The predictability of behaviour in Connect 4

Keywords
Loading...
Thumbnail Image
Issue Date
2023-01-27
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
The ability to predict behaviour can boost the development and safe implementation of A.I.. However, the ability to predict behaviour is difficult task to accomplish. Therefore, the game Connect 4 was taken as a simplified environment in which predictions were made about what move players would play. A computer system was created that was able to play Connect 4 on different levels using the Monte Carlo Tree Search algorithm. Neural networks were provided with the task to predict which move the computer would play given the current game-state after the networks were trained on already collected training data. The experiment included a part where additional data was provided to the networks and the accuracies of the predictions were measured to monitor how proficient the networks would be at adapting to players from a different level. The initial accuracies could reach 50% depending on the network and play level. By providing data of games played on a different level than the collected training data it was possible to monitor the changes of accuracies of networks. The biggest change in accuracy measured was an improvement of 31,6% compared to the initial accuracy with solely the original training data.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen