MLPacMan is a remake of the classic arcade game Pac-Man, enhanced with artificial intelligence to explore game dynamics and decision-making through Deep Q-Networks (DQN) and Dueling Deep Q-Networks (Dueling DQN). This project aims to demonstrate the application of reinforcement learning techniques and provide insights into human-computer interaction within a controlled game environment.
- Introduction
- Game Description
- Programming Language and Software
- Artificial Intelligence
- Conclusions and Future Work
- Installation
- Credits
This project integrates AI into the game of Pac-Man to evaluate AI decision-making and strategic planning in dynamic and constrained environments. Another goal was to get and insight of game development and enhance the user experience through effective human-computer interaction.
The AI models, DQN and Dueling DQN, can be trained using an epsilon-greedy strategy, ensuring a balance between exploration of new strategies and exploitation of known strategies to optimize gameplay.
MLPacMan offers both manual play and AI-driven modes, where players can engage directly or watch an AI navigate the maze. The game introduces complexities like power pellets and ghost strategies, providing a rich interaction layer.
- Initialization and Setup: The game initializes with a menu allowing players to select between playing modes.
- Game Mechanics: The game simulates Pac-Man's maze exploration, dot collection, and ghost avoidance.
- AI Interaction: In AI mode, Pac-Man is controlled by either a DQN or Dueling DQN model that learns from the game environment to optimize decision-making.
The game utilizes detailed sprites for Pac-Man and ghosts, dynamically adjusting based on game events. These include different sprites for ghost modes and Pac-Man's directions.
Python was selected due to the developer’s familiarity, which streamlined development and debugging. Key libraries include:
- Pygame: For rendering and game state management.
- PyTorch: For implementing and training neural network models.
- Matplotlib: For visualizing training results.
The game architecture is modular, separating game mechanics, AI components, and rendering to facilitate maintenance and scalability. The AI models interact directly with the game environment, processing states and calculating rewards based on game dynamics.
DQN and Dueling DQN models were integrated with an epsilon-greedy strategy. This setup helps the AI navigate the game's maze efficiently and learn optimal strategies through continuous interaction with the game environment.
The optimize_model method of the PacmanAgent class encapsulates the AI training process, employing experiences and
rewards to adjust and improve the AI strategies dynamically.
The project highlighted significant insights into AI capabilities and limitations within video game environments. Future work will focus on refining AI strategies, enhancing reward mechanisms, and exploring more complex AI models to better emulate human gameplay nuances.
To set up the MLPacMan game on your local machine for development and testing purposes, follow these steps:
git clone https://github.com/BugGreen/MLPacMan.git
cd MLPacMan
pip install -r requirements.txt
python pacman.pyIf you find this project useful for your research or development, consider giving it a star on GitHub to encourage future updates and enhancements!
Developed by Nicolás Rojas Bernal at Technische Universität Dresden, Interactive Media Lab.