Variable Structure Representation Based Genetic Programming and Its Application in Self-Adaptive Modeling

摘要

In the application of genetic programming, the tree-like individual often adopts 'S' expressions in the Lisp language to be presented, but this kind of coding method is inconvenient for genetic operations. This thesis proposed an effective representation based on the structure data and symbol objects of Matlab, using structures and symbol objects to express the tree-like individual expressions. Each individual structure has four fields: left, right, operator and value, the first two stands for the left and right branches of the individual tree, the field of "operator" is a symbol vector including the operators or functions which connect the elements in left and right branches, the last field of "value" is used to store the fitness value of each individual. The whole population can be expressed by a structure array whose number of elements equals to the population size. The size of individual structure may be changed during genetic operations. To keep the diversity of the population, a re-initialization process is combined with the evolution. At the end of this thesis, we established a model to analyze the relation between the total values of imports & exports of China from 1985-1997 and the corresponding number of containers used, then compared the results with 'S' expression based GP and the normal methods in data modeling including linear regression and two-order polynomial regression. Experimental results indicate that variable structure based GP can express the individual solution more directly and effectively than those mentioned.