A CUOA-Based Parallel Algorithm Design And Implementation Of Symbolic Regression

With the rapidly growth of information data, people expect so badly to find a method to extractuseful information and knowledge from these magnanimous amount of data. This experience can bewidely used in business management, production control, market analysis, engineering design andscientific exploration. Symbolic Regression is a promising research aspect in the era of scientificcomputing for data mining. In short, the general purpose of Symbolic Regression is to find anoptimal symbol expression that best fits a target sample set, which has much experimental data.Currently, the mainstream methods to solve the problem of Symbolic Regression are almostbased on genetic programming (GP). For the reason that individuals in GP are extremely difficult toreproduce with modification, we adopt a popular evolutionary algorithm, gene expressionprogramming (GEP), in this paper.The GEP individuals are encoded as fixed length linearchromosomes (genotype), which are then translated into different sized and shaped non-linearentities named expression trees (ETs)(phenotype). GEP inherits the manipulative and expressiveadvantages of GAs and GP. In this paper, we will introduce a hybrid variant of GEP namedMIN-GEP, which improves this traditional GEP algorithm in performance and correctness. In theface of large scale datasets, we select the GPU as parallel computing platform and a newgeneral-purpose parallel programming model, namely Compute Unified Device Architecture(CUDA).We have allocated the task of fitness calculation on the GPU, which is one of majorbottlenecks in terms of execution time for GEP. It makes full use of GPU’s computing performanceand obtains a satisfactory speedup. Meanwhile, in order to improve the individual’s fitness value, wehave embedded Method of Least Square (MLS) into GEP, in which MLS focuses on finding theoptimal constant coefficients locally on the fixed function structure.In the part of experiments, we have selected thirty problems as test cases, which appeared in thepublished literature on constant creation issues in GP or GEP. In the performance experiment, theCUDA-based GEP has almost250speeds up against the serial version.Then we compare MIN-GEPwith other existing well-known constant creation methods in terms of performance and accuracy.TheMLS is far less time consuming than other local optimization methods, and provides a significantimprovement in performance.