Deep Reinforcement Learning in Serious Games: Analysis and Design of Deep Neural Network Architectures

Serious games present a noteworthy research area for artificial intelligence, where automated adaptation and reasonable NPC behaviour present essential challenges. Deep reinforcement learning has already been successfully applied to game-playing. We aim to expand and improve the application of deep learning methods in SGs through investigating their architectural properties and respective application scenarios. In this paper, we examine promising architectures and conduct first experiments concerning CNN design and analysis for game-playing. Although precise statements about the applicability of different architectures are not yet possible, our findings allow for concluding some general recommendations for the choice of DL architectures in different scenarios. Furthermore, we point out promising prospects for further research.