Reinforcement Learning Optimizes Scenario Tree Construction for Multistage MPC

A recent preprint on arXiv (2608.09335) introduces a method aimed at enhancing control for building scenario trees in multistage stochastic model predictive control (MPC). Unlike conventional methods, such as those based on Wasserstein scenario reduction that prioritize aligning with the probability distribution, this approach emphasizes that better distributional accuracy does not guarantee superior control outcomes. The authors propose a sequential assignment of sampled scenarios to tree leaves, utilizing an attention-based policy trained via reinforcement learning, with the goal of maximizing closed-loop control profit. An asymmetric critic is employed to stabilize training by utilizing actual future trajectories. This method seeks to optimize tree construction by considering its effects on subsequent decisions, potentially boosting control efficacy in uncertain settings.