Tackling temporal-dynamic service composition in cloud manufacturing systems: A tensor factorization-based two-stage approach

Service composition (SC) is a critical matching process in cloud manufacturing (CMfg) mainly to organize geographically-scattered manufacturing services for manufacturing tasks. Currently, most of the existing CMfg SC approaches assume that cloud service data are always static without considering the dynamics in cloud environments. Meanwhile, existing approaches rely heavily on easy-to-get dense historical service data for SC. However, adapting such ideal approaches to fast-changing temporal-dynamic cloud environments possibly leads to unsatisfactory matching results. This paper proposes a hybrid two-stage approach to deal with temporal-dynamic CMfg SC problem by incorporating tensor factorization (TF). The approach has two stages: temporal-dynamic prediction (Stage I) and many-objective SC optimization (Stage II). For Stage I, a non-negative TF based CMfg SC predictor is proposed to obtain average attribute values of candidate services in desired time slots easily and conveniently. The predictor extracts temporal-dynamic QoS values of candidate services into low-dimensional latent factor space and build a low-rank approximation. Based on the predicted service data, Stage II optimizes temporal-dynamic CMfg SC solutions by an improved many-objective evolutionary optimizer, GrEA-X. GrEA-X is reinforced with multiple variation operators and a modified mating pool selection procedure to enhance its exploration ability in a given time limit. Numerical experiments are then carried out to validate the efficacy of the proposed approach. Based on these results, we confirm the superior performance of our approach over other benchmarks.