Value function gradient learning for large-scale multistage stochastic programming problems

• A stagewise decomposition algorithm for large-scale stochastic program is proposed. The approximates value functions by predetermined parametric forms. fits the gradients of function to those function. Numerical examples show computational efficiency proposed algorithm. called “value gradient learning” (VFGL) multistage convex programs. VFGL finds parameter values that best fit within a given family. Widely used algorithms programming, such as dual dynamic programming (SDDP), approximate adding linear subgradient cuts at each iteration. Although this approach has been successful problems, nonlinear problems may suffer from increasing size subproblem iteration proceeds. On other hand, fixed number parameters; thus, subproblems remains constant throughout Furthermore, can learn parameters means descent, which it be easil0y parallelized and does not require scenario tree approximation underlying uncertainties. was compared with deterministic equivalent formulation problem SDDP approaches three illustrative examples: production planning, hydrothermal generation, lifetime financial planning problem. generates high-quality solutions computationally efficient.