michaelwayman/python-genetic

Genetic Mutation

How to use the process of evolution and natural selection to solve problems.

What you will find here is original content, my take on natural selection and genetic mutation that you won't find anywhere else. In the example you can see it solve a real world 4.29x10^15 choices np-complete maximization problem.

In computer science, a genetic algorithm (GA) is a metaheuristic inspired by the process of natural selection that belongs to the larger class of evolutionary algorithms (EA). Genetic algorithms are commonly used to generate high-quality solutions to optimization and search problems by relying on bio-inspired operators such as mutation, crossover and selection. -- Wikipedia

Usage

3 easy steps:

subclass Evolvable to model your problem
create your gene_pool
Use the Evolve class to simulate the evolution and natural selection over n number of generations.

First

Subclass Evolvable in a way that models the problem you are trying to solve. There are only 2 simple functions you need to implement.

class MyEvolvable(Evolvable):

    def can_survive(self):
        # TODO: Return a boolean

    def fitness_level(self):
        # TODO: Return a number (bigger means better)

Second

Create a gene_pool that maps all your genes to their corresponding alleles. For the sake of simplicity, a Gene is just an idea, like 'hair color'. Whereas, 'alleles', are the actual mutation/implementation of a gene.

For example,

gene_pool = {
    'hair_color': ['blonde', 'brunette', 'black'],
    'height': ['short', 'tall', 'fun-size'],
    'weight': ['skinny', 'normal', 'overweight'],
    ...
}

Third

Perform the natural selection on your Evolvable using your gene_pool.

e = Evolve(gene_pool, MyEvolvable)
e.run(n=75000)

# 75,000 generations later, e.best will have our best answers.
print(str(x) for x in e.best))

The classes

There are only 2 classes that you need to familiarize yourself with to make GA/EA algorithms work.

Evolvables

class Evolvable: - represents 2 things: something that can "evolve", and a solution to a particular problem

An evolvable has a few pieces of information:

genes representation of the solution
can_survive something to tell us if this is a valid solution or not
fitness_level a number that tells us how good of a solution this is
unique some hashable representation to uniquely identify this solution (maybe a serialization of the genes?) it could be anything

Evolve

class Evolve: - Class that takes a type of Evolvable and simulates the "natural selection" process.

The important pieces of the Evolve class:

gene_pool all the different genes and alleles available to create Evolvable instances with.
best list of the best Evolvables throughout the generations
population the current generation of evolvables
cross_over a function that takes 2 or more evolvable instances (parents) and creates new evolvables from them (children).
mutate a function that takes an evolvable and switches a couple of the alleles with alleles from the gene pool.

Background

I had a notion that I could make a lot of $$ money $$ by creating an algorithm that could win at fantasy sports (gambling on fantasy sports was just made legal in NY, I was pretty excited)

Depending on the night, there could be up to 4.5 quadrillion (4.5x10^15) possible combinations of teams.

After doing some estimations, looking at all these combinations would take days to weeks to complete.

I came up with this genetic mutation algorithm and was able to calculate the best combination of players in less than a minute :)

Gotchas

There is no big-Oh notation or O(n^2) way of describing the performance of this algorithm
- In fact, we actually give the algorithm a stopping point ourselves by saying "evolve 50,000,000 generations and then quit"
If the gene_pool isn't big enough, it's possible to run into an infinite loop during natural selection.
- Basically, if we are unable to create n_children unique children from the gene_pool, we will run into an infinite loop. It would be easy to throw an error in such situations, but really, just make sure your gene_pool is large enough to create a large population.
There is no guarantee that the next generation will be better than the previous. It is very possible to evolve from something terrible, to something great, back to something terrible. This is why it is important to keep track of the best Evolvables throughout the generations.
It's possible to never come across the correct answer. There is so much randomness involved in this algorithm that sometimes it feels like we are just picking a random answer. But the results speak for themselves and it turns out this actually works, quite well.

michaelwayman / python-genetic