Use of Surrogate Models

[unaiLopez]

Genetic algorithms are extremely powerful for symbolic regression, but have you considered a hybrid approach using surrogate ML or deep learning models to quickly estimate the potential of candidate trees? This could drastically reduce the number of full evaluations required, accelerating convergence, especially on large datasets. The surrogate could be used to pre-select the most promising trees before performing crossover, focusing computational resources on candidates with higher likelihood of success.

[MilesCranmer]

Evaluating an expression on an dataset of 100 examples only takes about 1us. The main bottleneck is not the evaluation speed but really the vast combinatorial space of possibilities. Its such a massive space that it is actually intractable to learn a compressed general representation.