Towards adaptive online RTS AI with NEAT

Abstract

Real Time Strategy (RTS) games are interesting from an Artificial Intelligence (AI) point of view because they involve a huge range of decision making from local tactical decisions to broad strategic considerations, all of which occur on a densely populated and fiercely contested map. However, most RTS AI used in commercial RTS games are predictable and can be exploited by expert players. Adaptive or evolutionary AI techniques offer the potential to create challenging AI opponents. Neural Evolution of Augmenting Technologies (NEAT) is a hybrid approach that applies Genetic Algorithm (GA) techniques to increase the efficiency of learning neural nets. This work presents an application of NEAT to RTS AI. It does so through a set of experiments in a realistic RTS environment. The results of the experiments show that NEAT can produce satisfactory RTS agents, and can also create agents capable of displaying complex in-game adaptive behavior. The results are significant because they show that NEAT can be used to evolve sophisticated RTS AI opponents without significant designer input or expertise, and without extensive databases of existing games.