Skip to content

This repository is the official implementation of the Hybrid Self-Attention NEAT algorithm. It contains the code to reproduce the results presented in the original paper: https://link.springer.com/article/10.1007/s12530-023-09510-3

License

Notifications You must be signed in to change notification settings

SamanKhamesian/Hybrid-Self-Attention-NEAT

Repository files navigation

Hybrid-Self-Attention-NEAT

Abstract

This repository contains the code to reproduce the results presented in the original paper.
In this article, we present a “Hybrid Self-Attention NEAT” method to improve the original NeuroEvolution of Augmenting Topologies (NEAT) algorithm in high-dimensional inputs. Although the NEAT algorithm has shown a significant result in different challenging tasks, as input representations are high dimensional, it cannot create a well-tuned network. Our study addresses this limitation by using self-attention as an indirect encoding method to select the most important parts of the input. In addition, we improve its overall performance with the help of a hybrid method to evolve the final network weights. The main conclusion is that Hybrid Self-Attention NEAT can eliminate the restriction of the original NEAT. The results indicate that in comparison with evolutionary algorithms, our model can get comparable scores in Atari games with raw pixels input with a much lower number of parameters.

NOTE: The original implementation of self-attention for atari-games, and the NEAT algorithm can be found here:

Neuroevolution of Self-Interpretable Agents: https://github.com/google/brain-tokyo-workshop/tree/master/AttentionAgent
Neat Package: https://github.com/CodeReclaimers/neat-python
Pure python library for the NEAT and other variations: https://github.com/ukuleleplayer/pureples

Execution

To use this work on your researches or projects you need:

  • Python 3.7
  • Python packages listed in requirements.txt

NOTE: The following commands are based on Ubuntu 20.04

To install Python:

First, check if you already have it installed or not.

python3 --version

If you don't have python 3.7 in your computer you can use the code below:

sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get install python3.7
sudo apt install python3.7-distutils

NOTE: To create a virtual environment, you can use the following link:
Creation of virtual environment: https://docs.python.org/3.7/library/venv.html

To install packages via pip install:

python3.7 -m pip install -r requirements.txt

To run this project on Ubuntu server:

You need to uncomment the following lines in experiments/configs/configs.py

_display = pyvirtualdisplay.Display(visible=False, size=(1400, 900))
_display.start()

And also install some system dependencies as well

apt-get install -y xvfb x11-utils

To train the model:

  • First, check the configuration you need. The default ones are listed in experiments/configs/.
  • We highly recommend increasing the number of population size, and the number of iterations to get better results.
  • Check the working directory to be: ~/Hybrid_Self_Attention_NEAT/
  • Run the runner.py as below:
python3.7 -m experiment.runner

NOTE: If you have limited resources (like RAM), you should decrease the number of iterations and instead use loops command (n is the number of iterations)

for i in {1..n}; do python3.7 -m experiment.runner; done

To tune the model:

  • First, check you trained the model, and the model successfully saved in experiments/ as main_model.pkl
  • Run the tunner.py as below:
python3.7 -m experiment.tunner

NOTE: If you have limited resources (like RAM), you should decrease the number of iterations and instead use loops command (n is the number of iterations)

for i in {1..n}; do python3.7 -m experiment.tunner; done

Citation

For attribution in academic contexts, please cite this work as:

Khamesian, S., Malek, H. Hybrid self-attention NEAT: a novel evolutionary self-attention approach to improve the NEAT algorithm in high dimensional inputs. Evolving Systems (2023). https://doi.org/10.1007/s12530-023-09510-3

About

This repository is the official implementation of the Hybrid Self-Attention NEAT algorithm. It contains the code to reproduce the results presented in the original paper: https://link.springer.com/article/10.1007/s12530-023-09510-3

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages