Learning board game rules by observing game play, a comparison of symbolic and non-symbolic AI

Abstract

Usually the focus of Artificial Intelligence (AI) game research is on learning strategies for specific games. This thesis reversed this focus by looking for methods capable of learning game rules in general. The goal is to learn the rules of the simple board game Tic Tac Toe by observing played games in a way that can be used for more complicated games. This will be done by testing the performance of the symbolic AI algorithm ProbFOIL+ and two non-symbolic AI algorithms, the kNearest-Neighbor (KNN) algorithm and a Decision Tree (DT) algorithm. Both ProbFOIL+ and DT succeed in learning the win rule of Tic Tac Toe, but no algorithm succeeds in learning a more complex rule. In this case the Recall of ProbFOIL+ is very low whereas KNN and DT both overfit on the data. The strengths and weaknesses of both symbolic and non-symbolic AI seems to supplement each other and therefore it is suggested that future work focuses on combining these two.