Speeding up Genetic Programming Based Symbolic Regression Using GPUs

Abstract

Symbolic regression has multiple applications in data mining and scientific computing. Genetic Programming (GP) is the mainstream method of solving symbolic regression problems, but its execution speed under large datasets has always been a bottleneck. This paper describes a CUDA-based parallel symbolic regression algorithm that leverages the parallelism of the GPU to speed up the fitness evaluation process in symbolic regression. We make the fitness evaluation step fully performed on the GPU and make use of various GPU hardware resources. We compare training time and regression accuracy between the proposed approach and existing symbolic regression frameworks including gplearn, TensorGP, and KarooGP. The proposed approach is the fastest among all the tested frameworks in both synthetic benchmarks and large-scale benchmarks.

M5GP: Parallel Multidimensional Genetic Programming with Multidimensional Populations for Symbolic Regression

“…Hyperparameter tuning of M5GP should also be expanded, to configure the cuML methods used to estimate the parameters of the linear models, particularly for the regularized regression methods that did not perform well in these tests. Comparisons with methods not currently included in SRBench should also be performed, using different problems and performance criteria, considering, for example, previous iterations of the M5GP approach [9,23] or more recent methods based on different formulations of the SR problem [64][65][66]. Finally, the main bottleneck of the current implementation is the reliance on cuML, which does not allow for an efficient estimation process for multiple models concurrently, and, therefore, population evaluation is sequential, even though it is independent for each individual and could be done using either a parallel or distributed computation.…”

SymFormer: End-to-End Symbolic Regression Using Transformer-Based Architecture

“…Genetic Programming approaches are a traditional way of solving SR [8], [18], [19]. Genetic programming (GP) evolves expressions encoded as trees using selection, crossover, and mutation.…”

Prediction of the Fatigue Strength of Steel Based on Interpretable Machine Learning

“…To enhance the interpretability of the model, this study introduces SR [42][43][44], which establishes mathematical expressions between inputs and outputs. Gplearn [45][46][47], a Python open-source library based on genetic programming (GP), is utilized for SR. The process is as follows: Initially, a population of mathematical expression individuals is randomly generated, typically consisting of basic operators and functions.…”