One day I thought about fitting a surface to 3 dimensional data points, and solving this via a genetic algorithm seemed fitting. The reason I thought about this is that if one only has real valued features in a dataset one could just fit a polynomial with the number of variable of interest and then perform regression. A standard genetic algortihm is implemented from "Biologically Inspired Optimization Methods: An Introduction" by Mattias Wadhe using python, using roulette wheel selection, crossover and mutation.
An arbitrary bivariate function was constructed which the genetic algorithm will try to find the parameters of.
This surface can be seen as the orange surface in the gif below.
The algorithm got to evolve for 500 generations with a population size of 50 and a mutation probability of 0.02. The results is that it works but it seems like a quite inefficient way to solve this problem. However a real number encoding instead of a binary encoding would probably increase the speed of the algorithm, alternatively writing the decoding functions in a faster language.
