Real-time neuroevolution to imitate a game player

Abstract

In this paper, we present an algorithm to imitate a game player's play patterns using a real-time neuroevolution (NE); the examples of the patterns can be moving and firing units. Our algorithm to learn and imitate is possible to be executed during gameplay. To test effectiveness of our algorithm, we made an application similar to the Starcraft (TM). By using our method, a game player can avoids tediously repeating labors to control units. Moreover, applying this to enemy agents makes it possible to play more difficult and exciting games. From experimental results, we found that agents' ability to imitate a game player's unit control patterns could make human-like agents, and also we found that adaptive game AIs, especially the real-time NE, are efficient in such imitation problems.