Bridging Reinforcement Learning and Planning to Solve Combinatorial Optimization Problems with Nested Sub-Tasks

Combinatorial Optimization (CO) problems have been intensively studied for decades with a wide range of applications. For some classic CO problems, e.g., the Traveling Salesman Problem (TSP), both traditional planning algorithms and the emerging reinforcement learning have made solid progress in recent years. However, for CO problems with nested sub-tasks, neither end-to-end reinforcement learning algorithms nor traditional evolutionary methods can obtain satisfactory strategies within a limited time and computational resources. In this paper, we propose an algorithmic framework for solving CO problems with nested sub-tasks, in which learning and planning algorithms can be combined in a modular way. We validate our framework in the Job-Shop Scheduling Problem (JSSP), and the experimental results show that our algorithm has good performance in both solution qualities and model generalizations.