For several years there has not only been a lot of interest in enhancing the user experience within video games but also in the development of algorithms to allow AI agents to better assist humans. It is felt that if it were possible to cluster and imitate playstyles, this would lead to an enhanced human AI interaction. The main aim of this thesis is to lay out a three-step process to allow the encoding, clustering, and imitating of playstyles.
To encode the gameplays into a representation space, the STDIM-VAE architecture has been proposed. This architecture was then compared to both its individual components. Further, a more in-depth analysis of the STDIM-VAE learnt representation space was conducted and it was shown that the architecture was able to encode playstyle relevant features across the three games investigated.
Clustering was then attempted using a novel method, namely, Reference-Based Clustering. The results of which were then compared to clustering within the other encoder type representation spaces as well as clustering on extracted gameplay variables. It has been shown that the combination of the STDIM-VAE and Reference-Based Clustering techniques allows good quality clustering across all considered games.
Finally, imitation of three different playstyles was explored within a stealth game. A new technique called DTWI was used to imitate across a range of altered levels, to investigate the ability to correctly imitate the playstyle. It has been shown that it was possible to correctly imitate playstyles across many level variations.
Although the majority of this work has focused on use cases within the video game domain, it was expected that the methods created would have use cases beyond this application area. To this end, the STDIM-VAE was used in the generation of representations from which model design novelty was evaluated.
