A lightweight framework for writing genetic algorithms in Julia
Clone or download
WestleyArgentum Merge pull request #18 from Ayush-iitkgp/master
Update README.md to render markdown properly
Latest commit 7c9074c Jun 16, 2017
Type Name Latest commit message Commit time
Failed to load latest commit information.
src fix sizehint and require deprecations Dec 22, 2016
test change include to work in parallel Dec 23, 2016
.travis.yml use runtests in travis May 14, 2014
CONTRIBUTING.md fix links May 12, 2016
LICENSE.md contributors Dec 16, 2013
README.md Update README.md Jun 6, 2017



Build Status

This is a lightweight framework that simplifies the process of creating genetic algorithms and running them in parallel.

Basic Usage

What's your problem???

Let's say you've got a simple equality a + 2b + 3c + 4d + 5e = 42 that you'd like come up with a solution for.

Create a Module

Start by creating a file and a module for your ga. Your module will be loaded into the framework and things inside it will be used to run your algroithm.

module equalityga
    # implement the GA interface inside here

Define an Entity

Your entity should inherit from the abstract GeneticAlgorithms.Entity. The framework will look for a create_entity function and will use it to create an initial population.

type EqualityMonster <: Entity

    EqualityMonster() = new(Array(Int, 5), nothing)
    EqualityMonster(abcde) = new(abcde, nothing)

function create_entity(num)
    # for simplicity sake, let's limit the values for abcde to be integers in [-42, 42]
    EqualityMonster(rand(Int, 5) % 43)

Note that EqualityMonster has a field fitness. By default this field will be used by the framework to store the entities calculated fitness, so that you have access to it elsewhere in your GA. If you'd like to change the behavior you can overload fitness!(entity::EqualityMonster, score).

Create a Fitness Function

The framework will expect a fitness function. It should take in a single entity and return a fitness score.

function fitness(ent)
    # we want the expression `a + 2b + 3c + 4d + 5e - 42`
    # to be as close to 0 as possible
    score = ent.abcde[1] +
            2 * ent.abcde[2] +
            3 * ent.abcde[3] +
            4 * ent.abcde[4] +
            5 * ent.abcde[5]

    abs(score - 42)

Note that isless(l::Entity, r::Entity) will return l.fitness < r.fitness, but that in this case entities with scores closer to 0 are doing better. So we should define a specialized isless.

function isless(lhs::EqualityMonster, rhs::EqualityMonster)
    abs(lhs.fitness) > abs(rhs.fitness)

Group Entities

group_entities operates on a population (an array of entities sorted by score) and will be run as a task and expected to emit groups of entities that will be passed into a crossover function. group_entitites also provides a nice way to terminate the GA; if you want to stop, simply produce no groups.

function group_entities(pop)
    if pop[1].fitness == 0

    # simple naive groupings that pair the best entitiy with every other
    for i in 1:length(pop)
        produce([1, i])

Define Crossover

crossover should take a group of parents and produce a new child entity. In our case we'll just grab properties from random parents.

function crossover(group)
    child = EqualityMonster()

    # grab each element from a random parent
    num_parents = length(group)
    for i in 1:length(group[1].abcde)
        parent = (rand(Uint) % num_parents) + 1
        child.abcde[i] = group[parent].abcde[i]


Define Mutation

mutate operates on a single entity and is responsible for deciding whether or not to actually mutate.

function mutate(ent)
    # let's go crazy and mutate 20% of the time
    rand(Float64) < 0.8 && return

    rand_element = rand(Uint) % 5 + 1
    ent.abcde[rand_element] = rand(Int) % 43

Run your GA!

using GeneticAlgorithms

model = runga(equalityga; initial_pop_size = 16)

population(model)  # the the latest population when the GA exited