Space Invaders DQN — Hugging Face Deep RL Course Unit 3
Project
Training a Deep Q-Network (DQN) agent to play Space Invaders using Stable-Baselines3 and RL-Baselines3-Zoo.
Environment
- SpaceInvadersNoFrameskip-v4 (training configuration)
- Atari preprocessing
- 4-frame stacking
- CnnPolicy
Training configuration
See dqn.yml for the training configuration.
Evaluation
- Evaluation environment: ALE/SpaceInvaders-v5
- Evaluation episodes: 10
- Mean reward: 384.5
- Standard deviation: 130.55
- Mean minus standard deviation: 253.95
- Evaluation target: 200
- Target achieved: Yes
Detailed results: evaluation_results_final.json
Model checkpoints
Trained DQN model checkpoints are available in the logs/dqn/ directory.
Status
Training checkpoints and evaluation results are available. The measured evaluation score exceeds the target of 200. Official course completion has not yet been independently verified.
- Downloads last month
- 95