Negatively Correlated Search for Constrained Optimization

Evolutionary algorithms (EAs) combined with constraint handling techniques (CHTs) are very effective in solving constrained optimization problems (COPs). Recently, negatively correlated search (NCS) has been shown to be powerful for real-world multimodal optimization problems (MMOP). However, there has been few works on utilizing NCS to solve COPs. In this paper, we present a novel constrained optimization evolutionary algorithm (COEA) named NCS-E which combined NCS with a new CHT ensemble method to deal with COPs. It integrates three complementary CHTs as voters, and individuals are first voted by each voter, then the individual with more weighted votes is considered better. In addition, the negative correlation information of the voted individuals is utilized to adaptively adjust the voting weights. To demonstrate the performance of NCS-E, NCS-E is compared with four NCS-based COEAs and two state-of-the-art methods, MAgES and MAgES-VMCH, on 57 real-world COPs. The statistical results show that NCS-E exhibits the best performance among the compared methods. Moreover, NCS-E was applied to solve the Deep Neural Network pruning problem and compared with the state-of-the-art pruning method OLMP on three reference models. Empirical results show that the proposed method improves the performance of OLMP on all three DNN pruning tasks. © 2023 IEEE.