Ensemble of Dynamic Resource Allocation Strategies for Decomposition-Based Multiobjective Optimization

Evolutionary algorithms via decomposition, namely, DEAs, decompose the original challenging problem and evolve a number of subproblems/subspaces concurrently in a cooperative fashion. Adaptive computational resource allocation (CRA) strategy is able to identify the efficiency of different subspaces and invest search effort on them accordingly in an online manner. A crucial issue for CRA is to measure the efficiency of subspaces. Unfortunately, existing approaches for efficiency measurement are either fitness improvement oriented or contribution oriented, which struggle to capture the potentials of subspaces accurately. To mitigate such drawback, we present an ensemble method for CRA, based on the recent fitness contribution rates (FCRs) and fitness improvement rates (FIRs) of subspaces simultaneously. In order to dynamically track the potential of each subregion, we adopt two memory matrices to record FIR and FCR for multiple subspaces over recent generations, respectively. Afterward, an aptitude vector indicating the potentials of subspaces is defined by exploiting FCR and FIR with memory and decaying scheme. On the basis of above strategies, an ensemble CRA (ECRA) scheme is designed, which is then embedded into an adaptive objective space partition-based DEA, termed ECRA-DEA, for solving the multi/many-objective optimization. Extensive experimental studies for ECRA-DEA on various types of challenging problems have been carried out and the results confirm that ECRA is effective. Besides, the competence of ECRA-DEA is empirically validated in comparison with state-of-the-art designs. The proposed ECRA paves a new way to leverage the capability of DEAs on handling complex problems.