A deep reinforcement learning framework for solving two-stage stochastic programs

In this study, we present a deep reinforcement learning framework for solving scenario-based two-stage stochastic programming problems. Stochastic programs have numerous real-time applications, such as scheduling, disaster management, and route planning, yet they are computationally challenging to solve and require specially designed solution strategies such as hand-crafted heuristics. To the extent of our knowledge, this is the first study that decomposes two-stage stochastic programs with a multi-agent structure in a deep reinforcement learning algorithmic framework to solve them faster. Specifically, we propose a general two-stage deep reinforcement learning framework that can generate high-quality solutions within a fraction of a second, in which two different learning agents sequentially learn to solve each stage of the problem. The first-stage agent is trained with the feedback of the second-stage agent using a new policy gradient formulation since the decisions are interconnected through the stages. We demonstrate our framework through a general multi-dimensional stochastic knapsack problem. The results show that solution time can be reduced up to five orders of magnitude with sufficiently good optimality gaps of around 7%. Also, a decision-making agent can be trained with a few scenarios and can solve problems with many scenarios and achieve a significant reduction in solution times. Considering the vast state and action space of the problem of interest, the results show a promising direction for generating fast solutions for stochastic online optimization problems without expert knowledge.