Statistics of On overfitting and asymptotic bias in batch reinforcement learning with partial observability