I discovered evolutionary algorithms in my teens and was fascinated. After a bit of research, I somehow managed to hack together a neural network in GameMaker, attach it to the brains of a virtual car and watch it crash into walls.

I raced hundreds of these cars against each other. The best ones went on to the next round (with randomly generated modifications) and they all raced again. The outcome of this brutal process was a passable motorist, albeit not one likely to pass their driving test first time.

I was reminded of my obsession recently watching Code Bullet demonstrating things I’d never heard of, such as NEAT and Q learning. Anyway, it got me wanting to implement this stuff again.

So let’s get some definitions out of the way: an evolutionary algorithm is an algorithm inspired by biological evolution. It follows these steps:

  1. Generate a population of random individuals.
  2. Repeat the following:
    1. Assess each individual based on how well it solves a problem (the fitness of the individual).
    2. Breed a new population, based on mutation and crossover of the fittest individuals in the current population.

Mutation refers to randomly altering an individual, such that it behaves slightly differently. Crossover means taking two “parent” individuals and combining them to make a new “child” individual.

These two operations can improve fitness and maintain genetic diversity: mutation introduces diversity, allowing new solutions to be explored. Crossover avoids reducing genetic diversity by combining promising individuals together instead of allowing one fit individual to out-compete another.

To be, or not to be

I’ll use Rust for my examples, but follow along in any language you like! The full code is available on my GitHub.

It’s said that if you put a monkey in front of a typewriter, given enough time it will produce the complete works of Shakespeare.

We haven’t infinite time (or easy access to monkeys!), so let’s try something simpler, reproducing just this fragment from Hamlet:

const HAMLET: &str = "To be, or not to be--that is the question";

Our individuals are going to try their best to write Hamlet. So we’ll represent them with strings:

#[derive(Debug)]
struct Individual(String);

Let’s write out the basic algorithm:

// If you're unfamiliar with Rust, `usize` is just an integer!
const POPULATION_SIZE: usize = 1000;

fn evolve() {
    let mut population = gen_initial_population(POPULATION_SIZE);

    loop {
        population = breed_next_generation(population);
    }
}

Yes, that is an infinite loop! We’ll come back to it later. Next we’ll implement these:

fn gen_initial_population(size: usize) -> Vec<Individual> {
    unimplemented!()
}

fn breed_next_generation(population: Vec<Individual>) -> Vec<Individual> {
    unimplemented!()
}

Generating an initial population

Our initial population can just be random strings with the same length as the target text: I counted and that’s 41 characters!

const LENGTH_HAMLET: usize = HAMLET.len();

fn gen_initial_population(size: usize) -> Vec<Individual> {
    let mut population = Vec::new();

    for _ in 0..size {
        let mut individual = Individual(String::new());
        for _ in 0..LENGTH_HAMLET {
            individual.0.push(rand::random::<char>());
        }
        population.push(individual);
    }

    population
}

To generate random characters, we need to depend on rand:

# Cargo.toml
[dependencies]
rand = "0.7.3"

Let’s generate three individuals:

fn main() {
    dbg!(gen_initial_population(3));
}
$ cargo run
[src/main.rs:4] gen_initial_population(3) = [
    Individual(
        "\u{495d5}\u{990aa}\u{ed152}톘\u{5a426}𭼟...\u{e7849}\u{477d8}",
    ),
    Individual(
        "\u{102f12}\u{760ff}\u{cdffa}\u{ee293}...𘉳\u{3ae05}퍬\u{1a846}",
    ),
    Individual(
        "\u{d78bb}𪶎\u{be97e}\u{41b4f}...\u{820da}\u{1084f8}\u{d4ea7}",
    ),
]

Eww… let’s stick to printable ASCII characters:

use rand::Rng; // for `gen_range` method

fn gen_initial_population(population_size: usize) -> Vec<Individual> {
    ...
				
        for _ in 0..196 {
            individual.0.push(gen_ascii_char());
        }
				
    ...
}

fn gen_ascii_char() -> char {
    // printable ASCII ranges from 32 (space, ' ') to 126 (tilde, '~')
    rand::thread_rng().gen_range(32u8, 126u8) as char
}
$ cargo run
[src/main.rs:4] gen_initial_population(3) = [
    Individual(
        ",J>g\"\\dLPq9E{$JnIAO[;Q)0Zel):fM/Iu!fB)/Tu",
    ),
    Individual(
        "<?kO_gb|.SHw9ZAg=v{gfy4|JF/GG_X&/Sbw\'?PrJ",
    ),
    Individual(
        "Mu2``(w<.IG)cF)o@\\u:Tf f#ZNaOt5\'Vrg;Q$I`#",
    ),
]

Beautiful… kind of!

Crossover and mutation

Now the meaty part, creating the next generation. We need to do three things:

  1. Select the fittest individuals in the current population
  2. Generate new individuals by crossover
  3. Randomly mutate each new individual
fn breed_next_generation(mut population: Vec<Individual>) -> Vec<Individual> {
    select_fittest_individuals(&mut population);
    let mut children = crossover_population(&population);
    mutate_population(&mut children);
    children
}

Let’s select the fittest half of the population:

fn select_fittest_individuals(population: &mut Vec<Individual>) {
    population.sort_by_cached_key(|i| -i.fitness());
    population.truncate(population.len() / 2);
}

Our fitness function is simply “how similar is it to HAMLET”? We can count how many characters match those in HAMLET:

impl Individual {
    fn fitness(&self) -> i32 {
        let my_chars = self.0.chars();
        let target_chars = HAMLET.chars();
        let chars_same = my_chars.zip(target_chars).filter(|(x, y)| x == y);
        chars_same.count() as i32
    }
}

The basics of crossover are, we randomly select two parents and combine them to make a new individual, repeating until we have a full population:

fn crossover_population(parents: &[Individual]) -> Vec<Individual> {
    let mut population = Vec::new();

    use rand::seq::SliceRandom; // for `choose` method
    let mut random = rand::thread_rng();

    for _ in 0..POPULATION_SIZE {
        let parent1 = parents.choose(&mut random).unwrap();
        let parent2 = parents.choose(&mut random).unwrap();
        let child = parent1.crossover(parent2);
        population.push(child);
    }

    population
}

Crossover of individuals is interesting. We can do it in a few ways. I’m going with single-point crossover. We randomly pick an index in the string: everything left of that index comes from the first parent, everything to the right comes from the second. For example:

    CROSSOVER POINT |
                    |
parent1: "hello I am|a string"
parent2: "yo it's me|parent 2"
                    |
child:   "yo it's me|a string"

Here’s our implementation:

impl Individual {
    fn crossover(&self, other: &Individual) -> Individual {
        let index = rand::thread_rng().gen_range(0, LENGTH_HAMLET);

        let left = self.0[0..index].chars();
        let right = other.0[index..LENGTH_HAMLET].chars();
        let child_string = left.chain(right).collect();

        Individual(child_string)
    }
}

Like crossover, there are many ways to do mutation. Let’s pick something simple: each character has a 1% chance to be mutated.

fn mutate_population(population: &mut [Individual]) {
    for individual in population {
        individual.mutate();
    }
}

impl Individual {
    fn mutate(&mut self) {
        let mut random = rand::thread_rng();

        self.0 = self.0.chars().map(|current_char| {
            if random.gen_bool(0.01) {
                gen_ascii_char()
            } else {
                current_char
            }
        }).collect()
    }
}

Unleash the monkey!

Now for some finishing touches… we’ll print the fittest individual of each generation, and add our termination condition:

fn main() {
    evolve();
}

fn evolve()  {
    let mut population: Vec<Individual> = gen_initial_population(POPULATION_SIZE);

    loop {
        population = breed_next_generation(population);

        let fittest = population.iter().max_by_key(|i| i.fitness()).unwrap();
        println!("{}", fittest.0);

        if &fittest.0 == HAMLET {
            return;
        }
    }
}

And we can run it:

$ cargo run
&( R>ngurk>Oug}l V5*KSbE;&(bC&?m7JKe+A[#m
r?nb4?iiY,O<2\pk 1eJe}w6+t6p*.2e`2umRG{Pq
TDbGj%L+Z_><2\pk 1eJe}w6+t6p*.2e`2umRG{Pq
...
To bx,Zor}no+ 9o $0--t ab .smt72 quXst'on
To pe, or.nf# p0 be4-t ab i? 72e quesVion
To Re, or not 9o $e--twQr is`t!l quXsti+n
...
To be, or not to be--t]at is the question
To be, or not Oo be--that is the question
To be, or not to be--that is the question

Pretty quick too!

Is this… useful?

I have no idea! This is my hobby. Certainly, ones as simple as this won’t get you very interesting results, but the great thing about this algorithm is it’s so general.

Just about every part of this algorithm can be changed:

  • The fitness function could put an individual in a complex simulation, or test against a huge amount of training data, or even have individuals directly compete.
  • Selection in each generation can retain the fittest individuals unchanged (elitism).
  • Way too many options to list for mutation or crossover.

Most importantly, what we produced above was a genetic algorithm, these are evolutionary algorithms where individuals are fixed-length strings. But there’s no need to limit yourself to that! Instead of strings, we could encode individuals as:

I’m especially interested in neuroevolution - stay tuned for a post about that in the next decade 😃