Improvement And Research On Parallel Of Gene Expression Programming

Automatic programming is one of central goals in computer science. In 1998, renownedcomputer scientist, Professor John Holland lifted the flag of "Genetic programming is automaticprogramming", pushed the research on automatic programming to the high tide. The geneticprogramming is a new branch of evolutionary computation. It operates on the independentcomputer program which expressed by individual of population (It is not fixed length binarycharacter string in GA). This overcomes limitation in expression method of traditional GA anduses more nimble variable hierarchy structure. According to the request of problem, GP useslevel description method to produce programming automatically. So it is one kind "genetic orevolutionary" technology that does not limit in certain sophere. At present, the research on GPhas already permeated into engineering and technology science, life science and every sophere insocial science and got huge result. And now it still develops swiftly and violently.However, GP has some deficiency. The construction of solution is very complex and thereare some restrictions on genetic operation so that genetic operation is correct. According to thedeficiency of GE based on GA and GE Gene Expression Programming(GEP) was first proposedby Candida Ferreira in 2001 which uses gene expression law of biological heredity for referenceand is a kind of computational models discovering knowledge .It has intelligent characteristic ofself-organized, self-adaptation,self-learning what intelligent computation have. Now it hasbecome a powerful tool of automatic programming design. In this new system,the complexcomputer programs(the phenotype )evolved by GEP are totally encoded in simple strings of fixedlength (the genotype, which is beneficial to select, crossover and mutate etc and can ensurecorrect exchange of genotype and phenotype. It overcomes the shortcoming of traditional geneticalgorithm in expression method and traditional genetic programming in complexity ofconstructing solution and complex genetic operation. It has simpler code expression method,more powerful genetic operator (type and quantity of genetic operator have not any limitiation),easier genetic operation and the ability of producing more complex function.At present, GEP has become a hot point problem. With the rapid development of networktechnology, the parallel computation has become a feasible method. So the algorithm research suitable to parallel computation once becomes a hot point of front research. Based on these twohot points, this article mainly develops work on present research condition, principle,insufficiency, improvement and parallel realization of GEP.The main content of this article is as follows:(1) Elaborate present GEP technology, theory and application at home and abroadcomprehensivly and systematically;(2) Research on the core technology and discuss good and bad point of standard GEP;(3) Based on the research of GEP,Improve the gene structure of standard GEP firstly,the gene structure of "head and tail" is changed into "head, body and tail", which isadvantageous to introduce learning mechanism;(4) To improve the convergence speed and the learning ability of GEP,Introduce theEstimation of Distribution Algorithm based on the improved gene structure andpropose Gene Estimation of Distribution Gene Expression Programming (GEDGEP).The results of experiments indicate that this algorithm has higher forest precisionand quicker convergence speed;(5) A new hybrid evolution algorithm based on multi-Layer chromosomes GeneExpression Programming (M-GEP-GA) is proposed. In M-GEP-GA, multi-layerchromosome is introduced in GEP and reusing mechanism of programming isrealized. M-GEP-GA use two levels evolutionary method which uses level modelchromosomes constructing to express individual, uses gene expression programmingoptimize model structure, uses genetic algorithm to optimize model parameters. Thismethod can keep some good structure which may be eliminated from population;these individuals with good structure will get lower fitness due to unproperparameters what they select. So the ability of seeking the superior individual isimproved and this method can avoid the phenomenon of "converge prematurely".(6) Discuss parallel genetic algorithm firstly. And based on GEDGEP.designsynchronized and asynchronous distributed parallel algorithm based on MPI. Theresults of experiments indicate that the speed and quality of solution of these parallelalgorithms are improved and the synchronized distributed parallel algorithm haslinear acceleration rate. We do some experiments and analysis on synchronized andasynchronous parallel algorithm at last.