This project implements an autonomous agent capable of playing the classic game Bomberman. The primary agent, Deep Quapsel, utilizes Deep Q-Learning (DQN) to navigate the grid, collect coins, and strategically eliminate opponents.
The framework simulates a full game environment including destructible crates, power-ups, and hostile adversaries.
- Deep Quapsel: The main RL-agent trained with a Deep Q-Network and Experience Replay.
- Rule-Based Agent: A heuristic agent used as a baseline and training adversary.
- Coin Collector: A simple agent focused solely on pathfinding and gathering loot.
- Clone the repo
git clone https://github.com/libaum/bomberman-rl.git
- Install dependencies
pip install -r requirements.txt
To watch the agent play against reference agents, run:
python main.py play --agents deep_quapsel rule_based_agent coin_collector_agentTo train the agent (disable the GUI for faster performance):
python main.py train --agents deep_quapsel --no-guiYou can analyze the agent's training progress and decision-making metrics using the provided Jupyter Notebook:
- Open
agent_code/deep_quapsel/train_eval.ipynbto visualize loss convergence and reward statistics.