Understanding reinforcement learning for model training from scratch