Temporal Difference Learning for Model Predictive Control