A Tensor-Based Genetic Programming Framework for Symbolic Regression on Structured Domains

Genetic Programming is an evolutionary method for finding symbolic models that fit data, a task called symbolic regression. GP typically requires substantial computational resources, since every candidate program must be tested on many data points. This work introduces a new GP approach that uses tensor algebra and parallel hardware to efficiently represent and evolve mathematical expressions, focusing on structured and high-dimensional data. Experimental comparisons with a standard GP system are conducted on several symbolic regression benchmarks. The new method matches or outperforms traditional GP in terms of accuracy and convergence, and achieves significant gains in runtime, especially for large datasets. The analysis includes a discussion of the method’s scalability, the advantages for large-scale problems, and considerations related to overhead and memory use. The proposed approach allows symbolic regression tasks to be handled more efficiently and at larger scales, supporting new applications of GP to structured data domains.