Intelligent Bomberman with reinforcement learning

Bomberman is a strategical, maze-based game where the players defeat their enemies by placing a multi-direction count-down bomb that would explode and destroy obstacles and other players. In this paper, a simplified version of Bomberman is implemented in Java, where different controlled agents are placed. Each agent represents one of five reinforcement learning methods: Q-Learning, Sarsa, Double Q-Learning, and Deep Q Neural Network with two state representations: 5-tiles information and complete information. Then, we investigate whether a specific reinforcement learning method can successfully learn to play Bomberman efficiently by evaluating them with ad-hoc agents and finally against each other. The configuration of 5-tiles information with Sarsa archives the best overall quantitative results.