Advanced genetic and evolutionary algorithm library written in TypeScript. Special thanks to Sub Protocol for writing the intial JavaScript version.
The existing Javascript GA/EP library landscape could collectively be summed up as, meh. All that I required to take over the world was a lightweight, performant, feature-rich, nodejs + browser compatible, unit tested, and easily hackable GA/EP library. Seamless Web Worker support would be the icing on my cake.
Until now, no such thing existed. Now you can have my cake, and optimize it too. Is it perfect? Probably. Regardless, this library is my gift to you.
Have fun optimizing all your optimizations!
npm install @glavin001/genetic-jsThe genetic-js interface exposes a few simple concepts and primitives, you just fill in the details/features you want to use.
| Function | Return Type | Required | Description |
|---|---|---|---|
| seed() | Individual | Yes | Called to create an individual, can be of any type (int, float, string, array, object) |
| fitness(individual) | Float | Yes | Computes a fitness score for an individual |
| mutate(individual) | Individual | Optional | Called when an individual has been selected for mutation |
| crossover(mother, father) | [Son, Daughter] | Optional | Called when two individuals are selected for mating. Two children should always returned |
| optimize(fitness, fitness) | Boolean | Yes | Determines if the first fitness score is better than the second. See Optimizer section below |
| select1(population) | Individual | Yes | See Selection section below |
| select2(population) | Individual | Optional | Selects a pair of individuals from a population. Selection |
| shouldContinue(pop, gen, stats) | Boolean | Optional | Called for each generation. Return false to terminate end algorithm (ie- if goal state is reached) |
| notification(pop, gen, stats, isFinished) | Void | Optional | Runs in the calling context. All functions other than this one are run in a web worker. |
The optimizer specifies how to rank individuals against each other based on an arbitrary fitness score. For example, minimizing the sum of squared error for a regression curve Genetic.Optimize.Minimize would be used, as a smaller fitness score is indicative of better fit.
| Optimizer | Description |
|---|---|
| Genetic.Optimize.Minimizer | The smaller fitness score of two individuals is best |
| Genetic.Optimize.Maximizer | The greater fitness score of two individuals is best |
An algorithm can be either genetic or evolutionary depending on which selection operations are used. An algorithm is evolutionary if it only uses a Single (select1) operator. If both Single and Pair-wise operations are used (and if crossover is implemented) it is genetic.
| Select Type | Required | Description |
|---|---|---|
| select1 (Single) | Yes | Selects a single individual for survival from a population |
| select2 (Pair-wise) | Optional | Selects two individuals from a population for mating/crossover |
| Single Selectors | Description |
|---|---|
| Genetic.Select1.Tournament2 | Fittest of two random individuals |
| Genetic.Select1.Tournament3 | Fittest of three random individuals |
| Genetic.Select1.Fittest | Always selects the Fittest individual |
| Genetic.Select1.Random | Randomly selects an individual |
| Genetic.Select1.RandomLinearRank | Select random individual where probability is a linear function of rank |
| Genetic.Select1.Sequential | Sequentially selects an individual |
| Pair-wise Selectors | Description |
|---|---|
| Genetic.Select2.Tournament2 | Pairs two individuals, each the best from a random pair |
| Genetic.Select2.Tournament3 | Pairs two individuals, each the best from a random triplett |
| Genetic.Select2.Random | Randomly pairs two individuals |
| Genetic.Select2.RandomLinearRank | Pairs two individuals, each randomly selected from a linear rank |
| Genetic.Select2.Sequential | Selects adjacent pairs |
| Genetic.Select2.FittestRandom | Pairs the most fit individual with random individuals |
import Genetic from "@glavin001/genetic-js";
//
type Entity = string;
type UserData = {
solution: string;
};
// Extend the abstract class Genetic.Genetic
class CustomGenetic extends Genetic.Genetic<Entity, UserData> {
// more likely allows the most fit individuals to survive between generations
public select1 = Genetic.Select1.RandomLinearRank;
// always mates the most fit individual with random individuals
public select2 = Genetic.Select2.FittestRandom;
// ...
public notification({
population: pop,
isFinished,
}: {
population: Population<Entity>;
generation: number;
stats: Stats;
isFinished: boolean;
}) {
if (isFinished) {
console.log(`Solution is ${pop[0].entity} (expected ${this.userData.solution})`);
}
}
}
// ...
const userData: UserData = {
solution: "thisisthesolution",
};
const config: Partial<Genetic.Configuration> = {
crossover: 0.4,
iterations: 2000,
mutation: 0.3,
size: 20,
};
// ...
const genetic = new CustomGenetic(config, userData);
genetic.evolve();| Parameter | Default | Range/Type | Description |
|---|---|---|---|
| size | 250 | Real Number | Population size |
| crossover | 0.9 | [0.0, 1.0] | Probability of crossover |
| mutation | 0.2 | [0.0, 1.0] | Probability of mutation |
| iterations | 100 | Real Number | Maximum number of iterations before finishing |
| fittestAlwaysSurvives | true | Boolean | Prevents losing the best fit between generations |
| maxResults | 100 | Real Number | The maximum number of best-fit results that webworkers will send per notification |
| webWorkers | true | Boolean | Use Web Workers (when available) |
| skip | 0 | Real Number | Setting this higher throttles back how frequently genetic.notification gets called in the main thread. |
Feel free to open issues and send pull-requests.