poca-SoccerTwos

MA-POCA SoccerTwos with self-play

Trained from random initialization by Kay Zheng (bestdive), with AI coding and execution assistance, for the Hugging Face Deep RL course. Training used free Colab CPU, seed 42, ML-Agents 1.1.0, PyTorch 2.2.2 and 200096 steps. No externally pretrained policy or fabricated metrics were used.

The original course download was unavailable; the environment is Unity's official registry SoccerTwos scene from https://storage.googleapis.com/mlagents-test-environments/1.1.0/linux/Startup.zip . This is the official SoccerTwos simulation, not an unrelated relabeled environment.

Independent evaluation

100 completed agent episodes, seed 100003, deterministic self-play with this same policy controlling both teams. Rewards include each agent's individual plus group reward. Mean -0.04633200287818909, population standard deviation 0.9566068004261077. These are self-play returns, not a ranking against external opponents or a claim of soccer mastery. The course Unit 7 requires a genuinely trained model submission without a minimum reward.

The ONNX model, full evaluation records, evaluation script, training log and configuration are included.

Reproduce

Python 3.10.12, mlagents==1.1.0, torch==2.2.2, onnx==1.15.0, setuptools<81, onnxruntime.

mlagents-learn training-config.yaml --env=PATH_TO_STARTUP --run-id=Kay-SoccerTwos --results-dir=results --no-graphics --seed=42 --torch-device=cpu --env-args --mlagents-scene-name Assets/ML-Agents/Examples/Soccer/Scenes/SoccerTwos.unity