A PyTorch-based Deep Q-Network (DQN) agent trained to master Flappy Bird and other games offered on the Gymnasium platform using modern reinforcement learning techniques.
This project demonstrates how AI agents can learn optimal policies through self-play using experience replay, epsilon-greedy exploration, and advanced Q-learning strategies like Double DQN and Dueling DQN.
FlappyBot learns to play the classic side-scrolling game Flappy Bird by interacting with a Gym-style environment. Over thousands of episodes, the bot improves its decision-making by learning from rewards and penalties.
✅ Deep Q-Learning
✅ Experience Replay
✅ Epsilon-Greedy Exploration
✅ Target Network Syncing
✅ Double DQN (optional)
✅ Dueling DQN (optional)
✅ PyTorch-based architecture
✅ Fully configurable via YAML
✅ Reward and epsilon tracking graphs
✅ Model saving & loading
- Python 3.7+
pippackage manager- PyTorch
git clone https://github.com/yourusername/flappybot.git
cd flappybot
pip install -r requirements.txt
### Running the bots/training
- Add config to hyperparams.yml for a new game
- For easy training add the new game to the launch.json file and run from the run and debug tab of vscode
- Model training info is added to the runs directory
- To run a model just run python agent.py {game_name}