Research On Ensemble Optimization Algorithm

Ensemble optimization algorithms can make full use of the advantages of different algorithms or strategies,and has been widely used in complex optimization problems.However,the current ensemble algorithms still have some shortcomings.The first is that random allocation strategies or algorithms to update individuals have a certain blindness,it is not conducive to improve the convergence speed of ensemble algorithms.The second is that according to historical information,such as success probability or posterior probability,select strategies or algorithms to update individuals’ positions,it is required to implement different strategies or algorithms in each generation,which can increase the computational cost.The third is that individuals are usually grouped into subpopulations according to their fitness or with a random strategy in the conventional multipopulation method,and the information exchange between different subpopulations is generally based on a specific design method,which makes increase complication of algorithm and probability of local convergence.The fourth is that the basic mechanisms of sub-algorithms used are the same for solving complex problems,due to their inherent limitations it is difficult to make a breakthrough for improving the performance of the algorithms.The fifth is that current researches mainly focus on high-level ensemble or low-level ensemble,and there are few reports on collaboration between the two ensemble frameworks.Therefore,for static single-objective optimization problem,this paper proposes two ensemble algorithms to improve the overall performance of algorithm from four aspects of grouping strategy,algorithm allocation,heterogeneous ensemble algorithm and compound ensemble optimization framework.For dynamic multi-objective optimization problem(DMOP),fast and effective tracking of Pareto front is one of the keys to solving DMOPs,a dynamic multi-objective optimization algorithm based on prediction strategy ensemble is proposed,which improves the convergence and diversity of the algorithm under different degrees of environmental change through region division and ensemble of different methods.The main research contents of this paper are as follows:(1)Due to the limitations of homogenous algorithms in solving complex problems,a framework of heterogeneous ensemble algorithms(EHA)with multiple populations is developed.First,to maintain the diversities of the subpopulations,a new grouping strategy considering the convergence accuracy and acceleration is designed to divide the population into three subpopulations.This can decrease the probability of local convergence of the algorithm.Second,three algorithms considering the characteristics of the individuals are distributed for the subpopulations to balance the exploitation and exploration abilities of the framework.Third,a method for chromosome modification is designed to realize the smooth evolution of individuals in different subpopulations.Finally,the individuals are regrouped to realize the information exchange among three subpopulations.The performance of EHA is evaluated on two data suites IEEE CEC2005 and IEEE CEC2014,and the results are compared with those of some other algorithms.The results indicate that EHA has excellent performance.(2)To improve the performance of ensemble algorithms with fully using the advantages of the two ensemble methods,a compound ensemble optimization framework(CEOF)with multi-population is developed.Firstly,the population is divided into some sub-populations,and the size of each sub-population changes adaptively according to the success rate of the constituent algorithm in high-level,which realizes adaptive allocation of algorithm resources.Secondly,the ensemble mutation strategy is designed in the low level to balance the exploration and exploitation of the population.The fitness of the individuals and the cosine similarity between the individual and the best individual jointly determine which mutation strategy the individuals adopt.Finally,the performance of CEOF is evaluated on two benchmark data suites IEEE CEC2017 and IEEE CEC2020.The simulation results are compared with those of other algorithms.The results indicate that CEOF outperforms other algorithms in terms of the average performance.(3)To track the Pareto frontier quickly and effectively in dynamic environments,a dynamic multi-objective optimization algorithm based on prediction strategy ensemble(PSE)is proposed.First,the objective space is divided into several subspace according to knee points.Second,the severity of changes in subspace are estimated by the information of the current population and historical population.After identifying the degree of change in different subspace,the knee-based prediction,center-based prediction and indicator-based local search are used as change response mechanisms to improve the convergence and diversity of the algorithm under different degrees of environmental changes.Finally,the proposed prediction strategy is introduced into MOEA/D-DE.PSE is tested on three dynamic multi-objective benchmark problems(FDA,d MOP and UDF)by varying the severity and frequency of the changes to construct three different mixed change types.Experimental results show that this method can successfully identify different types of changes,dynamically track and adapt to complex environmental changes.In summary,this thesis has carried on the thorough research to and analysis to ensemble optimization methods and frameworks,and several ensemble optimization algorithms are proposed to provide new ideas and methods for improving the performance of ensemble optimization algorithm.