DQN Agent Playing Space Invaders
Trained using Stable-Baselines3 and RL Baselines3 Zoo.
Environment
- Game: ALE/SpaceInvaders-v5
- Algorithm: Deep Q-Network (DQN)
- Policy: CnnPolicy
- Training checkpoint: 700,000 timesteps
- Full experiment: 1,000,000 timesteps
Evaluation
30 episodes, using stochastic action selection.
- Mean reward: 573.33
- Standard deviation: 207.10
- Minimum reward: 205.00
- Maximum reward: 1345.00
- Mean minus standard deviation: 366.23
Environment Processing
- AtariWrapper
- 4 stacked frames
- VecTransposeImage
- Base environment frameskip: 1
- Repeat action probability: 0.0
Files
- model.zip: trained DQN model
- replay.mp4: gameplay demonstration
- evaluation_results.json: 30-episode results
- dqn_config.yml: training configuration
Notes
The gameplay video uses greedy action selection. The evaluation used stochastic action selection.
This experiment uses ALE/SpaceInvaders-v5, whereas the original Hugging Face course used SpaceInvadersNoFrameskip-v4. Results are not directly interchangeable.
- Downloads last month
- 25