"Evaluating the Performance of AI Techniques in the Domain of Computer Games"

Supervised by A.Sharkey

Abstract

This project concerns the development of an extensible software system that tests and compares three different types of artificial intelligence algorithms for their suitability as the control systems of agents in computer games. Game intelligence normally revolves around the use of either fuzzy state machines or more ad-hoc solutions. Three types of system are introduced, goal subsumption architecture, fuzzy state machines and neural networks, each of which is designed to deal with a homogeneous set of problems associated with games.

The project is split into two main sections. The first section deals with the development of a software framework in which the three systems can be tested. The second details analysis of each system, by looking at quantitative measurements of each control systems performance in terms of complexity, and examines their behavioural performance with some discussion of how the various behaviours are achieved.

Development covers the use of 'fuzzy' decision criteria in both the goal subsumption and fuzzy state approaches, showing how it allows an agent's behaviour to be unpredictable, yet still suitable for the purpose.

Each system has a set of benefits and problems, each of which has a trade off in terms of complexity. This project highlights the problems found when using neural networks. Problems detailed are caused by the used of genetic algorithms and the fitness function each leads to comparatively high development time and variable results. Specifically, defining a fitness function involved developing a large part of the solution to a problem.