LordN77/dqn-SpaceInvadersNoFrameskip-v4

DQN for Space Invaders

Vanilla Stable-Baselines3 DQN trained from random initialization on SpaceInvadersNoFrameskip-v4 for the Hugging Face Deep RL Course, Unit 3.

The selected checkpoint was saved at 20,000 agent timesteps. The training run used a compact CNN, 42×42 grayscale observations, four stacked frames, and a fixed training seed. Evaluation uses full games and unclipped rewards. Checkpoint selection and evaluation use separate seeds.

Evaluation	Mean reward	Population standard deviation	Score
10 episodes	210.00	0.00	210.00
Additional 30 episodes	210.00	0.00	210.00

The score is mean reward minus standard deviation. results.json also includes a separate 5,000-agent-step evaluation. replay.mp4 shows the first of the 10 evaluation games.

Use the evaluation environment settings in config.json or run the included training code. Only load model archives from sources you trust.