ss24_ReinforcementLearning_BomberMan

Repository for the implementation of a reinforcement learning agent for the game of Bomber Man for the subject on Machine Learning Essentials at Heidelberg University.

Bomberman Reinforcement Learning Agent

Project Description

This project implements two reinforcement learning agents for the classic Bomberman game. The agent is designed to navigate the game board, collect coins, and compete against other agents. The goal is to develop an intelligent agent that can adapt to different game scenarios (specified in settings.py) and outperform its opponents.

Table of Contents

• Agent Overview
• Callbacks
• Command Line Instructions
• Requirements
• Hardware Requirements
• Installation
• Usage
• Contributing
• License

Agent Overview

The agent operates in a discrete step environment where it can perform actions such as moving, dropping bombs, or waiting. The agent learns through reinforcement learning techniques, adapting its strategy based on the game state and feedback from the environment.

Our first agent (deep q-learning) presented for our project can be found in branch qdeepleaning -> agent_code/DALBomber/ Our second agent (q-learning) presented four our project can be found in branch qlearning -> agent_code/my_agent/

The Q-table of our q-learning agents can be found on: https://heibox.uni-heidelberg.de/d/57a78fe6abda48378ec6/. Additional files to reproduce graphs as well.

Key Features

• Efficient navigation to collect coins.
• Bomb usage to destroy crates and opponents.
• Adaptability to different game scenarios and opponents.

Callbacks

The agent interacts with the game environment through a series of callbacks. These callbacks handle the setup, action decisions, and event logging during the game. Key functions include:

• setup(self): Initializes the agent before the first round.
• act(self, game_state): Determines the best action based on the current game state.
• state_to_features(self, game_state): Creates a feature vector for our state representation. Different implementations have been discussed in our final report.

Event Logging

During training, various events are logged to provide feedback for the agent's learning process. Events include actions taken, coins collected, and opponents eliminated.

Command Line Instructions

To run the project, use the following command line instructions:

# To play the game
python main.py play

# To run in training mode
python main.py my_agent --train 1 --n-rounds 10000

# To save a replay of the game
python main.py replay <stored-replay>

# To skip frames for faster execution
python main.py --skip-frames

# To run without GUI for efficient training
python main.py --no-gui

Requirements

    Python 3.7 or higher
    Required libraries specified in requirements.txt

Hardware Requirements

    Processor: AMD Ryzen 5 2600 or equivalent
    RAM: Minimum 8 GB
    GPU: Recommended for deep learning (NVIDIA preferred)

Installation

    Clone the repository:

    git clone https://github.com/anacarsi/ss24_ReinforcementLearning_BomberMan.git
    cd bomberman-rl-agent

    Install the required libraries:

    pip install -r requirements.txt

Usage

After installation, you can start training your agent or playing the game using the command line instructions provided above. Customize your agent's behavior by modifying the callbacks.py file.
Contributing

Contributions are welcome! Please fork the repository and submit a pull request with your changes.
License

This project is licensed under the MIT License. See the LICENSE file for details.