998) A temporal description logic for reasoning about actions and plans. Journal of Artificial Intelligence Research 9:463-506 Baev K (1997) Highest level automatisms in the nervous system: A theory of functional principles underlying the highest forms of brain function. Progress in Neurobiology 51:129-166 Baluja S, Pomerleau DA (1997) Expectation-driven selective attention for visual monitoring and control of a robot vehicle. Robotics and Autonomous Systems 22:329-344 Beer RD (1997) The dynamics of adaptive behavior: a research program. Robotics and Autonomous Systems 20:257-289 Bellman R (1969) Modern analytic and computational methods in science and mathematics. American Elsevier Publishing Company Inc., New York Bench-Capon TJM (1990) Knowledge representation; an approach to artificial intelligence. APIC series, No. 32, Academic Press, London, UK Benkhoff J, Boice DC (1996) Modeling the thermal properties and the gas flux from a porous, icedust body in the orbit of P/Wirtanen. Planet. Space Sci. 44:665-673 19
Bieszczad A, Pagurek B (1998) Neurosolver: Neuromorphic general problem solver. Journal of Information Sciences 105:239-277 Bolles RC (1993) The story of psychology: a thematic history. Brooks/Cole Publishing Company, Pacific Grove, California Bonnet A, Haton J-P, Truong-Ngoc J-M (1988) Expert systems principle and practice. PrenticeHall Inc., New Jersey Bouslama F (1999) Neural networks in the recognition of machine printed Arabic. International Journal of Pattern Recognition and Artificial Intelligence 13:395-414 Brafman RI, Tennenholtz M (1996) On partially-controlled multi-agent systems. Journal of Artificial Intelligence Research 4:477-507 Broggi A, Bertè S (1995) Vision-based road detection in automotive systems: a real-time expectation-driven approach. Journal of Artificial Intelligence Research 3:325-348 Brooks RA, Maes P (eds) (1996) Artificial life IV. The MIT Press, Cambridge, MA Brosnan T, Sun D-W (2004) Improving quality inspection of food products by computer vision – A review. Journal of Food Engineering 61:3-16 Brouwer RK (1995) A method for training recurrent neural networks for classification by building basins of attraction. Neural Networks 8:597-603 Buessler J-L, Urban J-P (1998) Visually guided movements: learning with modular neural maps in robotics. Neural Networks 11:1395-1415 Bugmann G (1997) Biologically plausible neural computation. BioSystems 40:11-19 Buluswar SD, Draper BA (1998) Color machine vision for autonomous vehicles. Engineering Applications of Artificial Intelligence 11:245-256 Burgard W, Cremers AB, Fox D, Hähnel D, Lakemeyer G, Schulz D, Steiner W, Thrun S (1999) Experiences with an interactive museum tour-guide robot. Artificial Intelligence 114:3-55 Cadutal JT (1998) Artificial intelligence support for the United States armed forces' "System of systems" concept. U.S. Army War College, Pennsylvania Carmignoto G (2000) Reciprocal communication systems between astrocytes and neurones. Progress in Neurobiology 62:561-581 Castellano G, Attolico G, Distante A (1997) Automatic generation of fuzzy rules for reactive robot controllers. Robotics and Autonomous Systems 22:133-149 Changeux J, Dehaene S (2000) Hierarchical neuronal modeling of cognitive functions: from synaptic transmission to the Tower of London. International Journal of Psychophysiology 35:179187 Chella A, Frixione M, Gaglio S (1997) A cognitive architecture for artificial vision. Artificial Intelligence 89:73-111 Chen CLP, Cao Y, LeClair SR (1998) Materials structure-property prediction using a selfarchitecting neural network. Journal of Alloys and Compounds 279:30-38 Cheng H, Liu L, Li G, Shao L, Zhou C (1997) Second-order interpattern neural networks for optical pattern recognition. Optics Communications 139:182-186 Chialvo DR, Bak P (1999) Learning from mistakes. Neuroscience 90:1137-1148 Choi H, Rhee P (1999) Head gesture recognition using HMMs. Expert Systems with Applications 17:213-221 Cipolla R, Pentland A (eds) (1998). Computer vision for human-machine interaction. Cambridge University Press, Cambridge, UK Cohn D, Ghahramani Z, Jordan MI (1996) Active learning with statistical models. Journal of Artificial Intelligence Research 4:129-145 Dailey MN, Cottrell GW (1999) Organization of face and object recognition in modular neural network models. Neural Networks 12:1053-1073 Damper RI, French RLB, Scutt TW (2000) ARBIB: An autonomous robot based on inspirations from biology. Robotics and Autonomous Systems 31:247-274 Darwiche A, Provan G (1997 Query DAGs: a practical paradigm for implementing belief-network inference. Journal of Artificial Intelligence Research 6:147-176 Davis GW (1995) Long-term regulation of short-term plasticity: a postsynaptic influence on presynaptic transmitter release. J. Physiology 89:33-41 Daya B, Chauvet GA (1999) On the role of anatomy in learning by the cerebellar cortex. Mathematical Biosciences 155:111-138 De Jong H, Rip A (1997) The computer revolution in science: steps towards the realization of computer-supported discovery environments. Artificial Intelligence 91:225-256 De la Rosa D, Mayol F, Moreno JA, Bonsón T, Lozano S (1999) An expert system/neural network model (ImpelERO) for evaluating agricultural soil erosion in Andalucia region, southern Spain. Agriculture, Ecosystems and Environment 73:211-226 20
De Oliveira KA, Vannucci A, da Silva EC (2000) Using artificial neural networks to forecast chaotic time series. Physica D 284:393-404 Dean J (1998) Animats and what they can tell us. Trends in Cognitive Sciences 2:60-67 Di Sciascio E, Donini FM, Mongiello M (2002) Structured knowledge representation for image retrieval. Journal of Artificial Intelligence Research 16:209-257 Duch W (1996) Computational physics of the mind. Computer Physics Communications 97:136153 Dunbar R (1996) Grooming, gossip and the evolution of language. Harvard University Press, Cambridge, Massachusetts, USA Erichsen R, Theumann WK (1995) Learning and retrieval in attractor neural networks with noise. Physica A 220:390-402 Ezhov AA, Vvedensky VL (1996) Object generation with neural networks (when spurious memories are useful). Neural Networks 9:1491-1495 Faller WE, Schreck SJ (1996) Neural networks: applications and opportunities in aeronautics. Prog. Aerospace Sci. 32:433-456 Fedorenko YV, Husebye ES, Ruud BO (1999) Explosion site recognition; neural network discriminator using single three-component stations. Physics of the Earth and Planetary Interiors 113:131-142 Fernández M, Caballero J (2006) Bayesian-regularized genetic neural networks applied to the modeling of non-peptide antagonists for the human luteinizing hormone-releasing hormone receptor. Journal of Molecular Graphics and Modelling 25(4):410-422 Flood I (1998) Modeling dynamic engineering processes when the governing equations are unknown. Computers and Structures 67:367-374 Floreano D, Mondada F (1998) Evolutionary neurocontrollers for autonomous mobile robots. Neural Networks 11:1461-1478 Franco L, Treves A (2001) A neural network face expression recognition system using an unsupervised local processing. In: Proceedings of The Second International Symposium on Image and Signal Processing and Analysis (ISPA'01), Croatia, 2001 François O, Zaharie D (1999) Markovian perturbations of discrete iterations: Lyapunov functions, global minimization, and associative memory. Mathematical and Computer Modelling 29:81-94 Freeman RD (1996) Studies of functional connectivity in the developing and mature visual cortex. J. Physiology 90:199-203 French RM (2000) The Turing test: the first 50 years. Trends in Cognitive Science 4:115-122 Friedlander MJ, Hersanyi K, Kara P (1996) Mechanisms for regulating synaptic efficiency in the visual cortex. J. Physiology 90:179-184 García-Pedrajas N (2006) Cooperative coevolution of neural networks and ensembles of neural networks. Studies in Computational Intelligence 16:465-490 Gaussier P, Joulain C, Banquet JP, Leprêtre S, Revel A (2000). The visual homing problem: an example of robotics/biology cross fertilization. Robotics and Autonomous Systems 30:155-180 Gaussier P, Revel A, Joulain C, Zrehen S (1997) Living in a partially structured environment: how to bypass the limitations of classical reinforcement techniques. Robotics and Autonomous Systems 20:225-250 Ghaboussi J, Sidarta DE (1998) New Nested Adaptive Neural Networks (NANN) for constitutive modeling. Computers and Geotechnics 22:29-52 Gicquel N, Anderson JS, Kevrekidis IG (1998) Noninvertibility and resonance in discrete-time neural networks for time-series processing. Physics Letters A 238:8-18 Giles LC, Horne BG, Lin T (1995) Learning a class of large finite state machines with a recurrent neural network. Neural Networks 8:1359-1365 Glymour C, Ford KM, Hayes PJ (1998) Ramón Lull and the infidels. AI Magazine 19:136 Grigore O (1997) Syntactical self-organising map. Lecture notes in Computer Science 1226:101109 Gupta P, Sinha NK (1999) An improved approach for nonlinear system identification using neural networks. Journal of The Franklin Institute 336:721-734 Harvey I, Husbands P, Cliff D, Thompson A, Jacobi N (1997) Evolutionary robotics: the Sussex approach. Robotics and Autonomous Systems 20:205-224 Heiduschka P, Thanos S (1998) Implantable bioelectric interfaces for lost nerve functions. Progress in Neurobiology 55:433-461 Hirsch MW (1997) On-line training of a continually adapting adaline-like network. Neurocomputing 15:347-361 Horiuchi TK, Koch C (1999) Analog VLSI-based modeling of the primate oculomotor system. Neural Computation 11:243-265 Horneck G (1996) Life sciences of the Moon. Adv. Space. Res. 18(11):95-101 21
Husmeier D (2000) Learning non-stationary conditional probability distributions. Neural Networks 13:287-290 Ibnkahla M (2000) Applications of neural networks to digital communications - a survey. Signal Processing 80:1185-1215 Ilg W, Berns K (1995) A learning architecture based on reinforcement learning for adaptive control of the walking machine LAURON. Robotics and Autonomous Systems 15:321-334 Illi OJ (1996) Future diagnostics technology. Expert Systems with Applications 11:147-155 Jerbic B, Grolinger K, Vranjes B (1999) Autonomous agent based on reinforcement learning and adaptive shadowed network. Artificial Intelligence in Engineering 13:141-157 Johannet A, Sarda I (1999) Goal-directed behaviours by reinforcement learning. Neurocomputing 28:107-125 Kaiser M, Dillman R (1997) Hierarchical refinement of skills and skill application for autonomous robots. Robotics and Autonomous Systems 19:259-271 Kamm C, Walker M, Rabiner L (1997) The role of speech processing in human-computer intelligent communication. Speech Communication 23:263-278 Kavanau JL (1997) Memory, sleep and the evolution of mechanisms of synaptic efficacy maintenance. Neuroscience 79:7-44 Kilmer W (1997) A command computer for complex autonomous systems. Neurocomputing 17:47-59 Kinzel W (1999) Statistical physics of neural networks. Computer Physics Communications 121122:86-93 Kirchberg KJ, Jesorsky O, Frischholtz RW (2002) Genetic model opimization for Hausdorff Distance-based face localization. Lecture Notes in Computer Science 2359:103-111 Kozma R (1997) Multi-level knowledge representation in neural networks with adaptive structure. Systems Research and Information Science 7:147-167 Krebs F, Bossel H (1997) Emergent value orientation in self-organization of an animat. Ecological Modelling 96:143-164 Kühn S, Cruse H (2005) Static mental representations in recurrent neural networks for the control of dynamic behavioural sequences. Connection Science 17(3-4):343-360 Ladunga I (2000) Large-scale predictions of secretory proteins from mammalian genomic and EST sequences. Current Opinion in Biotechnology 11:13-18 Langton CG (ed) (1996) Artificial Life: an overview. The MIT Press, Cambridge, Massachusetts, USA Leahey TH (1980) A history of psychology: main currents in psychological thought. Prentice-Hall Inc., New Jersey Lek S, Guégan JF (1999) Artificial neural networks as a tool in ecological modelling, an introduction. Ecological Modelling 120:65-73 Levine ER, Kimes DS, Sigillito VG (1996) Classifying soil structure using neural networks. Ecological Modelling 92:101-108 Lin C-K, Wang S-D (1998) A self-organizing fuzzy control approach for bank-to-turn missiles. Fuzzy Sets and Systems 96:281-306 Lin L-J, Hancock TR, Judd JS (1998) A robust landmark-based system for vehicle location using low-bandwidth vision. Robotics and Autonomous Systems 25:19-32 Lin X, Ohtsubo J, Mori M (1998) Capacity of optical associative memory using a terminal attractor model. Optics Communications 146:49-54 Liu P (2000) Max-min fuzzy Hopfield neural networks and an efficient learning algorithm. Fuzzy Sets and Systems 112:41-49 Liu X, Wang DL (2001) Appearance-based recognition using perceptual components. In: Proceedings of the International Joint Conference on Neural Networks 2001 (IJCNN-01), Washington D.C., USA, 2001 Mackay DS, Robinson VB (2000) A multiple criteria decision support system for testing integrated environmental models. Fuzzy Sets and Systems 113:53-67 Maeda M, Shimakawa M, Murakami S (1995) Predictive fuzzy control of an autonomous mobile robot with forecast learning function. Fuzzy Sets and Systems 72:51-60 Mahajan A, Figueroa F (1997) Four-legged intelligent mobile autonomous robot. Robotics & Computer-Integrated Manufacturing 13:51-61 Marchant JA (1996) Tracking of row structure in three crops using image analysis. Computers and Electronics in Agriculture 15:161-179 Markram H, Tsodyks M (1996) Redistribution of synaptic efficacy: a mechanism to generate infinite synaptic input diversity from a homogenous population of neurons without changing absolute synaptic efficacies. J. Physiology 90:229-232 22
McCafferty JD (1990) Human and machine vision: computing perceptual organisation. Ellis Horwood, New York, McCarthy J, Hayes PJ (1969) Some philosophical problems from the standpoint of artificial intelligence. In: Michie D (ed) Machine Intelligence 4, American Elsevier, New York McCorduck P (1979) Machines Who Think. W. H. Freeman and Company, San Francisco McCulloch W, Pitts W (1943) A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics 7:115-133 McNamara S, Cunningham P, Byrne J (1998) Neural networks for language identification: a comparative study. Information Processing & Management 34:395-403 Mehr I, Sculley TL (1996) A multilayer neural network structure for analog filtering. IEEE Transactions on Circuits and Systems II - Analog and Digital 43:613-618 Minasny B, McBratney AB, Bristow KL (1999) Comparison of different approaches to the development of pedotransfer functions for water-retention curves. Geoderma 93:225-253 Minsky M (1986) The Society of Mind. Simon and Schuster, New York Moriarty DE, Schultz AC, Grefenstette JJ (1999) Evolutionary algorithms for reinforcement learning. Journal of Artificial Intelligence Research 11:241-276 Morimoto T, Hashimoto Y (2000) AI approaches to identification and control of total plant production systems. Control Eng. Practice 8:555-567 Morimoto T, Takeuchi T, Miyata H, Hashimoto Y (1996) Intelligent control for a plant production system. Control Eng. Practice 4:773-784 Nakasuka S, Tanabe T (1996) New control problems associated with a proposed future space transportation infrastructure. Control Eng. Practice 4:1703-1714 Nechyba MC, Xu Y (1994) Neural network approach to control system identification with variable activation functions. In: Proceedings of the IEEE International Symposium on Intelligent Control, Columbus, Ohio, USA, 1994 Neubig M, Destexhe A (2000) Are inhibitory synaptic conductances on thalamic relay neurons inhomogeneous? Are synapses from individual afferents clustered? Neurocomputing 32-33:213218 Ng KT, Feng J (2001) Dynamical associative memory based on an oscillatory neural network. Journal of Intelligent Systems 11:155-171 Nikravesh M, Farell AE, Stanford TG (1997) Dynamic neural network control for non-linear systems: optimal neural network structure and stability analysis. Chemical Engineering Journal 68:41-50 Noever DA, Brittain A, Matsos HC, Baskaran S, Obenhuber D (1996) The effects of variable biome distribution on global climate. BioSystems 39:135-141 Nolfi S (1997) Evolving non-trivial behaviours on real robots: a garbage collecting robot. Robotics and Autonomous Systems 22:187-198 Nordby VJ, Hall CS (1974) A guide to psychologists and their concepts. W. H. Freeman & Son, San Francisco O' Malley PD, Nechyba MC, Arroyo AA (2002) Human activity tracking for wide-area surveillance. In: Proceedings of 2002 Florida Conference on Recent Advances in Robotics, Miami, USA, 2002 Okamoto M, Sekiguchi T, Tanaka K, Maki Y, Yoshida S (1999) Biochemical neuron: hardware implementation of functional devices by mimicking the natural intelligence such as metabolic control systems. Computers and Electrical Engineering 25:421-438 Olson RL, Sequeira RA (1995) Emergent computation and the modeling and management of ecological systems. Computers and Electronics in Agriculture 12:183-209 Paraskevas PA, Pantelakis IS, Lekkas TD (1999) An advanced integrated expert system for wastewater treatment plants control. Knowledge-Based Systems 12:355-361 Pasquariello G, Satalino G, la Forgia V, Spilotros F (1998) Automatic target recognition for naval traffic control using neural networks. Image and Vision Computing 16:67-73 Pedrycz W (1991) A referential scheme of fuzzy decision-making and its neural network structure. IEEE Transactions on Systems Man and Cybernetics 21:1593-1604 Pentland A, Liu A (1999) Modeling and prediction of human behaviour. Neural Computation 11:229-242 Penumadu D, Zhao R (1999) Triaxial compression behavior of sand and gravel using artificial neural networks (ANN). Computers & Geotechnics 24:207-230 Pigford DV, Baur G (1990) Expert systems for business: concepts and applications. Boyd & Fraser, San Francisco Pulasinghe K, Watanabe K, Izumi K, Kiguchi K (2004) Modular Fuzzy-Neuro Controller Driven by Spoken Language Commands. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics 34(1):293-302 23
Quinn K, Didier AJ, Baker JF, Peterson BW (1998) Modeling learning in brain stem and cerebellar sites responsible for VOR plasticity. Brain Research Bulletin 46:333-346 Ray SR, Hsu WH (1998) Self-organized-expert modular network for classification of spatiotemporal sequences. Intelligent Data Analysis 2:287-301 Ricotti ME, Zio E (1999) Neural network approach to sensitivity and uncertainty analysis. Reliability Engineering & System Safety 64:59-71 Rietman E (1994) Genesis redux: experiments creating artificial life. McGraw-Hill, New York Rodrigue J-P (1997) Parallel modelling and neural networks: an overview for transportation/land use systems. Transpn Res.-C. 5:259-271 Rosenberg JR, Halliday DM, Breeze P, Conway BA (1998) Identification of patterns of neuronal connectivity - partial spectra, partial coherence, and neuronal interactions. Journal of Neuroscience Methods 83:57-72 Rusakov DA, Stewart MG, Davies HA, Harrison E (1995) Population trends in the fine spatial reorganization of synaptic elements in forebrain regions of chicks 0.5 and 24 hours after passive avoidance training. Neuroscience 66:291-307 Saksida LM, Raymond SM, Touretzky DS (1997) Shaping robot behavior using principles from instrumental conditioning. Robotics and Autonomous Systems 22:231-249 Salmela P, Lehtokangas M, Saarinen J (1999) Neural network based digit recognition system for dialling in noisy environments. Information Sciences 121:171-199 Samejima K, Omori T (1999) Adaptive internal state space construction method for reinforcement learning of a real-world agent. Neural Networks 12:1143-1155 Schaal S (1999) Is imitation learning the route to humanoid robots? Trends in Cognitive Sciences 3:233-242 Schaap MG, Leij FJ (1998) Using neural networks to predict soil water retention and soil hydraulic conductivity. Soil & Tillage Research 47:37-42 Schenker B, Agarwal M (1997) Dynamic modelling using neural networks. International Journal of Systems Science 28:1285-1298 Schleiter IM, Borchardt D, Wagner R, Dapper T, Schmidt K-D, Schmidt H-H, Werner H (1999) Modelling water quality, bioindication and population dynamics on lotic ecosystems using neural networks. Ecological Modelling 120:271-286 Schoonhoven R Prijs VF, Frijns JHM (1997) Transmitter release in inner hair cell synapses: a model analysis of spontaneous and driven rate properties of cochlear nerve fibres. Hearing Research 113:247-260 Sette S, Boullart L, van Langenhove L (1998) Using genetic algorithms to design a control strategy of an industrial process. Control Engineering Practice 6:523-527 Seung HS (1998) Continuous attractors and oculomotor control. Neural Networks 11:1253-1258 Smith CS (1980) From art to science. Seventy-two objects illustrating the nature of discovery. The MIT Press, Cambridge, Massachusetts Smith RE, Cribbs HB (1997) Combined biological paradigms: a neural, genetics-based autonomous systems strategy. Robotics and Autonomous Systems 22:65-74 Starrenburg JG, van Luenen WTC, Oelen W, van Amerongen J (1996) Learning feedforward controller for a mobile robot vehicle. Control Eng. Practice 4:1221-1230 Steels L (1997) A selectionist mechanism for autonomous behaviour acquisition. Robotics and Autonomous Systems 20:117-131 Stoecker M, Reitboeck HJ, Eckhorn R (1996) A neural network for scene segmentation by temporal coding. Neurocomputing 11:123-134. Sumpter N, Bulpitt A (2000) Learning spatio-temporal patterns for predicting object behaviour. Image and Voice Computing 18:697-704 Talukder A, Casasent D (2001) Adaptive activation function neural net for face recognition. In: Proceedings of The IEEE Intl Joint Conf. on Neural Networks, Washington, D.C., USA, 2001. Thrun S, Mitchell TM (1995) Lifelong robot learning. Robotics and Autonomous Systems 15:2546 Timmermans AJM, Hulzebosch AA (1996) Computer vision system for on-line sorting of pot plants using an artificial neural network classifier. Computers and Electronics in Agriculture 15:41-55 Tipping E, Woof C, Rigg E, Harrison AF, Ineson P, Taylor K, Benham D, Poskitt J, Rowland AP, Bol R, Harkness DD (1999) Climatic influences on the leaching of dissolved organic matter from upland UK moorland soils, investigated by a field manipulation experiment. Environmental International 25:83-95 Treves A, Rolls E, Simmen M (1997) Time for retrieval in recurrent associative memories. Physica D 107:392-400 24
Tsaih R, Hsu Y, Lai CC (1998) Forecasting S & P 500 stock index futures with a hybrid AI system. Decision Support Systems 23:161-174 Tsodyks M (2005) Attractor neural networks and spatial maps in hippocampus. Neuron 48(2):168-169 Tyler L, Czarnecki CA (1999) A neural vision based controller for a robot footballer. In: Proceedings of The 7th IEE Int. Conference on Image Processing and its Applications, Manchester, UK, 1999 Von Wichert G (1998) Mobile robot localization using a self-organized visual environment representation. Robotics and Autonomous Systems 25:185-194 Von Wichert G (1999) Can robots learn to see? Control Eng. Practice 7:783-795 Walley WJ, Fontama VN (1998) Neural network predictors of average score per taxon and number of families at unpolluted river sites in Great Britain. Wat. Res. 32:613-622 Weiss M, Baret F (1999) Evaluation of canopy biophysical variable retrieval performances from the accumulation of large swath satellite data. Remote Sens. Environ. 70:293-306 Weng J, Chen S (1998) Vision-guided navigation using SHOSLIF. Neural Networks 11:15111529 Wolff JR, Laskawi R, Spatz WB, Missler M (1995) Structural dynamics of synapses and synaptic components. Behavioural Brain Research 66:13-20 Wong JC, McDonald KA, Palazoglu A (1998) Classification of process trends based on fuzzified symbolic representation and hidden Markov models. J. Proc. Cont. 8:395-408 Wong PM, Jang M, Cho S, Gedeon TD (2000) Multiple permeability predictions using an observational learning algorithm. Computers & Geosciences 26:907-913 Yang H-L (1997) A simple coupler to link expert systems with database systems. Expert Systems With Applications 12:179-188 Yeh I-C (1997) Application of neural networks to automatic soil pressure balance control for shield tunneling. Automation in Construction 5:421-426 Yun C-B, Bahng EY (2000) Substructural identification using neural networks. Computers and Structures 77:41-52 Zardeki A (1995) Fuzzy controllers in nuclear material accounting. Fuzzy Sets and Systems 74:73-79 Zhai Y, Thomasson JA, Boggess III JE, Sui R (2006) Soil texture classification with artificial neural networks operating on remote sensing data. Computers and Electronics in Agriculture 54(2):53-68 Zhang M, Fulcher J, Scofield RA (1997) Rainfall estimation using artificial neural network group. Neurocomputing 16:97-115 Zhao Y, Collins Jr EG (2005) Robust automatic parallel parking in tight spaces via fuzzy logic. Robotics and autonomous systems 51(2-3):111-127
25