Simple Genetic Programming

For Symbolic Regression

This Python 3 code is a simple implementation of genetic programming for symbolic regression, and has been developed for educational purposes.

Dependencies

Numpy. The file test.py shows an example of usage, and requires Scikit-learn.

Installation

You can install it with pip using python3 -m pip install --user simplegp, or locally by downloading the code and running python3 setup.py install --user.

Random

Create RelU node

Experiments

Measure

Number of generations
Time taken
Maximum Fitness
Some measure of diversity

Optimise general params first for Symbolic Regression

Different initial population sizes
Different mutation probabilities
Tournament selection size
Subtree Size

Optimise the gradient descent local search

Uniformly randomly select individuals to do gradient descent
Top k% of nodes (Implementation could be tricky because it requires another evalutaiont round)
When to apply graident descent (every x generations)
Different graident descent learning rates for different generations, maybe course vs fine grained
Combinations of above (Just pass the corresponding parameters in defining backprop function)
Iterations of gradient descent, 2 levels, lower for everyone, and then higher for top k

Reoptimise solution with optimal graident descent params

Questions For Marco

What sort of parameters should we test for in the general GP (not including gradient descent)? Do we even need to optimise these?

cseale/SimpleGP