Research On The Subproblems And Local Search Strategies In Decomposition-based Multiobjective Evolutionary Algorithm

Over the last several years,the decomposition-based multi-objective evolutionary algorithm(MOEA/D)has received much attention in the field of evolutionary computation.MOEA/D has been widely used in a wide range of real-world applications.This thesis identifies some common shortcomings of the existing MOEA/D variants and proposes several approach-es to overcome them for improving the performance of MOEA/D.The major contribution includes:1.Since the traditional decomposition approaches are not originally designed for pop-ulation based algorithms,direct use of them in MOEA/D may not suitable for some multi-objective optimization problems(MOP).We have analyzed a set of test prob-lems which are different for traditional decomposition approaches and a novel decom-position approach,the constrained decomposition approach.Our experimental studies show that MOEA/D with our proposed approach can balance the population conver-gence and the diversity better than other approaches.Comparison demonstrates that our MOEA/D variant outperforms others.2.It could be difficult to determine the constraint-degree in the constrained decomposi-tion approach for some MOPs beforehand.To address this issue,an adaptive strategy for online adjusting the constraints is proposed for promoting the population diversity.In this strategy,a divergence degree of the current solution for each subproblem is first defined,and then the average divergence of all the subproblems is used to mea-sure the population diversity.Using these metrics,we determine how to penalize the constraints.We have shown that our proposed strategy is effective.3.A generic problem transformation method is proposed for dealing with some difficult MOPs.The Pareto front shapes of some MOPs could cause difficulties for MOEA/D.Our proposed method can change the shape of the PF and make it easier for MOEA/D.We have theoretically studied this approach and tested it on a number of test problems.We also show that this method can accommodate the decision maker's preference,which can be very important for solving many objective problems.4.A hybrid method of MOEA/D with a gradient based local search is proposed.We dy-namically change the step size for balancing the convergence and diversity when the gradient descent search is conducted.The frequency of using the local search strategy is also dynamically adjusted during the evolution process.Besides,the quadratic in-terpolation is used to find an approximate optimal solution.The resultant algorithm is compared with MOEA/D-DRA to show its effectiveness.