This repository contains code used for Bomberman game with reinforcement learning and is a project of Gradient AI Science Club.
- Gradient Science Club - gradient.eti.pg.gda.pl
- Project website - RL vs Bomberman
In this project, we compare various agents and their behaviour in different environments of the classic console game Bomberman and present our analyse of exploration methods. The aim of Bomberman is to eliminate all of oponents from the map by placing bombs in their range and collecting additional bonuses. We employ multilayer convolutional neural networks (CNNs) and reinforcement learning (RL) to train agents, which achieve challenging results in the gameplay. Our work is based on model-free algorithms: simple Q-learning, fixed Q-targets, double DQN and dueling DQN. Furthermore, we present our own game environment in confrontation with Pommerman, popular AI competition. Finally, we discuss existing exploration methods such as Max-Boltzmann, Random-Walk, Greedy, E-Greedy and explain systems of rewards implemented in agents as well as tactics they have learned.
Our own environment "Bomberman - Brics" is available at GitHub repository: Environment Bomberman - Brics.
Project was presented on several events and conferences such as FOKA 2019 and GUT Day for AI.
- Pommerman competition - pommerman.com
- Bartłomiej Borzyszkowski
- Marcin Świniarski
- Hubert Skrzypczak
- Grzegorz Opoka
- Maksymilian Bubula