A learning-based two-stage optimization method for customer order scheduling

This paper addresses the customer order scheduling problem in parallel production environment commonly appearing in the pharmaceutical and paper industries. The problem aims to minimize the total completion time of the orders with their jobs processed on dedicated machines in parallel. To deal with the computational challenge of large-scale problems, we propose a learning-based two-stage optimization method consisting of a learned dispatching rule in the first stage and an adaptive local search in the second stage. The new dispatching rules are automatically generated by the proposed feature-enhanced genetic programming method in an offline learning manner. Based on the high-quality initial solutions provided by the learned dispatching rule, we develop an adaptive local search to further improve the solution quality. Numerical results indicate the superiority of the learned dispatching rule and show the proposed two-stage optimization method significantly outperforms state-of-the-art methods in the literature.