evolutionary-optimization

A generic evolutionary algorithm for function optimisation.


Keywords
artificial-intelligence, evolutionary-algorithms, machine-learning-algorithms, machinelearning-python
Install
pip install evolutionary-optimization==3.0.0

Documentation

CircleCI CircleCI

Evolutionary-Optimisation

A generic evolutionary algorithm for function optimisation.

Introduction

This package allows the user to optimise a function using an evolutionary algorithm. An evolutionary algorithm uses the principles of evolution to find optimal solutions. At a high level diagram of the process is as follows:

flowchart TD;

    A[Generate population] --> B[Evaluate population];
    B --> C{Is exit condition met?};
    C --No--> D[Update population];
    C --Yes--> E[Exit algorithm];
    D --> B

Currently, the only exit condition for the code is running the desired number of generations.

Using the Package

Getting Started

To get started with this package clone this repo:

git clone https://github.com/mwolinska/Evolutionary-Optimisation

Then enter the correct directory on your machine:

cd Evolutionary-Optimisation

This package uses poetry dependency manager. To install all dependencies run:

poetry install

Running Experiments

To run the code type the following in your terminal. The default experiment is a simple optimization of the $x^{2}$ using integers.

run_evolution

The parameters used for the run can be edited within the main.py file.

    genotype_class = Genotype.get_genotype(Genotypes.INTEGER_LIST)
    phenotype_class = Phenotype.get_phenotype(Phenotypes.PARABOLA)

    evolutionary_algorithm = Evolution(
        phenotype=phenotype_class(genotype_class()),
        number_of_individuals=10,
        number_of_generations=5,
        fitness_function=FitnessFunctions.MINIMIZE,
        ratio_of_elite_individuals=0.1
    )

This will produce the following output:

The value of the best individual is [1]

And the graph of fitness over time will be generated:

Personalising Experiments

To personalise your experiment you can either use the prebuilt phenotypes and genotypes using our interface, or you can build your own. To do so, you simply need to create a new phenotype / genotype class that inherits from the corresponding abstract class and implement the methods to suit your needs.

The classes structure is outlined in the diagram below

The implemented fitness functions are outlined in the diagram below (please note ApproachValueFitnessFunction is not currently operational.)

Documentation

You can consult our documentation here.