ddehueck/pytorch-neat

Evolving Activation function choices

Opened this issue · 3 comments

Is your feature request related to a problem? Please describe.
I would like to see if we can evolve activation function choices like in the original Python NEAT

Describe the solution you'd like
A setting in the config for activation takes in the value 'random' and then evolves the activation function from a set of activations available.

This makes sense to do - these are that will need to be changed:

  • neat/phenotype/feed_forward.py - will need to use the activation from each unit not directly from config
  • neat/genotype/node_gene.py - will need to hold a value for the activation function
  • neat/genotype/genome.py - may need to add a mutation rate for activation functions
  • neat/activations.py - will need to add a way to randomly select an activation function, should also add more activation functions

Thanks for the response! I was waiting to see if this was already in the works. Thanks for letting me know which files need changes. I can also start working on these.

@GautamSridhar That would be great! It would be a couple of days until I could get some good work done on this, but I am available to answer questions!