The Feasibility of Ignoring Opponents in Multi-Player Games

Abstract

The application of AI in board games is an interesting area in the diverse field of Artificial Intelligence research. Board games allow for experimentation in an isolated fashion. Their clearly defined rules and the wide-ranging differences in complexity levels are ideal to test the performance of different AI approaches in various settings.

Board games with a large action space complexity prove difficult for AIs. This thesis introduces Agent Impact as a method to reduce this complexity and it investigates the impact of nine proposed agent properties have on the Agent Impact. This thesis further discusses whether these properties are indicative of an opponent’s opportunity to lower the player’s chances of winning, and if ignoring opponents based on these agent properties can improve player strength.

The experiments show that the performance of Agent Impact combined with the examined agent properties is strongly dependent on the domain. In Rolit ignoring opponents always leads to a lower performance, while in Blokus the performance differs per agent property. This thesis further concludes that Agent Impact has potential, but its application in the investigated domains does not improve the performance of MonteCarlo Tree Search when the algorithm it is given enough computing time.