Research And Application On Covariance Learning Model Driven Evolutionary Algorithm

Evolutionary algorithms are a type of heuristic random algorithms that iteratively search through a population,and their operation process can be seen as the iterative interaction between the algorithm and the problem to be solved,during which a large amount of data about the problem information is generated.In situations where prior knowledge of problems is difficult to obtain,how to use advanced machine learning methods to deeply analyze these data,obtain and utilize the obtained knowledge,in order to improve the problem solving ability of evolutionary algorithms,has become an important trend in evolutionary computing research in recent years.In view of this,the dissertation aims to study evolutionary algorithms driven by machine learning models,and conduct systematic research on their evolutionary generation operators,knowledge transfer strategies,and population diversity maintenance.The main innovative work is as follows:1.Aiming at the bottleneck of significantly decreasing search performance and easily falling into local optima when artificial bee colony algorithm solves variables that cannot be separated,an adaptive covariance preference learning artificial bee colony algorithm is proposed.First,a search equation based on fitness ranking learning is designed to improve the mining ability of the algorithm,because the search equation of the hired bee lacks the guidance of population preference information;Secondly,in response to the defect of rapidly decreasing improvement intervals in the search equation for observing bees,a feature coordinate system based on population covariance learning is designed to reduce the correlation between variables and thereby improve the efficiency of algorithm search;Finally,to reduce the risk of the algorithm falling into premature convergence,a delete restart learning strategy is proposed,which can compensate for the limitations of a single search mode and help the algorithm more easily jump out of local optima when solving complex problems.Numerical experiments were conducted to compare the accuracy of 16 improved artificial bee colony algorithms and 7 other similar algorithms on the CEC2014 test problems.The experimental results showed that the optimization performance of the adaptive covariance preference learning artificial bee colony algorithm was significantly better than various improved artificial bee colony algorithms,and it also had strong competitiveness compared to other similar algorithms.2.Aiming at the problem that existing knowledge transfer strategies are difficult to give consideration to similar tasks and dissimilar tasks,an adaptive differential evolution algorithm based on dual transfer learning model is proposed.First,in the early and middle stages of the evolution,the classification learning model is used to implement the knowledge transfer strategy,so as to explore the similarity between different task populations to improve the search efficiency of the algorithm;Secondly,in the middle and later stages of the evolution,the knowledge transfer strategy is realized by constructing a affine transformation learning model to improve the efficiency of knowledge transfer in low similarity tasks;Finally,design an adaptive adjustment strategy for knowledge transfer probability based on population success rate,which is beneficial for reducing the risk of negative transfer phenomena.Numerical experiments were conducted to compare the accuracy of the proposed algorithm with six different multi task evolutionary algorithms on nine sets of problems.The experimental results showed that the proposed algorithm has strong robustness in solving different types of problems.3.Aiming at the difficulty of finding all PSs in the decision space for imbalanced multimodal multi-objective optimization problems,a local outlier learning model driven imbalanced multimodal multi-objective evolution algorithm is proposed.First,the evolutionary population is grouped based on the principle of speciation,which lays the foundation for the decentralized search of the population;Secondly,for each individual in the subpopulation,an outlier factor learning model is used to evaluate its convergence in the target space and diversity in the decision space;Finally,using a random sorting environment selection method to independently update each subpopulation is beneficial for maintaining the diversity of the evolutionary population;Numerical experiments were conducted to compare the accuracy of six different multimodal multi-objective evolutionary algorithms on four types of imbalanced problems.The experimental results showed that the proposed algorithm has strong competitiveness in solving imbalanced multimodal multi-objective optimization problems.