Dual dynamic programming for stochastic programs over an infinite horizon

Caleb Ju, Guanghui Lan

Published 2023-03-03, updated 2023-04-04Version 2

We consider a dual dynamic programming algorithm for solving stochastic programs over an infinite horizon. We show non-asymptotic convergence results when using an explorative strategy, and we then enhance this result by reducing the dependence of the effective planning horizon from quadratic to linear. This improvement is achieved by combining the forward and backward phases from dual dynamic programming into a single iteration. We then apply our algorithms to a class of problems called hierarchical stationary stochastic programs, where the cost function is a stochastic multi-stage program. The hierarchical program can model problems with a hierarchy of decision-making, e.g., how long-term decisions influence day-to-day operations. We show that when the subproblems are solved inexactly via a dynamic stochastic approximation-type method, the resulting hierarchical dual dynamic programming can find approximately optimal solutions in finite time. Preliminary numerical results show the practical benefits of using the explorative strategy for solving the Brazilian hydro-thermal planning problem and economic dispatch, as well as the potential to exploit parallel computing.