The popular and well known classical computer game Bomberman is undoubtedly one of the best-suited environments to explore the realm of reinforcement learning and to test cutting edge machine learning models. Bomberman is certainly not unknown to the community and has al- ready been studied in terms of reinforcement learning. Not only does this game consist of a massive amount of possible discrete states, but also demands strategic behaviour in competition with multiple opponents. This is a difficult task to learn and to solve for a ordinary computer and there is no one-size-fits-all approach to this problem. However, recent developments in the on going research in reinforcement learning has shown promising results and brought up many remarkable strategies. Some of which we want to take a closer look at. In this group project report we want to present the elaboration of our investigations to rein- forcement learning methods as well as techniques to train a computer to play Bomberman in competition with AI agents, trained by other students.
Read more here