/BoxingV0-StableBaselines

An implementation of a DQN trained on Openai Gym's boxing-v0-ram using Stable Baselines. Credit goes to [StarAI](https://www.starai.io/course/) for Colab visualization code.

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

BoxingV0-StableBaselines

An implementation of a DQN trained on Openai Gym's Boxing-ram-v0 using Stable Baselines.

The agent is playing the white character below:

An agent trained for 200,000 episodes beats the built-in AI handily.

Performance

It starts beating the built-in AI reliably after around the 70,000th episode.

Unfortunately, training got really unreliable (probably because of the vanilla DQN architecture) and performance varied a lot from episode to episode.

The highest-performing agent (saved on episode 149,000) achieved an average score of 40.54 over 37 test episodes.

Training

Want to train your own agent?

  • git clone this repository and run the ipython file here.

Alternatively:

  • Click here to access the colab file directly.
  • Click File -> Open in Playground or Save copy to Drive -> Run All.

Credits

Credit goes to StarAI for Colab visualization code. Check out the Stable Baselines library for awesome reinforcement learning algorithms!

License

License