Artificial Neural Network Learning via Genetic Algorithm

Overview

This project simulates learning a neural network using a genetic algorithm rather than the conventional error backpropagation method. Experiment with this using the Cart Pole provided by the open-ai gym library. We confirmed that simple fully connected networks evolve well enough in the environment of tournament selection, 1000 individuals per population, and 0.05 mutation probability.

Installation

$ git clone https://github.com/leibniz21c/Artificial-Neural-Network-Learning-via-Genetic-Algorithm
$ cd Artificial-Neural-Network-Learning-via-Genetic-Algorithm/
$ pip install -r requirements.txt

Training

$ python train.py [OPTIONS]

Argument	Type	Default	Description
--num_populations	int	10	Number of individuals per population
--p_mutation	float	0.05	Probability of mutation for each individual
--max_generations	int	10	Max iteration
--selection_method	str	tournament	Selection method. (Available : wheel, tournament)
--num_validations	int	1	Number of simulations per individual

tournament use every candidates.

Test

$ python test.py [SAVED_MODEL_PATH]

Default trained models are in saved/

Simulations

1. Best Model in Generation 1

2. Best Model in Generation 10

3. Best Model in Generation 100

4. Best Model