A Study Of Interactive Preference Multi-objective Evolutionary Algorithm Through Decomposition

Multi-objective optimization evolutionary algorithms(MOEAs) are widely applied in many engineering areas as a useful tool. Evolutionary algorithms, as a heuristic method,has become a hot spot. Since 1990, with the development of the evolutionary algorithms, many evolutionary multi-objective optimization(EMO) methodologies have been proposed. As the most popular evolutionary algorithms, decomposition based multi-objective evolutionary algorithms has attracted extensive attention. Interaction is a new trend, in which the Decision Maker(DM) dynamically directs the searching of EMO algorithm using preference information provided by DM during the optimization process. In this paper, we discuss interactive preference based multi-objective evolutionary algorithm through decomposition, the main work of this paper is as follows:1. An interactive reference region based evolutionary algorithm through decomposition is proposed, which uses MOEA/D as a basic framework. First, a new preference pattern which based on preference region is introduced to the proposed algorithm. By dealing with the subproblems in the preference region only, the computational complexity has been decreased to a great extent. Second, a flexible interactive condition is proposed in this work and it can cope with various preference information including uncertain and inconsistent human decision and modifications of decision.2. An interactive multi-objective evolutionary algorithm based on decomposition of the objective space in the proposed algorithm. First of all, the objective space can be decomposed into a set of subobjective space by a set of direction vectors and each subobjective space consists a best solution found so far in this sub-region. The subobjective space can be treated as an external archive. In the second place, a Maximin operator is used to calculate individual fitness, which can take diversity and domination into consideration. At last, an easy interactive preference method which based on preference region is proposed in this work, which can fast capture the corresponding subspace and optimize it. The experiment results illustrate that the proposed algorithm has good diversity and convergence and easily deals with the preference information the DM provided.3. A convergence acceleration operator for multi-objective algorithm based on decomposition of the objective space is proposed. In this work, by a local improvement in objective space and objective space to decision space mapping using RBF neural networks, the proposed algorithm consistently improve the speed of convergence of the original algorithm while maintaining the desired distribution of solutions.