15M steps

Retrained to 10M steps, with higher play_against_latest_model_ratio (0.75 instead of 0.25) this helped the model to learn to play defense better

Trained with optuna hyperparam optimization with ml-agents-optuna