Gene Expression Programming: Theory And Supervised Machine Learning

Gene Expression Programming (GEP) is a new member of Evolutionary Algorithms, although it has been widely applied in many fields, there are not systematic and comprehensive theoretical research and mathematic analysis of the operation mechanism, and the improving of the key technologies lack of theoretical basis, which is very unfavorable to the research and development of gene expression programming and evolutionary algorithm. In addition, there are still many problems for the classical machine learning methods in solving classification and complex function finding problems in supervised machine learning field, such as difficult to understand, not very precise and lack of generalization.To bridge the gap, in the paper, we studied the theory of GEP thoroughly. By analyzing the operation mechanism of GEP, using a series of mathematical reasoning, we constructed a theory of GEP schema and presented a building block hypothesis. Through analysis of the convergence of GEP schema and further study of GEP schemas, we obtained the necessary conditions of schema theorem, which conducted to the design of adaptive genetic operators. Finally, some key technologies of GEP, such as the production of chromosome, the chromosome decoding and fitness evaluation had been studied, which will be the mainly part of the Revising Gene Expression Programming (RGEP).In the supervised machine learning, we applied the RGEP to construct the model of supervised machine learning, studied the fitness function and stopping criterion, finally, we provided the experimental results of two types of classification, multi-dimensional classification, and complex function finding which verified the validity of the RGEP model.The main contributions of the paper include:(1) The work proved that the GEP is effective as well as laid a theoretical foundation for further study of GEP. After analysis of genetic algorithms, evolution programming, evolution programming, genetic programming, gene expression programming, we presented a formal definition of GEP. The legitimacy of the individual coding was proposed and the gene schema and GEP schema were also provided. After a series of mathematical reasoning, we presented a GEP schema theorem and building block hypothesis, and proved that the convergence of GEP schema and proposed that elitist strategy of GEP can prevent the loss of the optimal schema. (2) Proposed an adaptive genetic operator algorithm. After studied the survival rate of the better individuals, we presented the requirement of schema theorem and the parameter limit of genetic operator, also, we proposed a adaptive genetic operator algorithm and conduct a experiment for adaptive mutation operator, which shown the algorithm is effective.(3) Proposed a revising gene expression programming. After a fundamental analysis of the GEP, we presented a different structure chromosome generation algorithm and individual quick decoding and evaluating algorithm to improve the population diversity and efficiency. Relevant analysis and experiment were made to verify the validity of the improving, later we proposed a revising gene expression programming based on these two algorithms.(4) Presented an efficient supervised machine learning model. We construct a supervised machine learning model for solving classification and function finding problems based on RGEP. The model used a unique fitness function and cross-validation method to obtain the conditions for early stopping criterion so as to improve the noise immunity and generalization ability. Finally, we verified the validity of the RGEP based model by experiments of monk's problems, detection of micro-calcification in mammogram, wine recognition, complex function finding and quantitative structure activity relationship modeling.