Improvement And Application Of Gene Expression Programming Algorithm

In order to solve a variety of optimization problems,many optimization algorithms,such as gradient descent,newton method and so on,have been proposed.But limited to the algorithm itself lead to the objective function must satisfy the conditions such as continuous,differentiable optimization,so these optimization algorithms are not suitable for high dimension,nonlinear,multi-objective and other large-scale complex problems,and easy to fall into local optimum.Gene expression programming algorithm is an evolutionary algorithm which can solve the optimization problem effectively.It has the characteristics of general applicability,group intelligence,high robustness and potential parallelism,and is not limited by the nature of the problem to be solved.However,the algorithm still has some problems,such as slow convergence speed,premature convergence,poor generalization ability.In addition,it is necessary to enhance the support of its mathematical theory.This paper makes a comprehensive and in-depth study on the content,algorithm flow and key technologies of gene expression programming algorithm.And put forward the further improvement of the algorithm.The work and contents are summarized in 6aspects.(1)Through summarizing the research significance,the thought content and the algorithm process of the evolutionary algorithm and GEP algorithm,this paper points out the deficiency of the GEP algorithm,and puts forward some suggestions to improve it.(2)The adaptive evolutionary parameter is designed.According to the number of generations and the ranking of fitness value in the population,the recombination rate and mutation rate are dynamically adjusted.It can solve the problem of slow convergence rate at the late of evolution stage,protect good individuals and enhance the evolutionary parameters of inferior individuals to adapt to evolution.So that the algorithm can jump out of local optimum as soon as possible.(3)Age stratified structure is designed.It can make sure that the initial new individual has certain space and time to fully evolve,search the extreme point in the vicinity,diversify of good genes,further expand the search space,improve the convergence speed,and largely avoid premature,finally get the global optimal solution;(4)GEP is ported to the Spark framework for distributed parallel computing.A large population is divided into a number of sub populations,which are distributed to different working nodes to perform the calculation,selection,recombination and mutation of the fitness values of each other.At the same time,select the best individuals to migrate to other nodes to realize the exchange and sharing of genetic information.The algorithm has the ability to analyze and deal with large data,effectively improve the training speed and efficiency.(5)The improved algorithm is applied to the short-term forecast of China's Shanghai Composite Index,and the historical time series data are analyzed.Compared with the traditional GEP,the predicted results of MAPE decreased by 0.213%and R~2increased by 0.1095,and the algorithm has linear speedup in the Spark cluster environment.This improved algorithm has faster execution speed and higher prediction accuracy and stability.(6)A content-based image retrieval system is designed and implemented.The problem of image matching is transformed into the function discovery of gene expression programming algorithm.GEP algorithm is used to optimize the weights of multiple image features,and a new image similarity model is obtained and written to the index library to quickly respond to user queries.The system can improve the image matching accuracy,to meet the needs of users to search for images with images.