Research Report - 2001
Intelligent Systems
Academic Staff
Prof Tom Downs
A/Prof Janet Wiles
Dr Kevin Gates
Dr Helen Purchase
Dr Mark Schulz
Dr Gordon Wyeth
Dr Marcus Gallagher
Research Staff
Dr Paul Darwen
Research Students
Mr Ajantha Atukorale
Mr Robert Andrews
Mr Scott Bolland
Mr Brett Browning
Mr Phillip Chan
Mr Mark Chang
Ms Jean Chen
Mr Stuart Clarke
Mr David Collins
Dr David Cook
Mr Kristedjo Kurnianto
Mr Michael Neville
Mr Michael Norris
Mr Mark Pedersen
Mr Peter Stratton
Mr Ashley Tews
Mr Bradley Tonkes
Mr Michael Turner
Mr Ian Wood
Ms Peta Wyeth
Mrs Mehreen Ahmed
Ms Adelina Tang
Mr Daniel Johnson
Mr James Watson
Mr Kai Willadsen
Mr Andrew Janke
Mr Grant Larcombe
Mr Gregg Buskey
Mr Brad Byrne
Contact Details
Prof Tom Downs
Email: td@csee.uq.edu.au
Tel: 3365 3791
A/Prof Janet Wiles
Email: janetw@csee.uq.edu.au
Tel: 3365 2902
Dr Mark Schulz
Email: marks@csee.uq.edu.au
Tel: 3365 4136
Dr Helen Purchase
Email: hcp@csee.uq.edu.au
Tel: 3365 2697
Dr Gordon Wyeth
Email: wyeth@csee.uq.edu.au
Tel: 3365 3770
Dr Marcus Gallagher
Email: marcusg@csee.uq.edu.au
Tel: 3365
This group is interested in a variety of areas of research involving intelligent systems and computational intelligence. In many ways, intelligent systems research is at the interface of humans and computers. Generally speaking, the goals of intelligent systems research are:
To produce computer software and hardware that is capable of intelligent behaviour and/or problem solving
To model aspects of human intelligence using computational techniques, and
To investigate the interaction between humans, sensory information from the environment and computer systems.
Computational intelligence is concerned with the development of algorithms and problem-solving techniques in areas such as supervised and unsupervised learning, optimization, pattern recognition and function approximation. The application of this technology is often to problems that are difficult to solve with conventional computing methods, yet are accomplished easily by humans. In other cases, the aims are to relieve humans from tedious or hazardous tasks, as well as solving problems where exact solutions are not possible and effective approximate methods are desirable.
Artificial Neural Networks (ANNs) are computing devices designed to emulate some of the learning processes that are believed to take place in the brain. In the last decade or so ANNs have emerged as powerful computational techniques for identifying patterns in data and for predicting future events, and have been widely successful in solving many problems in engineering and information technology. ANNs are data driven, and typically require less knowledge about the problem domain, meaning less critical assumptions and approximations that have to be made. In recent years, statistics and computational learning theory have given this field a solid theoretical foundation and produced a number of powerful modelling techniques and algorithms. These include Support Vector Machines, the technique of boosting and other methods for combining machine learning models. Probabilistic modelling methods such as Bayesian networks and graphical models have also created recent interest in relation to machine learning and evolutionary computation.
Cognitive Science is a rapidly expanding, interdisciplinary field of study aimed at understanding the mental processes that underlie cognitive abilities. Its defining technique is to bring expertise gained from the disciplines of computer science, psychology, linguistics, neuroscience and philosophy to bear on a set of common questions:
What are the basic components of cognitive processes?
Are they subsumed by a common mental mechanism?
What is the relationship between the physical apparatus and cognition?
To answer these questions Cognitive Scientists engage in empirical and conceptual studies aimed at assessing their formal and computational models of various aspects of cognition. The sorts of areas investigated include the information-acquisition and information-processing mechanisms underlying cognitive abilities like perception, recognition, information storage and information retrieval, language acquisition, comprehension and production, concept acquisition, problem solving, and reasoning. Cognitive science is concerned with modelling and understanding aspects of human intelligence via computer simulation. Evolutionary algorithms, game playing, human memory, learning dynamics, language learning and cognitive neurophyschology.
Robotics is the embodiment of intelligence in computer hardware. Here, broader studies of biological brains are being undertaken to assist in the development of sensing and locomotion schemes to enable intelligent systems to interact with real world environments. Research into navigation in dynamic environments and emergent multiagent cooperation is conducted through practical applications such as the RoboRoos Robot Soccer Team. The research aims to address a number of mechanical, electronic as well as computer systems issues. Topics of major interest include vision and navigation, control and multi-robot coordination.
Research in Human-Computer Interaction focuses on the interface between humans and computing technology. The effectiveness of computers is fundamentally tied to the capacity in which people can control, manipulate and receive information from them. Issues surrounding interface architecture, adaptability, design, development and tools are of major interest, as is human-computer interaction in education, tutoring and training. Finally, defining and developing multimedia is of major importance in advancing the ability for natural human-computer communication.
Evolutionary Computation and Probability Modelling for Continuous Global Optimization
Many problems in science and engineering can be formulated as global optimization problems, often involving continuous numerical variables. The aim of this project is to develop innovative algorithms for solving such problems, by combining various techniques developed in the fields of Evolutionary Computation and Probabilistic Graphical models. These algorithms are significant because they uniquely learn the structure of a given problem solution space and exploit this structural knowledge to offer improved performance. This is achieved through probabilistic modelling techniques. The project will provide algorithms for solving difficult practical optimization problems, and advance knowledge on the theoretical and dynamical properties of these kinds of techniques.
Marcus Gallagher. An Empirical Investigation of the User-Parameters and Performance of Continuous PBIL Algorithms. Proceedings IEEE Workshop Neural Networks for Signal Processing, pp 702-710, 2000. IEEE Press, New York.
Applying Semiotic theory to Multimedia Texts
Initial semiotic classifications were made in an era of simple communication devices. As technology advances, it is appropriate to reconsider these classifications and to extend them to include the additional dimensions that are suggested by increasingly complex methods of communication. The aim of this project is to suggest a definition of multimedia that is based on an extension to a common classification of semiotic representational systems, in both the visual and aural modalities. In addition, the differing nature of communication devices are considered, providing a theoretical framework and terminology for consistent and appropriate discussion about multimedia texts and the devices that are used to transmit then. The underlying principle behind this work is the importance of separating the classification of the nature of the message to be communicated from the nature of the device, in order to focus on the importance of appropriate matching of text with technology. The resultant model of multimedia communication will be empirically evaluated for its effectiveness as a tool to support the development of multimedia software.
Purchase, H.C. and Naumann, D., A Semiotic Model of Multimedia: Theory and Evaluation, Design and Management of Multimedia Information Systems: Opportunities and Challenges, Rahman, S.M. (ed)., Idea Group Publishing, 2001 (to appear).
Purchase, H.C. and Naumann, D., The TOMUS Model of Multimedia: an empirical investigation, Proceedings of the World Conference on Educational Multimedia, Hypermedia and Telecommunications, Bordeau, J. and Heller, R. (eds), Association for the Advancement of Computing in Education, 2000.
Evaluating Grammar Formalisms for Free Word Order Natural Language Processing
The formalisms with which grammatical theories are expressed, and subsequently with which particular grammars are implemented, are core components of the grammar development process, and have been the subject of ongoing research and development. While the majority of formalisms in common use have settled on a number of key technologies, such as (typed) feature structures and unification-based processing, a common approach to the treatment of free word order has not yet been reached. In response to this, our department has taken a systematic approach to evaluating the different approaches to free word order within the major syntactic representational styles, namely dependency structure, phrase structure, and hybrid structures (which incorporate features of both dependency and phrase structure) in order to determine which, if any, holds a particular advantage for representing free word order languages. In particular, focus has been on assessing Tree Adjoining Grammar, Lexical Functional Grammar, and Mark Pedersen's own variant of Dependency Grammar in terms of their quality in use in the development of computational grammars of Hindi and English.
Pattern Classification Using a Mixture of Labelled and Unlabelled Data.
In many classification problems there is an abundance of data, but classification (ie labelling) of each data item is expensive. In such situations, it is desirable to be able to build a classifier that is trained on all available labelled data and which can also in some way make use of the unlabelled data in order to improve classification performance. This is usually done by first using some scheme to form the unlabelled into clusters of similar items and then drawing upon the labelled data to provide some fine-tuning of the clusters. In this project, a hierarchical structure has been developed which is based upon the above principle but which employs the hierarchy to assist in distinguishing between patterns of different classes that are close together in pattern space. The structure contains many rudimentary classifiers whose outputs are combined using the fuzzy integral.
A.S. Atukorale and T. Downs, "Using Labeled and Unlabeled Data for Training", Proceedings Sixth International Conference on Pattern Recognition and Information Processing (PRIP-2001), Vol. 1, pp. 195-199, Minsk, Belarus, 2001.
A.~S. Atukorale and P.N. Suganthan, "Hierarchical Overlapped Neural Gas Network with Application to Pattern Classification", Neurocomputing, vol. 35, no. 1-4, pp. 165-176--707, November 2000.
A.S. Atukorale, P.N. Suganthan and T. Downs, "On the Performance of the HONG Network for Pattern Classification", Proceedings International Joint Conference on Neural Networks, Vol. 2,
pp.285-290, Como, Italy, July 2000.
Neural Networks and Related Methods for Optimization
This work encompasses a variety of techniques. These include extensions to relatively conventional methods such as simulated annealing, extensions to evolutionary techniques and exploration of error surfaces. The extensions to simulated annealing are concerned with the further development of a new method, known as the demon algorithm, that has emerged from statistical physics and this work is aimed at producing new and efficient algorithms for the training of Boltzmann machines. The work on evolutionary techniques is largely concerned with extending recent generalizations of genetic algorithms and applying them to problems in telecommunications. The error surface exploration work is aimed at finding new ways of speeding up the learning process.
M Gallagher, “Fitness distance correlation of neural network error surfaces: a scalable, continuous optimisation problem”, European Conference on Machine Learning, Freiburg, Germany, 2001, (in press).
Computational Methods for Support Vector Machines
Support vector machines have recently emerged as a new and powerful technique for learning from data and in particular for solving classification and regression problems. They have a major advantage over neural networks in that they formulate the learning problem as a quadratic optimization problem whose error surface is free of local minima and has a unique global optimum. They perform by nonlinearly transforming a given problem into a high dimensional space (by means of a kernel function) and then constructing a linear discriminant (for classification) or a linear regression in the transformed space. Those data vectors nearest to the constructed line in the transformed space are called the support vectors. For some problems (especially noisy ones) the technique generates a large number of support vectors and this leads to computational problems. These problems, along with issues such as how best to select the kernel function and the efficient solution of the quadratic programming problem, are being investigated in this project.
T Downs, K E Gates and A Masters, “Simplifying support vector solutions” Journal of Machine Learning Research, 2001, (in press).
Applications of Boosting and Support Vector Machines in Clinical Medicine
Boosting is an ensemble method in which a number of learning systems are connected together in order to implement a classification or regression task. It has been shown that the manner in which the ensemble is constructed under boosting leads to improving generalization performance as learning proceeds. The learning process in support vector machines also ensures improvement of generalization performance during learning. In this project, these two new techniques are being applied to a number of problems in clinical medicine. These problems include the prediction of outcomes for (I) patients in intensive care, (ii) patients requiring hip prostheses and (iii) patients requiring kidney transplants.
H.D. Navone & T. Downs, “Variations on a Kernel-Adatron Theme”, Proceedings of the VII International Congress on Information Engineering, Buenos Aires, Argentina, 562-572, 2001.
Robotics
Robotics research provides an opportunity to implement in hardware the relatively abstract ideas and learning techniques that are being developed in the wider machine learning community. A major vehicle and test bed for this within the Intelligent Systems Group is robot soccer, which is described elsewhere in this report. This provides an opportunity to study problems of navigation and cooperation between robots in a dynamic and visually noisy environment. Underpinning this is a fundamental study of brain behaviour to assist with the development of new sensing and locomotion schemes that are essential for improved performance.
Wyeth G.F., Tews A. and Browning B. UQ RoboRoos: Kicking on to 2000. RoboCup-2000: Robot Soccer World Cup IV. Lecture Notes in Artificial Intelligence 2019. Springer Verlag, Berlin, 2001
Chang M., Browning B. and Wyeth G.F ViperRoos 2000. RoboCup-2000: Robot Soccer World Cup IV. Lecture Notes in Artificial Intelligence 2019. Springer Verlag, Berlin, 2001.
Tews A. and Wyeth G.F. MAPS: A System for Multi-Agent Coordination. Advanced Robotics, VSP / Robotics Society of Japan, Volume 14 (1), pp. 37-50.
Collins D. and Wyeth G.F. Fast and Accurate Mobile Robot Control Using a Cerebellar Model in a Sensory Delayed Environment, International Conference on Robotics and Systems (IROS 2000), pp.233-238.
Tews A. and Wyeth G.F. Thinking as One: Coordination of Multiple Mobile Robots by Shared Representations, International Conference on Robotics and Systems, pp.1391-1396, 2000.
Wyeth G.F. and Brown B. Robust Adaptive Vision for Robot Soccer, Mechatronics and Machine Vision in Practice, ed. John Billingsley, Research Studies Press, pp 41 – 48, 2000.
Collins D. and Wyeth G.F. Utilising a Cerebellar Model for Mobile Robot Control in a Delayed Sensory Environment, Accepted to the Proceedings of the Sixth International Conference on Simulation of Adaptive Behaviour, Paris, 2000.
Wyeth G.F., Buskey G. and Roberts J. Flight Control Using an Artificial Neural Network, Proceedings of the Australian Conference on Robotics and Automation, Melbourne, 2000..
Wyeth G.F., Kennedy J. and Lillywhite J. Distributed Digital Control of a Robot Arm, Proceedings of the Australian Conference on Robotics and Automation, Melbourne, 2000.
Collins D. and Wyeth G.F. Overcoming the Effects of Sensory Delay By Using a Cerebellar Model, Proceedings of the Sixth Pacific Rim International Conference on Artificial Intelligence, Melbourne,
p 554, 2000.
Solving Translational Invariance in Neural Networks using Attention and Memory
In this project a neural model of sensory attention is being developed using expectation as a basis. A major aspect of this work is in attempting to explain how the brain is able to extract low level features from visual input and then combine then into higher level "super features". A local learning rules has been found that is able to perform both the functions of feature extraction and feature recombination in a biologically plausible fashion. An artificial neural network has been trained using this learning rule and is able to detect features and combine them into super features. It can be shown that the trained network treats the super features as single entities. This work is now being extended to deal with more general and complex inputs.
Machine Recognition of Mandarin Chinese Speech
This project aims to develop a system capable of recognizing Mandarin Chinese speech using the technique of keyword-spotting. Current systems for machine recognition of speech are incapable of recognizing continuous speech from arbitrary speakers and keyword spotting is an important and useful intermediate technique. A set of keywords is selected for a given application area and the recognizer is trained to pick out these words when they are uttered in the midst of continuous speech. In this project a keyword spotter for Mandarin Chinese is being developed using techniques that have proven successful with western languages (e.g., hidden Markov models). It is expected that various novel refinements to these techniques will be necessary in order to achieve the performance levels required for practical applications.
Evolutionary Computation for Scheduling and Game-Like Tasks
Board games like Backgammon provide a useful testbed for methods in artificial intelligence. While Chess computers work brute-force look-ahead methods, those do not work on Backgammon because of the randomness and uncertainty in Backgammon's dice. Real-world problems (such as schedule optimization) mean facing the uncertainties of the future. So artificial intelligence methods that work on Backgammon can provide clues on how to face realistic, dynamic rescheduling problems of the kind found in industry.
``Why Co-Evolution beats Temporal Difference learning at Backgammon for a linear architecture, but not a non-linear architecture'' by Paul J. Darwen. Congress on Evolutionary Computation (CEC'2001), pages 1003-1010.
``Genetic Algorithms and Risk Assessment to Maximize NPV With Robust Open-Pit Scheduling'' by Paul J. Darwen. Fourth Biennial Conference on Strategic Mine Planning, pages 29-34, 2001.
``Computationally Intensive and Noisy Tasks: Co-Evolutionary Learning and Temporal Difference Learning on Backgammon'' by Paul J. Darwen. Congress on Evolutionary Computation (CEC-2000),
pages 872-879, 2000.
``Unobtrusive Workstation Farming Without Inconveniencing Owners: Learning Backgammon with a Genetic Algorithm'' by Paul J. Darwen. IEEE International Workshop on Cluster Computing, pages 303-311, 1999.
``Co-Evolutionary learning on noisy tasks'' by Paul Darwen and Jordan Pollack, Congress on Evolutionary Computation (CEC-99), pages 1724-1731, 1999.
Emergence of Cooperation in Complex Systems
While economics and evolutionary biology has a popular image of a dog-eat-dog world, of nature red in tooth and claw, the surprising reality is that cooperative relationships are common, even without any central authority. The small fish that clean the teeth of sharks are a well-known example. How such cooperation first got started is usually a mystery: trying to clean a shark's teeth usually results in a shark's lunch. This work uses artificial intelligence agents to explore the mechanisms for how such mutual cooperation emerges without a central authority. This has applications in economics and biology, as well as providing insights into such current issues as George W. Bush's nuclear missile defence.
``Why More Choices Causes Less Cooperation in Iterated Prisoner's Dilemma'' by Paul J. Darwen and Xin Yao. Congress on Evolutionary Computation (CEC'2001), pages 987-994, 2001.
``Does extra genetic diversity maintain escalation in a co-evolutionary arms race'' by Paul Darwen and Xin Yao, International Journal of Knowledge-Based Intelligent Engineering Systems, volume 4, number 3, pages 191-200, 2000.
``Genetic algorithms and evolutionary games'' by Xin Yao and Paul Darwen. Pages 325-347, in "Commerce, Complexity and Evolution", Cambridge University Press, 2000.
``Initial Population Diversity as a Wild Goose Chase: Phenotypic Effects Sometimes Dominate'' by Paul J. Darwen and Xin Yao. Australia-Japan Joint Workshop on Intelligent and Evolutionary Systems, 1999.
``How important is your reputation in a multi-agent environment'' by Xin Yao and Paul J. Darwen. IEEE Conference on Systems, Man, and Cybernetics, pages 575-580, 1999.
<%-- end content --%><%@ include file="/include/footer.inc.jsp" %>