Evolving Neural Network through a Reverse Encoding Tree

News: Our Paper has been accepted to IEEE CEC 2020 for a lecture presentation. An updated version could be found out here. Feel free to contact us for experiement details. Video

Code for Python 3.7 implementation (in the PyCharm) of Reverse Encoding Tree from the paper.

Getting Started

The library is divided into two parts. In the benchmark part, you will easy easily understand the principle of our strategy and its difference from other strategies. In the evolution part, you can use it for many tasks of NeuroEvolution.

We have further integrated neat-python in evolution/bean. The files in the example folder describe how to use the original NEAT to finish the well-accepted tasks. tasks folder includes all the execution documents in the experiments mentioned in the paper.

Prerequisites

neat-python -- version 0.92
gym -- version 0.14.0
box2d -- version 2.3.2
matplotlib -- version 3.1.1
pandas -- version 0.25.1
numpy -- version 1.17.1

Building a Bi-NEAT

We have 6 additional hyper-parameters in the configure.

max_node_num in the network parameters: maximum numnber of node in all the generated neural networks, it describes the range of phenotypic landscape.
init_distance in the Reproduction: initial distance describes the minimum distance between each of the two neural networks in the initial (first) generation.
min_distance in the Reproduction: minimum distance describes the minimum distance between each of the two neural networks after the initial (first) generation.
correlation_rate in the Reproduction: correlation rate describes the demarcation line between positive and negative correlation coefficient. The default value is -0.5. If the correlation coefficient less than correlation rate, it is positive.
search_count in the Reproduction: search count describes the maximum number of searches required when adding a novel neural network.
cluster_method in the Reproduction: Alternative clustering methods, the default is kmeans++. We have kmeans, kmeans++, birch and spectral options.

You need to create a configure before running, the document including original settings is shown in https://readthedocs.org/projects/neat-python/.

After creating the configure:

from neat import population, config, genome, reproduction, species, stagnation

task_config = config.Config(genome.DefaultGenome, reproduction.DefaultReproduction, species.DefaultSpeciesSet, stagnation.DefaultStagnation, "your configure path")
task_population = population.Population(task_config)

Continous Learning Environments

Cartport-v0

LunarLander-v2

If you think this repo helps or being used in your research, please consider refer this paper. Thank you.

Evolving Neural Networks through a Reverse Encoding Tree, Arxiv 2002.00539, IEEE-CEC 2020 Oral.

@inproceedings{zhang2020evolving,
  title={Evolving neural networks through a reverse encoding tree},
  author={Zhang, Haoling and Yang, Chao-Han Huck and Zenil, Hector and Kiani, Narsis A and Shen, Yue and Tegner, Jesper N},
  booktitle={2020 IEEE Congress on Evolutionary Computation (CEC)},
  pages={1--10},
  year={2020},
  organization={IEEE}
}

Haoling Zhang, Chao-Han Huck Yang, Hector Zenil, Narsis A. Kiani, Yue Shen, Jesper N. Tegner

Contributors for this library