Co-evolving real-time strategy game players.

摘要

I aim to develop superior real-time strategy (RTS) game players. Most existing Artificially Intelligent (AI) players for RTS games use a variety of hand-coded approaches including expert systems, scripts and decision trees to play the game. These traditional approaches suffer from the knowledge acquisition bottleneck common to many AI systems. To improve on this, I investigate the use of genetic algorithms to co-evolve AI players for real-time strategy games, overcoming the knowledge acquisition bottleneck while allowing for more complicated players. I tackle four significant problems. First, existing commercial RTS games are not suitable for research, therefore I develop a new real-time strategy game as a platform for my work. LagoonCraft, the game I developed, consists of many interoperating systems making up over 300,000 lines of code. Second, I formulate the problem of playing an RTS game as solving a sequence of spatially resolved resource allocation problems. Third, I create a new representation that takes advantage of this problem formulation in order to provide an encoding of RTS game-playing strategy amiable to evolution. Last, I develop a new method based on co-evolution that generates competent RTS game-playing strategies that routinely beat human opponents.