A Hybrid Graph-Based Imitation Learning Method for a Realistic Distributed Hybrid Flow Shop With Family Setup Time
Junqing Li, Jiake Li, Kaizhou Gao, Peiyong Duan
Abstract
Prefabricated construction has attracted research interest as it can significantly save energy consumption. In this study, a distributed hybrid flow shop with family setup time in a typical prefabricated system is investigated. A hybrid graph-based imitation learning from multiple experts (hereafter called IML) is developed to minimize the makespan. Efficient input features with operation processing times are presented. Next, to enhance the training speed of the network, a less parameter encoder mechanism is developed. Subsequently, a multiexpert learning method is proposed, in which the solutions obtained by these experts are used as the ground truth values to enhance the convergence and searching capabilities. Moreover, a variable neighborhood search (VNS)-based local search method is embedded to further improve the performance. Finally, based on a realistic prefabricated component production horizon, a set of instances is generated to test the performance of the proposed algorithm. The comprehensive computational comparison and statistical analysis reveal that the proposed IML algorithm, when compared to two recently published efficient algorithms, yields an average improvement of about 8.66% and 13.78%, respectively. This highlights the efficiency of the proposed algorithm to solve large-scale instances.