Research On Diversity Enhancement Methods For Probability Distribution Estimation Algorithm

Optimization problems are ubiquitous and increasingly complex in big data and the Internet of Things,which leads to great challenges in the performance of the estimation of distribution algorithm(EDA).To enhance the performance of the algorithm in solving optimization problems in complex environments,first,this paper designs an adaptive covariance scaling estimation of distribution algorithm(ACSEDA)to address the problem that the covariance shrinks too fast,resulting in premature convergence of the population into local area.Secondly,for the lack of diversity caused by traditional single model,this paper proposes a layered learning estimation of distribution algorithm(LLEDA).ACSEDA mainly optimizes the probability distribution model evaluation method of the traditional ways,which calculates the mean vector and covariance matrix on the same highquality individuals.Since the traditional method brings the problem of covariance convergence,let the covariance matrix be calculated using more high-quality individuals than mean vector,thereby expanding the covariance and improving the diversity of sampling offspring.For alleviating the sensitivity of parameters,this paper further designs an adaptive individual selection strategy.That is,the number of individuals used to evaluate the mean and covariance decreases nonlinearly during optimization.in addition,the individuals of the covariance calculation are more than mean vector in whole process,until the number is the same in the later stage.By this mechanism,to balance the diversity and convergence of EDA.LLEDA maintains the diversity of the population by constructing multiple distribution models.Specifically,the individuals in the population are calculated and sorted,then divided into multiple layers.After that,the corresponding distribution models are constructed.Multiple models bring more sampling space and provide better diversity.In addition,to improve the quality of the sampling offspring.A novel learning mechanism is introduced,that is,to let the mean vector of the poor layer randomly learn to the better one,so that the sampled individuals are biased towards better areas,increasing the probability of capturing high-quality individuals.On constructing the parent population,this paper further designs a cross-generation population selection strategy.By combining the population offspring of the last generation and the current generation to form a larger sample,then select better individuals to estimate the probability distribution model,so as to facilitate the use of historical information and current information to guide the population search.Finally,to verify the effectiveness of two algorithms,this paper conducts lots of experiments on the CEC 2014 test set and compares them with the latest EDA variants.The experimental results show that two algorithms proposed in this paper have better optimization performance than the compare algorithms.Overall,this paper designs strategies to increase the search diversity of EDA from both single model and multi model perspectives,and conducts extensive experiments on the test set to demonstrate the effectiveness of the proposed improved strategy.