GeneticAlgorithms.jl

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
end

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
    abcde::Array
    fitness

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

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)
end

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)
end

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)
end

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
        return
    end

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

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]
    end

    child
end

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
end

Run your GA!

using GeneticAlgorithms
require("GeneticAlgorithms/test/equalityga.jl")

model = runga(equalityga; initial_pop_size = 16)

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

songroom2016/GeneticAlgorithms.jl