A public python implementation of the DeepHyperNEAT system for evolving neural networks. Developed by Felix Sosa and Kenneth Stanley. See paper here: https://eplex.cs.ucf.edu/papers/sosa_ugrad_report18.pdf
NOTE: This implementation is under development. Updates will be pushed over time, bringing in new functionality, tests, and various other elements. The purpose of this repo is to allow others to have a codebase to understand, use, or improve upon DeepHyperNEAT.
To run DHN in its current form, you need to create a task file. For reference, see xor_study.py.
This task file must contain:
rom genome import Genome # Genome class
rom population import Population # Population class
rom phenomes import FeedForwardCPPN # CPPN class
rom decode import decode # Decoder for CPPN -> Substrate
rom visualize import draw_net # optional, for visualizing networks
ub_in_dims = [1,2] # Is of type list
ub_sh_dims = [1,3] # Is of type list
ub_o_dims = 1 # Is of type integer
op_key = 0 # Key for population
op_size = 150
op_elitism = 2 # Number of members of pop to keep each generation
ef task(genomes):
task_inputs = [1,2,3]
expected_outputs = [2,4,6]
for key, genome in genomes:
cppn = CPPN.create(genome) # Create cppn from genome
substrate = decode(cppn,sub_in_dims,sub_o_dims,sub_sh_dims) # Decode cppn into substrate
error = 0.0 # Initialize error for current genome
for inputs, expected in zip(xor_inputs, expected_outputs):
inputs = inputs + (1.0,) # Append inputs with bias value
actual_output = substrate.activate(inputs)[0] # Query substrate
error += error_func(actual_output,expected) # Evaluate error
genome.fitness = 1.0 - error # Assign fitness
op = Population(pop_key, pop_size, pop_elitism)
olution = pop.run(task,num_generations) # Returns the solution to the task
These modules are associated with the primary function of the DeepHyperNEAT (DHN) algorihtm.
Contains all functionality of the genome, a Compositional Pattern Producing Network (CPPN) and its mutation operators.
Contains multiple representations for feed-forward and recurrent neural networks for the CPPN and the Substrate.
Contains all functionality and information of the populations used in DHN.
A library of activation functions that can be used for the CPPN and Substrate.
Contains all functionality needed for the reproductive behavior in DHN.
Contains all functionality needed for speciation in DHN.
Contains all functionality needed for stagnation schemes used in speciation.
Contains all functionality needed to decode a given CPPN into a Substrate.
These modules are intended for secondary functionality such as reporting evolutionary statistics, visualizing the CPPN and Substrate, and various utility functions used throughout the primary modules.
Contains various functions for reporting evolutionary statistics during and after an evolutionary run.
Contains functions for visualizing a CPPN or Substrate.
Contains common functions and iterators used throughout DHN.