Alternating Mixed-Integer Programming and Neural Network Training for Approximating Stochastic Two-Stage Problems