Principled Hierarchical Decision Making with Inverse Optimization.

Complex control tasks often benefit from breaking them down into smaller, manageable subgoals, a process known as hierarchical decision-making. Current approaches, however, face challenges: reinforcement learning methods struggle with guaranteeing strict adherence to constraints, while optimal control techniques can be computationally intensive and short-sighted. This new framework seeks to bridge this gap. The proposed method introduces a principled way to combine high-level goal setting with detailed low-level actions. It leverages inverse optimization, learning the structure of the lower-level problem from expert demonstrations. This ensures that the immediate objectives of the lower-level policies remain consistent with the overall, long-term task goals. Evaluations across tasks like network resource allocation and collision avoidance demonstrate that this framework outperforms existing methods, including end-to-end reinforcement learning and learning-augmented optimal control, in both efficiency and the quality of decisions made.