Application Of Gene Expression Programming Based On Layer Distance

Genetic Algorithm(GA)uses the characteristic that groups in environment are able to evolve with the time to find optimal solutions.As the optimal problems become much more complex,the encoding style of individuals and groups has turned into trees from linear strings.However,both of the two styles have shortcomings,such as the strings cannot express complex practical problems and trees will decrease the efficiency of genetic operators during evolutionary process.Gene Expression Programming(GEP)have built an expression mechanism between strings and the trees,thus it can change the tree's structure by modifying the components of strings.In this way,GEP can significantly improve the ability in solving complex problems and evolutionary efficiency.However,GEP still has some defects.This thesis aims at those defects in GEP and the main content is as follows:1.In order to solve low efficiency in genetic operation caused by the asymmetry tree structure of GEP,this thesis modifies the asymmetry structure into the symmetry structure,then defines the Layer Distance(LD)from this structure and proposes a Gene Expression Programming based on Layer Distance(LD-GEP).This algorithm combines the LD with genetic operators,proposes three new genetic operators: genetic recombination,layer recombination and layer mutation respectively.These three operators can reduce the number of parameters and the effects of human factors.2.This thesis proposes a method to ascertain the operation point of those three genetic operators by using the Hamming distance.Compared with the single evolutionary direction in GEP,this method is able to provide a new direction by the internal structure during the process,which can improve the efficiency of finding optimum points,specify the location of variation and the function of each individual.This mechanism can also balance the ability of exploration and exploitation during the evolutionary process.3.To verify the effectiveness of LD-GEP,this thesis adopts open datasets from UCI Machine Learning database in classification and function modeling problems respectively.For classification problems,LD-GEP has similar performance on 2-class and 3-class datasets with compared algorithms,all of them have F1 value over 0.9;whereas for the Avila dataset which contains more than 5 classes,LD-GEP still has a F1 value more than 0.9,exceeding 0.1 to 0.15 than compared algorithms.For function modeling problems,this thesis uses PM2.5 dataset in BeiJing and the sunspots dataset.The mean absolute error,the mean absolute percent error and the root mean squared error are selected for judging the performance of algorithms.The results show that,LD-GEP is better in all sectors than other compared algorithms.