Genetic Algorithm

Simple implementation of a genetic algorithm applied into gym environments

Getting Started

This implementation uses a jupyter machine inside a container, so to build and run the container just:

./run_notebook.sh

# Or manually
docker build -t ga-notebook .
docker run --rm -p 8888:8888 -e JUPYTER_ENABLE_LAB=yes -v "$PWD":/home/jovyan/work ga-notebook

Genetic Algorithm implementation

This genetic algorithm has some parameters that can be customized to explore different scenarios:

• population_size: number of genomes you'll have for each generation
• generations: Number of generations you'll run through
• mutation_variance: Percentage of mutation in each weight (so if 0.2 the weight has a probability of 20% of mutating)
• survival_ratio: Keeping the best <survival_ratio> genomes (if 0.1, only the best 0.1 genomes survive for the next generation)
• both_parent_percentage: Percentage of genomes that will come from crossover between two parents
• one_parent_percentage: Percentage of genomes that will come from one parent only (will not suffer from crossover, only from mutation)
• episodes: how many episodes it will run before extracting the fitness of the genomes

Results

The algorithm is surprisingly fast at converging, very different comparing to its backpropagation peers.

Results using the following parameters:

population_size 64
generations 200
mutation_variance 0.1
survival_ratio 0.1
both_parent_percentage 0.8
one_parent_percentage 0.1
episodes 15