A Reinforcement Learning-Artificial Bee Colony algorithm for Flexible Job-shop Scheduling Problem with Lot Streaming

Abstract

As a typical production model in manufacturing industry, Flexible Job-shop Scheduling Problem (FJSP) has an important impact on enhancing the productivity of enterprises. Flexible Job-shop Scheduling Problem with Lot Streaming (FJSP-LS) is an extension of FJSP that allows jobs to be split into multiple sublots so they can be processed and transported separately. Since FJSP-LS has a large solution space and it is difficult and unstable for many algorithms to find a high-quality solution, this paper proposes a hybrid algorithm combining Reinforcement Learning and Artificial Bee Colony (RL-ABC) algorithm. In RL-ABC, the utilities for solving FJSP-LS are divided into 2 stages: (1) determining the best dispatch scheme and (2) determining the best scheme of sublots. For stage 1, an algorithm with different initialization and local search strategies is proposed. For stage 2, reinforcement learning is developed by building mappings between the environment and schemes of sublots. The effectiveness and robustness of RL-ABC algorithm and its components are compared with five algorithms including three types (traditional heuristic algorithm, improved heuristic algorithm and new evolutionary algorithm) on nineteen benchmark instances and three real instances. The results show that although RL-ABC algorithm exhibits inferior performance in terms of CPU time, its effectiveness and robustness surpass all the other compared algorithms on all instances. Moreover, both components of the RL-ABC algorithm effectively reduce the Makespan. Therefore, it can be used as a new technique to solve large-scale and complex problems in scheduling domain.