DMEA: a direction-based multiobjective evolutionary algorithm
A novel direction-based multi-objective evolutionary algorithm (DMEA)is proposed, in which a population evolves over time along some directions of improvement. We [Bui, Liu, Bender, Barlow, Wesolkowski and Abbass] distinguish two types of directions: (1) the convergence direction between a non-dominated solution (stored in an archive) and a dominated solution from the current population; and, (2) the spread direction between two non-dominated solutions in the archive. At each generation, these directions are used to perturb the current parental population from which offspring are produced. The combined population of offspring and archived solutions forms the basis for the creation of both the next-generation archive and parental pools. The rule governing the formation of the nextgeneration parental pool is as follows: the first half is populated by non-dominated solutions whose spread is aided by a niching criterion applied in the decision space. The second half is filled with both nondominated and dominated solutions from the sorted remainder of the combined population. The selection of nondominated solutions for the next-generation archive is also assisted by a mechanism, in which neighborhoods of rays in objective space serve as niches. These rays originate from the current estimate of the Pareto optimal front’s (POF’s) ideal point and emit randomly into the hyperquadrant that contains the current POF estimate. Experiments on well-known benchmark sets have been carried out to investigate the performance and the behavior of the DMEA. We validated its performance by comparing it with four well-known existing algorithms. With respect to convergence and spread performance, DMEA is very competitive.