A deep reinforcement learning project that trains an AI agent to play Pokémon Blue using PyBoy emulation and deep Q-learning.

This project uses deep Q-learning to train an agent to navigate and play Pokémon Blue on the Game Boy. The system uses PyBoy for emulation and PyTorch for the deep learning model.

Inspired by Peter Whidden's Youtube Video.

• Custom Gym environment for Pokémon Blue
• Deep Q-Learning implementation with PyTorch
• Exploration-based reward system
• Automated training pipeline
• Model checkpointing

• Python 3.8+
• PyTorch
• Gymnasium
• PyBoy
• NumPy
• A legally obtained Pokémon Blue/Red ROM file

1. Clone Repo
2. Install dependencies:

pip install torch numpy gymnasium pyboy
``

3. Place your legally obtained Pokémon ROM file in the project directory as `POKEMONR.GBC`.

4. (Optional) Create a saved state file using PyBoy and save it as `state_file.state`.

## Usage

Run the training script:

```sh

python pokemon-emulator-project.py

The custom PokemonBlueEnv class wraps PyBoy to provide a Gym-compatible environment. It:

• Captures screen state as observations
• Translates actions to Game Boy button presses
• Calculates rewards based on exploration of new map tiles
• Manages episode termination

The DeepQLearningAgent implements deep Q-learning with:

• Convolutional neural network for processing screen images
• Experience replay for stable learning
• Epsilon-greedy exploration strategy
• Target network for stable Q-value estimation

The agent receives rewards for:

• Exploring new map tiles (visiting new coordinates)

Modify these constants in the script to adjust training:

• REPLAY_MEMORY_SIZE: Size of experience replay buffer
• STEPS_PER_EPISODE: Maximum steps per episode
• NUM_EPISODES: Number of episodes to train

After training, model weights are saved to:

• [MODEL_NAME]_model_state_dict.pth
• [MODEL_NAME]_optimizer_state_dict.pth

• Incorporate game-specific rewards (badges, Pokémon caught, etc.)
• Implement prioritized experience replay
• Add visualization tools for agent behavior
• Support for saving/loading training progress

• Built with the guidance of Claude AI
• Inspired by Peter Whidden's reinforcement learning projects
• Uses PyBoy for Game Boy emulation