An object-oriented genetic algorithm framework with support for asynchronous operations at any stage of the algorithm. It also unblocks execution between generations, allowing CPU-intensive genetic algorithms to be used as part of user-facing API's.
Check out the api docs or use the following guide to get started.
Start by creating a chromosome class and implementing the required methods:
const { Chromosome } = require('gene-lib');
class MyChromosome extends Chromosome {
constructor(...args) {
super();
// Initialize MyChromosome instance, as needed.
}
static create(...args) {
// Return a new instance of MyChromosome for the first generation.
// Argument(s) can be specified in settings later.
}
getFitness() {
// Return a number representing the chromosome's fitness.
// You must always implement this method.
}
crossover(other) {
// Return an array of children based on some crossover with other.
// You must implement this if and only if you set the crossover rate,
// or if the manualCrossoverCheck setting is true.
}
mutate(rate) {
// Return a mutated copy, based on the provided rate.
// You must implement this if and only if you set the mutation rate.
}
}
module.exports = MyChromosome;
Once your Chromosome class is created, you can easily run the GA using the
geneLib::run
method. The following will run 1000 generations of 100
individuals, with a crossover rate of 0.2 and a mutation rate of 0.05. The
selection method defaults to deterministic binary tournament:
const geneLib = require('gene-lib');
const MyChromosome = require('./path/to/my-chromosome');
geneLib.run({
chromosomeClass: MyChromosome,
generationSize: 100,
generationLimit: 1000,
createArgs: [ 'foo', 'bar' ],
crossoverRate: 0.2,
mutationRate: 0.05
})
.then((result) => {
console.log(result);
/*
{
generationCount: Number of generations that were run.
best: {
fitness: Highest fitness produced by the GA,
chromosome: Best MyChromosome instance produced by the GA
},
individuals: [
Array of all individuals in final generation, each
with the same properties as `best` above.
]
}
*/
});
The ::run
method will place processing for each generation at the end of the
event loop queue. This helps prevent it from blocking execution for too long,
which you definitely want in most cases. See the first entry in
this article
for a great explanation of why this is.
- chromosomeClass: This is your chromosome class constructor. It need not
actually inherit from
gene-lib
'sChromosome
class, but it must implement all required methods shown above, including the staticcreate
method. - createChromosome: As an alternative to chromosomeClass, you may specify
a factory function that simply returns a chromosome object each time it is
invoked. These objects must implement all required instance methods, including
getFitness
and potentiallycrossover
and/ormutate
. - createArg: Can be used to specify an argument for the
create
method orcreateChromosome
function, if any. - createArgs: If you want multiple create arguments, use this instead of
createArg
and specify them as an array. - selector: Specify the selection method, as a string. Defaults to
'tournament', indicating tournament selection. Also available is 'roulette'
for fitness-proportional selection. Custom selectors can be added using the
geneLib::registerSelector
method, described later in this README. - selectorClass: Instead of using the selector option, you may specify a custom selector class directly here.
- selectorSettings: Specify options specific to your selection method, as
a nested object. Check the documentation for your selector for more info.
In the case of tournament selection, you can specify the following:
- tournamentSize: Number of individuals selected for the tournament. Defaults to 2, for binary tournament.
- baseWeight: Probability of selecting the top individual in the tournament. Defaults to 1, for a deterministic tournament.
- generationSize: Specifies the number of individuals in each generation. This setting must be provided, and it must be a positive integer that is also a multiple of the childCount.
- generationLimit: Specifies the maximum number of generations to run. Defaults to infinity, meaning the algorithm will not stop until a solution is found, as specified by the solutionFitness setting.
- solutionFitness: If an individual ever meets or exceeds this fitness, the algorithm will stop and return that individual, even if the generationLimit has not yet been reached. Defaults to Infinity. Set to false to disable this behavior completely.
- crossoverRate: The fraction of individuals in each generation that should be produced through crossover operations, on average. Defaults to zero. You'll want to set either this, the mutationRate, or both, otherwise generations will not evolve and your GA will do nothing useful.
- manualCrossoverCheck: Set to true to bypass internal rate-checking of crossover operations. See 'Manual Crossover Checks' below for more information.
- parentCount: Can be used to specify the number of parents per crossover operation. Must be an integer greater than 1. Defaults to 2. See 'Unusual Crossovers' below for more information.
- childCount: Can be used to specify the number of children per crossover operation. Must be a positive integer that is also a factor of the generationSize. Defaults to 2. See 'Unusual Crossovers' below for more information.
- mutationRate: Fractional rate of mutation. Defaults to zero. You'll want to set either this, the crossoverRate, or both, otherwise generations will not evolve and your GA will do nothing useful.
- async: Set to specify asynchronous operation. Set any of its properties
to true to cause that operation to become asynchronous. Set a property to a
number to specify concurrency, which is otherwise assumed to be 1. See the
'Asynchronous Operations' section below for more information.
- async.add: Setting for selector
#add
operations. - async.selector: Setting for selector
#select
operations. - async.create: Setting for chromosome
::create
operations. - async.getFitness: Setting for chromosome
#getFitness
operations. - async.crossover: Setting for chromosome
#crossover
operations. - async.mutate: Setting for chromosome
#mutate
operations.
- async.add: Setting for selector
- onGeneration: A function that will be invoked after each generation, including the 0-th generation. Receives an argument indicating the current state of the algorithm, in the same format as the final result, described above.
Some of the settings above are directly tied to your chromosome and/or selector
implementations. For these, it could make more sense to specify the settings
on the chromosome or selector classes themselves. To do this, simply have a
static settings
property on the appropriate class. You can do this with a
getter:
const { Chromosome } = require('gene-lib')
class MyChromosome extends Chromosome {
static get settings() {
return {
crossoverRate: 0.2,
mutationRate: 0.05
};
}
// Other chromosome methods...
}
module.exports = MyChromosome;
or by simply assigning it:
const { Selector } = require('gene-lib');
class MySelector extends Selector {
// Implement selector here...
}
MySelector.settings = {
defaults: { foo: 'bar' }
};
module.exports = MySelector;
Settings provided directly to the ::run
method will always take priority over
class-level settings, and not all settings are supported on the class level.
Supported class-level settings for selectors include:
- defaults: Will become defaults applied to
selectorSettings
. - async.add
- async.select
Supported class-level settings for chromosomes include:
- solutionFitness
- crossoverRate
- manualCrossoverCheck
- parentCount
- childCount
- mutationRate
- async.create
- async.getFitness
- async.crossover
- async.mutate
As a convenience, gene-lib
performs checks against the crossover rate
internally. Normally, if two parents are selected and are determined to not
cross over with each other, they are simply copied into the next generation
without invoking the crossover
method at all. This is what allows you to
skip implementing the crossover
method if your crossover rate is zero.
This is great for most genetic algorithms, where crossovers either do or don't happen at the chromosome level. Sometimes, however, it's useful to break your chromosomes down into component genes, each of which may or may not cross over with its counterpart gene in another chromosome.
One such GA could be used to solve the ordered clustered traveling salesman problem. The path through each segment would be a gene, and crossover rate checks would need to occur at the gene level rather than the chromosome level.
To do this, set the manualCrossoverCheck
option to true. This will cause the
crossover
method to be invoked for every selected set of parents. The
crossoverRate will be passed to the crossover
method as its last argument.
You should do your rate checks against this argument in order to maintain the
ability to easily tweak the crossover rate through geneLib::run
settings:
const { Chromosome } = require('gene-lib');
class ManualChromosome extends Chromosome {
create() {
// Return a new instance, as usual.
}
getFitness() {
// Return fitness, as usual.
}
crossover(other, rate) {
// Check each gene for crossover individually while creating children.
}
}
module.exports = ManualChromosome;
const geneLib = require('gene-lib');
const ManualChromosome = require('./path/to/manual-chromosome');
geneLib.run({
chromosomeClass: ManualChromosome
generationSize: 100
generationLimit: 1000,
crossoverRate: 0.5,
manualCrossoverCheck: true
})
.then((result) => {
// ...
});
For most GA's, you'll have exactly two parents and exactly two chidren per
crossover. Just in case you need it, however, gene-lib
provides you with the
ability to change either of these numbers using the parentCount and childCount
settings.
The parentCount must always be greater than one, and the childCount must always be positive. In addition, the childCount must be a factor of the generationSize. Otherwise, it would not be possible to create a generation of that size through crossover operations.
Even though these are configurable, your crossover method must always accept the same number of parents and return the same number of children.
The following example does crossovers with three parents and one child:
const { Chromosome } = require('gene-lib');
class WeirdChromosome extends Chromosome {
create() {
// Return a new instance, as usual.
}
getFitness() {
// Return fitness, as usual.
}
crossover(a, b) {
// Return a crossover of this, a, and b, producing only one child.
// You may return this child by itself, without wrapping it in an array.
}
}
module.exports = WeirdChromosome;
const geneLib = require('gene-lib');
const WeirdChromosome = require('./path/to/weird-chromosome');
geneLib.run({
chromosomeClass: WeirdChromosome
generationSize: 100
generationLimit: 1000,
crossoverRate: 0.7,
parentCount: 3,
childCount: 1
})
.then((result) => {
// ...
});
Normally, the #getFitness
method is only called once on each chromosome.
Its result is stored internally as part of the ::run
method, so you don't
need to worry about caching it yourself. In some cases, however, you may want
access to the fitness in the other chromosome methods. As a convenience for
these cases, gene-lib
provides the CachingChromosome
base class:
const { CachingChromosome } = require('gene-lib');
const _ = require('lodash');
class MyCachingChromosome extends CachingChromosome {
calculateFitness() {
// Implement this instead of #getFitness. The first #getFitness call
// will invoke this method and return its result. All subsequent
// #getFitness calls will return the same result without invoking
// #calculateFitness again.
}
crossover(other) {
// Assume doCrossover produces an array of child instances.
let children = doCrossover(this, other);
// Return only the highest-fitness child. The fitness result will be
// reused later when gene-lib needs it.
return _.maxBy(children, (child) => child.getFitness());
}
}
module.exports = MyCachingChromosome;
gene-lib
comes with tournament and roulette selection methods. If you need
others, create one by extending the Selector
class:
const { Selector } = require('gene-lib');
class MySelector extends Selector {
constructor(settings) {
// Settings object will be provided by the settings.selectorSettings
// ::run argument.
super(settings);
// Do any additional initialization here.
}
static validateSettings(settings) {
// This method is optional, but can be implemented to ensure that the
// settings provided to instances make sense. It should throw an
// appropriate error if they don't.
}
add(individual) {
// Store individual in the selector. The individual will be an object
// with two properties: 'fitness' and 'chromosome'.
}
select() {
// Return a single stored individual.
}
}
module.exports = MySelector;
Then provide the selector to the run method:
const geneLib = require('geneLib');
const MyChromosome = require('./path/to/my-chromosome');
const MySelector = require('./path/to/my-selector');
geneLib.run({
chromosomeClass: MyChromosome,
selectorClass: MySelector,
generationSize: 100,
generationLimit: 1000,
crossoverRate: 0.7,
mutationRate: 0.01
})
.then((result) => {
// ...
});
Since most selectors will store individuals in an array, the ArraySelector
base class is provided as a convenience. Its constructor creates an empty
array on this.individuals
, and its #add
method simply pushes the provided
individual onto that array:
const { ArraySelector } = require('gene-lib');
class MyArraySelector extends ArraySelector {
select() {
// Return a single individual from this.individuals.
}
}
module.exports = MyArraySelector;
If you need your custom selection methods to be used more easily, you can
register them globally with gene-lib
:
const geneLib = require('geneLib');
const MySelector = require('./path/to/my-selector');
geneLib.registerSelector('my-selector', MySelector);
const geneLib = require('geneLib');
const MyChromosome = require('./path/to/my-chromosome');
geneLib.run({
chromosomeClass: MyChromosome,
selector: 'my-selector',
generationSize: 100,
generationLimit: 1000,
crossoverRate: 0.7,
mutationRate: 0.01
})
.then((result) => {
// ...
});
In Node frameworks it is common to support potentially asynchronous operations
by simply wrapping the user-provided result with
Promise::resolve
.
This allows the user to return a promise for asynchronous operation, or anything
else for synchronous operation.
For genetic algorithms, however, this approach has too great an impact on
performance, as it requires thousands of promises to be created every second.
Instead, gene-lib
assumes that every operation will be synchronous by default,
since this will be the case for most genetic algorithms. If any operation in
your GA needs to be asynchronous, you must declare it to be so explicitly using
the settings.async
::run
argument.
For example, if your #getFitness
method must make a request to a database,
have it return a promise that resolves with the fitness. Then set the
getFitness
property on settings.async
:
const geneLib = require('gene-lib');
const MyChromosome = require('./path/to/my-chromosome');
geneLib.run({
chromosomeClass: MyChromosome,
generationSize: 100,
generationLimit: 1000,
crossoverRate: 0.7,
mutationRate: 0.01,
async: {
getFitness: true
}
})
.then((result) => {
// ...
});
You can make any combination of operations asynchronous if you wish. Even selector operations, though this will of course require a custom selector:
const geneLib = require('gene-lib');
const MyChromosome = require('./path/to/my-chromosome');
const MySelector = require('./path/to/my-selector');
geneLib.run({
chromosomeClass: MyChromosome,
selectorClass: MySelector,
generationSize: 100,
generationLimit: 1000,
crossoverRate: 0.7,
mutationRate: 0.01,
async: {
// Chromosome operations.
create: true,
getFitness: true,
crossover: true,
mutate: true,
// Selector operations.
add: true,
select: true
}
})
.then((result) => {
// ...
});
By default, asynchronous operations are done in series, but concurrency is
supported. Simply provide the maximum number of simultaneous operations as the
value in the async
object:
const geneLib = require('gene-lib');
const MyChromosome = require('./path/to/my-chromosome');
geneLib.run({
chromosomeClass: MyChromosome,
generationSize: 100,
generationLimit: 1000,
crossoverRate: 0.7,
mutationRate: 0.01,
async: {
create: 4,
getFitness: 2,
}
})
.then((result) => {
// ...
});
If concurrent operations involve a request to some external service-- as they often will-- make sure not to set these concurrencies too high, otherwise you might overload that service with simultaneous requests.
If you're 100% sure that you won't need to unblock execution for the entire
run, you can force it to take place in a single event loop iteration using the
::runSync
method:
const geneLib = require('gene-lib');
const MyChromosome = require('./path/to/my-chromosome');
let result = geneLib.runSync({
chromosomeClass: MyChromosome,
generationSize: 100,
generationLimit: 1000,
createArg: 'create argument',
crossoverRate: 0.2,
mutationRate: 0.05
});
// ...
As you can see, the result is returned directly instead of being wrapped in a
promise. Of course, when using this method, none of the async
settings above
are supported. This method will throw if you attempt to use them.
This method will be slightly faster than ::run
and maybe a little more
convenient in some cases, but save yourself some trouble and avoid using it in
web services or other multi-user applications. You should only use it if
you're writing a command-line script or some other application that does not
need to share its thread.
In addition to everything described thus far, gene-lib
also comes with some
utility methods to help you with your chromosome methods. For example, to
perform two-point crossovers, try something like this:
const lodash = require('lodash');
const geneLib = require('geneLib');
const { Chromosome } = geneLib;
class MyChromosome extends Chromosome {
// ...
crossover(other) {
// Assuming 'genes' is an array property on each chromosome, containing
// its genetic data.
let children = geneLib.twoPointCrossover(this.genes, other.genes);
// Wrap children in new MyChromosome instances.
return children.map((child) => new MyChromosome(child));
}
}
module.exports = MyChromosome;
Utilities include:
::getRandomIndex
::getRandomRange
::getRandomIndices
::singlePointCrossover
::twoPointCrossover
::uniformCrossover
::pmx
For a simple example of gene-lib
in action, check out the
phrase solver
in the
integration tests.
Currently on npm there are a lot of competing projects in this space without a
clear winner. I created gene-lib
mainly for my own enjoyment and because all
of the existing projects had something about them I that didn't quite like. For
most, a major problem is blocking execution for the whole run unless you have a
true async operation somewhere. This can sometimes be remedied with a
setImmediate
in the right place, but in my opinion it's better for the
framework to handle it.
- darwin-js - Similar idea, but far simpler. As of this writing, it only supports async fitness calculation, and even that must be in series.
- genetic - I think this is a bit more targeted towards browsers. Everything is callback-based, which avoids the issue of creating lots of unnecessary promises, but that can make it a bit less nice to use.
- genetic-js - One of the better alternatives, though it could use some more documentation, and it's API is not my personal cup of tea.
There are a lot more that I couldn't hope to list here. Feel free to search
around on npm, if you're interested. I won't tell you which one to use, though
I hope I've made a pretty good case for gene-lib
.