Transfer learning assisted batch optimization of jobs arriving dynamically in manufacturing cloud

Abstract

Cloud resource assembly problem (CRAP) in smart manufacturing is characterized by NP-hard complexity. Evolutionary algorithms (EAs) have become the mainstream search paradigm for tackling CRAP. Unfortunately, most EAs start the search from scratch and ignore the useful knowledge carried by solving related CRAPs, which leads to repetitive search and a waste of limited computational budget. In light of this, we devise a novel transfer learning assisted EA (TAEA) to optimize a batch of heterogeneous CRAPs (jobs) in a parallel fashion. Overlap knowledge of tackling distinct CRAPs are extracted to not only expedite the search efficacy but also leverage the solution quality. Specifically, to perceive the relatedness between jobs, a new indicator considering service collaboration features is proposed. In addition, to capture and learn cross-job knowledge in common, resource quality data associated with optimal solutions are employed as carries of knowledge and the connection between jobs is built with the help of Auto-Encoder. Furthermore, an adaptive source job selection strategy is adopted by integrating similarity-based criterion and credit-based criterion together. To test the efficacy, we compare TAEA with other baselines on a series of CRAP instances with varying degrees of similarity. Empirical results substantiate the superiority of TAEA. By exploiting the traits of transfer learning to assist cloud resource scheduling, TAEA has a great potential of expediting the scheduling efficacy for manufacturing cloud platform in practical applications.