Achieving dynamic AI difficulty by using reinforcement learning and fuzzy logic skill metering

Abstract

The most important functional requirement of a video game is to provide entertainment. Players can always be entertained if they face a challenge according to their own level of skills. While different players owned different levels of skills, the game should not be very hard or very easy for different players with varying levels of skills. Artificial intelligence provides a number of methods to adaptively tune the playing agents in the game with respect to human players. In this paper we propose a method in which reinforcement learning is used to make learning agents as well as a dynamic AI difficulty system based on fuzzy logic. To validate our approach we applied our method to an action tower defense game to show how a player can have better experiences while playing against agents who can learn to adapt their behavior to the skill level of the player.