Super Mario Bros Reinforcement Learning Agent

This project implements a reinforcement learning agent that learns to play Super Mario Bros using Deep Q-Learning (DQN). The agent uses a Convolutional Neural Network (CNN) to process game states and make decisions.

Features

• Deep Q-Learning implementation for game control
• CNN-based state processing
• Custom reward system for optimal gameplay
• Automatic model saving and loading
• GPU acceleration support with CUDA 12.4
• Anti-freezing mechanism to prevent agent from getting stuck

Requirements

• Python 3.8+
• PyTorch with CUDA 12.4 support
• Gym Super Mario Bros
• NES-Py
• NumPy
• Pillow
• NVIDIA GPU with CUDA 12.4 support

Installation

1. Clone this repository
2. Create environment

conda create -n mario-rl python=3.8
conda activate mario-rl

3. Install PyTorch with CUDA 12.4 support:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

4. Install the remaining required packages:

pip install -r requirements.txt

Usage

Run the training script:

python mario_main.py

The agent will:

• Train for 100 episodes
• Save the model after training
• Display training statistics
• Render the game during training

Model Architecture

The CNN model consists of:

• 3 convolutional layers
• 2 fully connected layers
• ReLU activation functions
• Output layer for action selection

Training Parameters

• Learning rate: 0.0001
• Batch size: 128
• Gamma (discount factor): 0.99
• Epsilon (exploration rate): 1.0 with decay
• Freeze time limit: 60 frames

Hardware Requirements

• NVIDIA GPU with CUDA 12.4 support
• Minimum 8GB RAM
• CUDA Toolkit 12.4 installed

License

This project is open source and available under the MIT License.

Author

AnssiO