Statistics of Min max generalization for two-stage deterministic batch mode reinforcement learning: relaxation schemes