MOEA/D With Adaptive Weight Vectors

Many real-world problems can be defined as multi-objective optimization problems.These optimization problems usually involve more than one conflicting objectives.Multiobjective Optimization Evolutionary Algorithm Based on Decomposition(MOEA/D)is one of the representative decomposition-based multi-objective evolutionary algorithms.It has attracted a lot of attention in the evolutionary computation community due to its effectiveness on multi-objective and many-objective optimization.However,its performance depends on the specification of the weight vectors.This thesis analyzed the disadvantages of the existing weight vector adjustment methods.Based on the analysis,this thesis proposed a novel weight vector adjustment method and discussed the frequency specification problem in weight vector adjustment.Besides,this thesis also analyzed the existing computationally expensive multi-objective optimization algorithms and proposed a novel hybrid framework with a new weight adaptation strategy for computationally expensive many-objective optimization.The main contributions in this thesis are as follows:(1)This thesis proposed a novel method to adjust weight vectors only twice.The modified MOEA/D used a weight vector grid as the external archive to store all the solutions found during the search.At the latter stage of the algorithm,the original weight vectors and population in MOEA/D are replaced by the weight vectors and solutions in the external archive.The proposed method not only minimizes the effect of weight vector adjustment on the convergence of the algorithm,but also avoids the problem of insufficiency of the weight vector when it is being adjusted minimally.(2)This thesis proposed a novel hybrid framework.The framework comprises the classification-based algorithm and the decomposition-based algorithm together.A novel weight adaptation strategy is also adopted as the connecter between the two different computationally expensive multi-objective algorithms.The new hybrid algorithm inherits the diversity maintenance ability and the convergence ability from the two underlying algorithms.