tashobi02/dqn-SpaceInvaders-v5-5k

DQN Space Invaders - 5,000-step local run

This model was trained from scratch for exactly 5,000 environment steps with RL Baselines3 Zoo and Stable-Baselines3 on CPU, using seed 42. It completed 1,125 gradient updates after a 500-step replay warm-up. This short learning run is an early checkpoint; it is not presented as a solved or human-level Atari agent.

Evaluation over 10 deterministic, full-game episodes: 165.00 +/- 47.75 game points. Evaluation uses unclipped rewards and ends when the game ends, rather than when a single life is lost. Per-episode scores and the evaluation settings are in results.json.

Environment and preprocessing

The current environment ID is ALE/SpaceInvaders-v5, with frameskip=1, repeat_action_probability=0.0, and full_action_space=False, matching the old SpaceInvadersNoFrameskip-v4 settings. AtariWrapper applies four-frame action repeats, max pooling, 84x84 grayscale preprocessing, and training reward clipping. Four frames are stacked for the CNN. An environment step here is one wrapped action, not one raw emulator frame or one optimizer update.

Load and run

python -m pip install -r requirements.txt
python load_model.py

load_model.py downloads the model when it is not already present, constructs the matching evaluation environment, and runs a full demonstration episode. It can also be imported for make_env() and load_agent().

Files and provenance

• dqn-SpaceInvaders-v5-5k.zip: the locally trained Space Invaders DQN.
• replay.mp4: this model's post-training demonstration.
• config.yml, training.json, requirements.txt, and results.json: configuration, training evidence, dependencies, and evaluation.
• all-agents-5k.mp4, when present: the local Space Invaders agent beside the notebook's pretrained Beam Rider comparison, revision b9165071ded8a581add77dd56cfe37cd6d3134eb. Beam Rider was not trained locally and its weights are not included in this repository.

The demo JSON files state whether each clip contains a complete episode or reaches the 1,500-step recording limit. The combined video holds the shorter replay's last frame.

References

• Playing Atari with Deep Reinforcement Learning
• Human-level control through deep reinforcement learning
• Space Invaders environment documentation