Toward Fast Niching Evolutionary Algorithms: A Locality Sensitive Hashing-Based Approach

• Published in: IEEE transactions on evolutionary computation Vol. 21; no. 3; pp. 347 - 362
• Main Authors: Yu-Hui Zhang, Yue-Jiao Gong, Hua-Xiang Zhang, Tian-Long Gu, Jun Zhang
• Format: Journal Article
• Language: English
• Published: IEEE 01.06.2017
• Subjects: Algorithm design and analysis; Evolutionary algorithm (EA); Evolutionary computation; locality sensitive hashing (LSH); Maintenance engineering; multimodal optimization; niching technique; Optimization; random projections; Sociology; Statistics; Time complexity
• ISSN: 1089-778X
• DOI: 10.1109/TEVC.2016.2604362

Niching techniques have recently been incorporated into evolutionary algorithms (EAs) for multisolution optimization in multimodal landscape. However, existing niching techniques inevitably increase the time complexity of basic EAs due to the computation of the distance matrix of individuals. In this paper, we propose a fast niching technique. The technique avoids pairwise distance calculations by introducing the locality sensitive hashing, an efficient algorithm for approximately retrieving nearest neighbors. Individuals are projected to a number of buckets by hash functions. The similar individuals possess a higher probability of being hashed into the same bucket than the dissimilar ones. Then, interactions between individuals are limited to the candidates that fall in the same bucket to achieve local evolution. It is proved that the complexity of the proposed fast niching is linear to the population size. In addition, this mechanism induces stable niching behavior and it inherently keeps a balance between the exploration and exploitation of multiple optima. The theoretical analysis conducted in this paper suggests that the proposed technique is able to provide bounds for the exploration and exploitation probabilities. Experimental results show that the fast niching versions of the multimodal algorithms can exhibit similar or even better performance than their original ones. More importantly, the execution time of the algorithms is significantly reduced.