Vladprime/poca-SoccerTwos

SoccerTwos

Trained from random initialization for Vlad Kolomiiets (Vladprime) using the official Hugging Face Deep RL Course Unity environment and Unity ML-Agents 1.1.0. No pretrained weights were used. This is an initial training run; it is not claimed to have converged. Units 5 and 7 require publishing a trained model without a reward threshold.

Training: poca, 30000 configured environment steps, seed 20261101. Export: genuine ML-Agents ONNX plus PyTorch checkpoint and trainer configuration. Separate ONNX evaluation process: seed 1000000; 40 completed agent episodes, mean -0.1039, standard deviation 0.5858. Original environment, action and observation spaces were retained. See results.json for individual returns and evaluation limitations.

Offline MA-POCA self-play training and local evaluation only. No ranked AI-vs-AI challenge result is claimed; the course online leaderboard is unavailable. Player episodes within a soccer match are correlated.

The .onnx file is compatible with the course Unity browser viewer. Training/evaluation code is in train-unity.py; the exact run configuration is config.yaml.