Research On Grouping Strategy And Local Search Method For Large Scale Global Optimization

In the past decades,many meta-heuristic algorithms are proposed to solve global optimization problems,most of which show good ability to find the global optimal value and can get the solution needed.However,with the development of science and technology,problems with 5 or 10 decision variables turn to be problems with thousands of decision variables.Traditional small scale optimization problems begin to turn larger and more complex.As the increase of decision variables,search space of the problems become larger and it is hard for these evolutionary algorithms to find global optimal.These algorithms are suffering from the phenomenon called the curse of dimensionality.The reason for the deterioration of these problems are two aspects.First,search space shows an exponential increase when the number of decision variables increase linearly.Second,with the increase of decision variables,most problems show a change of characteristics,some of which change from unimodal problems to multimodal problems.And large scale global optimization problems can divide into two parts,white-box problems and black-box problems,according to whether the equations are known.And large scale global optimization problems become a hot and key topic in this research area and draw much attention from researchers.First,a main research area for large scale global optimization is how to divide a large scale problem into several small scale ones to make it easier to solve.If a decision variable is independent from other variables,it can be solve separately and treated as constant variable.Then,the key point of this problem is how to divide the related decision variables into a subcomponent.This paper proposes a new grouping strategy for large scale global optimization problems based on the divide-and-conquer strategy.Through finding the related decision variables,the related group can be formed to find global optimal solution.Partial correlation coefficient is used in this paper to find the relationship from every two decision variables,and the idea of machine learning is used to revise the grouping results through optimization procedure.Meantime,cooperative co-evolution frame is used to optimize the grouped decision variables.Second,for large scale global optimization with a fixed evaluation number,to speed up convergence rate,a new method based on SaNSDE is proposed which adds zooming strategy to SaNSDE and uses a new mutation method in zooming.Decreasing the search space through oprimization process.And with the use of slide window,we can also avoid converging to a local optimal solution.In large scale global optimization,the search space of every decision variable increases largely,so if we can decrease search space horizontally,the convergence speed will increase a lot.To test the efficiency of the proposed method,the grouping strategy is tested on Benchmark Functions for the CEC' 2010 Special Session and Competition on Large-Scale Global Optimization and compared with state-of-art grouping strategies.The new SaNSDE method is tested on Benchmark Functions 2005.And the two methods are also integrated to test on Benchmark Functions 2010 and compares with the other methods.Experimental results show that the methods proposed show good efficiency in solving large scale global optimization problems.