Logo

CIRG - Research  -  Games 



[Bullet] Home
About
News
[Bullet]

- NN
- DM
- SI
- EC
- MAS
- AIS
- IA
- Bioinf
- Games
- Opt
- FA
- Industry

Publications
People
Resources
Links
Contact Us

OVERVIEW

This focus area applies various computational intelligence approaches to the wide domain of games. Systems are developed for the effective playing of a variety of competitive games against humans and other computer-based systems.

Currently, the specific research focus is on the use of coevolutionary strategies to train game playing agents, as well as novel approaches for analysing the behaviour of these agents.

ACTIVE MEMBERS

List the current members actively doing research in this focus area. [ Show ]

ALUMNI MEMBERS

M Smit

M.Sc Started in 2007

W Duminy

M.Sc Completed in 2007

L Messerschmidt

M.Sc Completed

N Franken

PhD Started in 2005
M.Sc Completed in 2004
Hons-B.Sc Completed in 2002

GROUP PUBLICATIONS

List publications of this research focus area. [ Show ]

MEMBER PROFILE



 Name:

 Willem Duminy

Portrait photo

 E-mail:

 wduminy@mweb.co.za

 Group(s):

 Evolutionary Computation
 Games

 

 Degree specific information: M.Sc

 Title:

 An evolutionary approach to learn the components of evaluation functions

 Abstract:

The three topics: evolutionary computing, machine learning and perfect information games combine in an experiment that investigates a method to increase the skill of a game-playing agent from zero game knowledge. The research entails the definition of a learning framework that makes use of genetic- and evolutionary- techniques to improve the performance two-person perfect-information game agents. Initially, the agent is equipped with the game rules and the ability to "percieve" elementary properties of a game state. A example of such a property might be "how many of my kings are on row four of the board". The properties are combined using a set of primitive operations to form features that are used in the evaluation function. The work of the learning agent is to discover these features and then determine the weights assigned to the features for optimal playing performance. The idea is to start with no game knowledge and through repeated application of the discovery step and weight optimisation step, the agent will gradually improve. The features and the weights can be seen as two genes that adapt and influence each other during the learning process. Individuals with strong features cooperate in a co-evolutionary fashion with those with strong weights. Competitive co-evolution ensures the improvement of a single gene: playing agents against each other to improve weights; or eliminating weaker feature sets.

 Supervisor / Co-Supervisor:

 AP Engelbrecht

 Thesis:

 Not available for download yet.




You are visitor #7306
Contact webmaster
Back to top

QualNet Network Simulator University Program Valid XHTML 1.0! Valid CSS!


Computational Intelligence Research Group
University of Pretoria
Copyright © 2017