Implementation of DQN-Taxi reinforcement learning algorithm

Technical Details

• Framework: PyTorch
• Environment: Taxi
• Category: Other

  Implementation Details

DQN-Taxi: Deep Q-Network for OpenAI Gym Taxi-v3

This project implements a Deep Q-Network (DQN) agent to solve the classic Taxi-v3 environment from OpenAI Gym. The agent learns to efficiently pick up and drop off passengers in a grid world using reinforcement learning.

[]

Environment

• Taxi-v3 is a discrete environment with:
  • State space: 16 (or 500 for the full version)
  • Action space: 6 (South, North, East, West, Pickup, Dropoff)
• The agent receives positive rewards for successful drop-offs and negative rewards for illegal moves or time steps.

Features

• DQN with experience replay and target network
• Epsilon-greedy exploration
• One-hot encoding for discrete state representation
• Logging of Q-values, advantage, and value estimates
• Integration with TensorBoard and Weights & Biases (WandB) for experiment tracking

Logging & Visualization

• Training logs and metrics are saved for visualization in TensorBoard and/or WandB.
• Q-values, advantage, and value estimates are logged for analysis.

Customization

• Change hyperparameters and logging options in the Config class in train.py.
• You can switch between different exploration strategies or network architectures as needed.

Results

The DQN agent should learn to solve the Taxi-v3 environment, achieving high average rewards after sufficient training.

References

• OpenAI Gym Taxi-v3
• DQN Paper (Mnih et al., 2015)
• Stable Baselines3

License

MIT License

Source Code

📁 GitHub Repository: DQN Taxi (DQN Taxi)

View the complete implementation, training scripts, and documentation on GitHub.