Psygnosis AI Seminar

19th - 21st November 1996

University of Leeds

This document is available online as http://www.comp.leeds.ac.uk/ajh/psygnosis.html.


Purpose

Artificial Intelligence is perhaps one of the most misconceived areas of Computer Science, largely due to the density of buzz-phrases associated with the field and the trumped-up claims that the popular science media make about the potential of AI.

The aim of this 3-day seminar is to cover some of the most popular and widely-used AI techniques, describe what they do, how they work and where they can be used. The other main goal is to explain the jargon and dispell the myths. In this way we hope to be able to focus on what is practicable and what is infeasible.

For each topic there will be an `expert' from the School of Computer Studies at the University of Leeds who is either teaching or researching (or both) in the field, and who will give a talk on the key issues and in some cases give demonstrations of applications they have developed. Following this there will be discussions to try to draw out any potential uses within Psygnosis and to stimulate any ideas people have. The talks and discussions will be `round the table' and the atmosphere will be as informal as possible.

To round off the event we have invited Dave Cliff from Sussex University to give a talk on Artificial Life (a-life) and how it can be used in the computer entertainment industry. Dave is the UK's leading researcher into a-life and his talk promises to be really rather excellent.


Dates and Venue

The seminar is for 3 full days; 19th - 21st November 1996 (Tuesday - Thursday), and will take place on the University of Leeds Campus. Lunch and coffee breaks will be provided and there is a budget for some evening entertainment.

The venue for the first day has been changed. This will now be in the Merrion Hotel.


Accomodation

Overnight accomodation has been arranged in the Merrion Hotel in Leeds City Centre. Delegates are booked in for Monday, Tuesday and Wednesday nights.


Timetable

The following timetable is provisional. Talks are expected to last about an hour and all times are approximate, depending on how much there is to discuss on each topic.

Day 1 (19th Nov) - Merrion Hotel
9.00General WelcomeSam Brown
9.10Introduction to AIDavid Hogg
10.30Coffee
10.45Expert SystemsDave Ranyard
1.00Lunch
2.00Genetic AlgorithmsTony Heap
3.00Coffee
3.15Genetic Algorithms (continued)Tony Heap
5.00Close
7.30Beer

Day 2 (20th Nov) - Beech Grove Room
9.00Natural Language ProcessingGavin Churcher
11.15Coffee
11.30Fuzzy LogicDave Ranyard
1.15Lunch
2.15Neural NetsNeil Johnson
3.15Coffee
3.30Neural Nets (continued)Neil Johnson
5.00Close
7.30More Beer

Day 3 (21st Nov) - Active Learning Lab
9.00Hidden Markov ModelsRoger Boyle
10.00Coffee
10.15Hidden Markov Models (continued)Roger Boyle
12.00Lunch
1.00Artificial Life in Entertainment SoftwareDave Cliff
3.15Coffee
3.30Wrap UpSam Brown
4.00Close
7.30Stalwarts' Beer Session


Abstracts

Click on the speaker's name to see their biography.

Introduction to AI (David Hogg)

The field of Artificial Intelligence (AI) has spawned many different ways for simulating all aspects of human intelligence in machines - usually general purpose computers. These aspects include perceptual abilities such as vision and audition, reasoning capabilities, emotional behaviour, memory, and learning. The talk will briefly introduce the full range of approaches, including some that are often presented as radical alternatives to Good-Old-Fashioned-AI (GOFAI for short).

In particular, we will review classical approaches based on various flavours of logic, constraint satisfaction, pattern recognition, neural networks, evolutionary approaches, and artificial life. The aim is to provide an overall context for the presentation of most of these topics in greater depth during the seminar.

Examples of working systems based on these different approaches to the construction of intelligent systems will be used by way of illustration throughout.

Expert Systems (Dave Ranyard)

An expert system simulates human reasoning in a specific knowledge domain. For example, diagnosing a disease or configuring a computer system. In many situations traditional computing techniques or models are inefficient or infeasible. We cannot model the human body and the effects of an infective agent such as the common cold may have on it. However, we can define heuristics: rules of thumb that provide a core knowledge which can be applied to a problem domain.

This overview will cover the basic concepts behind expert systems, in particular:

We will look at some ways expert systems may be applicable to the development of computer games as well as discussing exactly what their limitations are.

Genetic Algorithms (Tony Heap)

Genetic Algorithms are essentially Darwin's theory of Evolution (survival of the fittest) as applied to computer-aided problem solving. A `population' of genetically-coded `solutions' are tested on a `problem' and are each given a score or `fitness'. The solution genes are then bred using concepts such as mating and mutation to produce a new and hopefully better population of solutions. Over many iterations or `genetarions', one or more optimal solutions are eventually found.

In this session we will be discussing:

There will also be a discussion on how this concept extends to `Genetic Programming' - instead of writing a program to perform a task, it may be possible to let it evolve.

Natural Language Processing (Gavin Churcher)

Natural Language Processing is a large field covering many of the computational aspects of language. Back in the 1950s when the future potential of computers was beginning to be realised, people believed that translating between different languages was attainable within a few years. At the time it seemed that computers could do almost anything given their speed and large memory (large at the time!). Initial experiments between Russian and English showed promise. But then the scientists hit problems. Translation was much more subtle than they had ever thought. Their mistake was to liken the human mind to a simple computer. Nowadays, we realise that the problems humans cope with everyday, from reasoning to having a conversation, are very difficult for computers to do.

Natural Language Processing covers many aspects of language used by humans. In my talk I shall be giving an overview of the following areas:

Fuzzy Logic (Dave Ranyard)

Fuzzy sets are a relatively new method of handling uncertain data and knowledge. They extend the concept of traditional sets to include the notion of partial membership to a set. For example, if we define a set of fast-cars as cars which go at least 150mph, fuzzy sets allows us to define a partial membership to this set so that a car which goes 140mph can be classed as "quite fast". Fuzzy logic operators have been defined on top of fuzzy set theory which correspond to traditional Boolean logic operators. Combining these techniques provides a simple method of dealing with uncertainty by linking them to linguistic concepts such as: tall, long, old and so on. By doing this fuzzy approaches are deemed to emulate the human thought process more closely than more traditional, rigid computational processes. This discussion of fuzzy sets will provide an overview of the concepts of fuzzy sets and fuzzy logic and how they can be used. They will be put into context by comparing them briefly with other uncertainty management techniques such as probabilities and Dempster-Shafer theory of evidence. Some applications will be discussed as well as the limitations and criticisms of the fuzzy approach.

Neural Nets (Neil Johnson)

Although the development of the digital computer has seen rapid increases in computing power, there are many tasks routinely performed by humans (such as pattern recognition) which have proved extremely difficult and computationally expensive to achieve. Research into the structure and dynamics of the human brain suggests that the basis of biological computation is a small number of serial processing steps within a huge, massively interconnected, network of relatively simple processing units.

Research into (Artificial) Neural Networks has been motivated by the recognition of this different approach to computation, resulting in the development of many different artificial neural structures, all based around the idea of large interconnected networks of relatively simple processing units where information is stored by virtue of the strengths of the interconnections.

This tutorial will cover the basic ideas behind neural networks, highlighting the types of tasks for which they can be used, the benefits of the neural network approach, and some of the problems involved in their use. Finally, as an example, some current research with possible applications in the development of computer games will be discussed and demonstrated.

Hidden Markov Models (Roger Boyle)

Hidden Markov Models are a mechanism for determining the behaviour of sequential systems by observing them through a (probably noisy) filter.

In this tutorial, we will explore the need for hidden models of sequential real world events, and concentrate on the Markov approach. The theory will be introduced via a simple example, and then delivered in detail. A `real world' application will then be explained that generalises the approach to cope with some of its shortcomings.

[God willing] a software demonstration will also be given.

Suitable references and supporting documentation will be available.

Artificial Life in Entertainment Software (Dave Cliff)

Artificial Life involves a variety of biologically inspired computing techniques which ease the development of complex adaptive systems. Typically, the complexity in the system is an "emergent phenomena", resulting from interactions between a number of comparatively simple components. In this talk I'll give a brief overview of the major artificial life (a-life) techniques, concentrating on the integration of these techniques into autonomous software agents which can perceive and act within virtual environments, possibly interacting with other synthetic agents or with human users. Following this, I'll discuss the current market penetration of a-life technologies in the computer games industry, and sketch some potential future applications.


Biographies

Click on a speaker's name to see their home page.

Dr Roger Boyle

roger@comp.leeds.ac.uk

Roger Boyle has a BA and DPhil in mathematics from the Universty of York. After a year in the civil service he joined the systems staff at the School of COmputer Studies in 1979, and became a lecturer in 1984, with interests in comms and PR. Since then he has developed interests in low level vision and neural networks - application areas of special interest are OCR and OMR. He is now a senior lecturer in AI and head of teaching.

Roger is 1.99m tall and weighs about 85kg. He was once a competitive swimmer but now restricts himself to casual exercise. He likes beer very much.

Mr Gavin Churcher

gavin@comp.leeds.ac.uk

Gavin Churcher is currently studying for a PhD with the Natural Language Processing Group at Leeds University. He has been working on building language models suitable for speech recognition and has recently moved into the field of dialogue management systems as a means to improve both the performance of speech recognition and to produce more natural ways of interacting with computers.

Dr Dave Cliff

davec@cogs.susx.ac.uk

Dave Cliff is a Lecturer in Computer Science and Artificial Intelligence at the University of Sussex. He has been involved in Artificial Life research since 1988, has published about 50 papers on his work, and has performed consultancy work for companies including Warner Interactive Entertainment, Inscape, and Hewlett-Packard. Over the last two years he has been actively involved in developing entertainment applications of artificial life: he attended the E3 Trade Show in Los Angeles in May, and has co-authored a paper on artificial life entertainment software to be presented at the First International Conference on Autonomous Agents, California, February 1997.

Mr Tony Heap

ajh@comp.leeds.ac.uk

Tony Heap is currently studying towards his PhD in the Computer Vision Group. He is investigating robust, model-based methods for tracking deformable objects. He is particularly interested in tracking human hands with a view to recognising gestures. Possible applications are computer-assisted sign language and human-computer interaction.

Prof David Hogg

dch@comp.leeds.ac.uk

David Hogg is Head of the School of Computer Studies at the University of Leeds. His research is broadly in the area of computer vision with emphasis on the analysis of motion in relation to non-rigid objects and generic shape classes (e.g. the human body).

David currently teaches an Introduction to Artificial Intelligence, available to first year students from any department. In the 1995/96 Session, around 250 students enrolled on this module.

Mr Neil Johnson

neilj@comp.leeds.ac.uk

Neil Johnson is currently studying for his PhD in the Computer Vision Group. He is researching the modelling of `object behaviours' using detailed, learnt statistical models. The techniques being developed will allow models of characteristic object behaviours to be learnt from the continuous observation of long image sequences.

Mr Dave Ranyard

dcr@comp.leeds.ac.uk

Dave Ranyard's interests centre on the development of RODOS, a decision support system for response to nuclear accidents. RODOS is a large software project being built by a consortium of European institutes and universities. It is being designed to work on a European wide scale, supporting decisions from the moment an accident threatens to months and even years later.

Dave is also lecturer in an Expert Systems course. In charge of a course with approximately 130 undergraduate and MSc students; responsible for all aspects of the course.


Stuff


This page written by Tony Heap - ajh@comp.leeds.ac.uk