Identifying Playstyles in Games with NEAT and Clustering

In the research of agents playing computer games, attempts have been made to create agents with particular playstyles. A typical approach adopted in previous studies is to define the reward functions so that the agent has the desired playstyle. However, based on this approach, it is difficult to identify other playstyles that have not been specified in advance. The purpose of this paper is to create agents with playstyles by combining a genetic algorithm called NeuroEvolution of Augmenting Topologies (NEAT) and clustering based on the logs of gameplay by agents. Here, individuals of NEAT are classified in some species by clustering according to the play log and crossed with the same species. We evaluated our method by an experiment with a roguelike game and observed that multiple playstyles were identified. Furthermore, by comparing the proposed method with the method using only NEAT, it was shown that the bias in the number of individuals among playstyles can be suppressed.