SwiftGenetics is a genetic algorithm library that has been engineered from-scratch to be highly extensible and composable, by abstracting away different pieces of functionality, while providing concrete implementations of certain use cases, such as tree-based genomes. SwiftGenetics is written in pure Swift, and makes proper use of functional programming and Swift's wonderful type system. This project is provided under the MIT License (see the LICENSE
file for more info).
SwiftGenetics can be added to your project as a dependency with the Swift Package Manager.
SwiftGenetics provides an abstracted base for genetic algorithms, intended to support different representations and genetic operators.
- Gene Represented by the
Gene
protocol, genes represent an individual piece of a genome. - Genome A collection of genes, represented by the
Genome
protocol. - Organism An individual in the population, represented by the
Organism
class (we want reference semantics here), and generic upon aGenome
subtype, where the organism's genotype is an instance of the type that conforms toGenome
. - Population The world in which organisms evolve, represented by the
Population
class. - Environment A configuration which is disseminated down to individual genes and controls various behaviors such as mutation, recombination, and selection. You can modify environments as a GA is running.
- Clade A "kind" of genetic algorithm (e.g., bit strings and trees might be two kinds clades). Clades can be built from broader clades to get more complex functionality. Clades conform to the
Genome
protocol. Clades are intended to be generic and composable. A specific GA can be thought of as being a "concretized" clade type (e.g., a tree of arithmetic operators).
- Living Strings Homogeneous sequences of genes. A classic within the world of genetic algorithms.
- Living Trees Trees whose structure and node types are evolved.
- Living Forests A fixed-size collection of coevolved trees. Evolution of Living Forests is analogous to evolution of a genome of chromosomes.
The library also provides the following primitives that you can use to build your own complex genomes:
ContinuousGene
to represent a single floating-point value.DiscreteChoiceGene
to represent a discrete choice in a set.
Everything you need to use SwiftGenetics is in the Sources/
directory. The main entry point you need is Population
. Generally, you can use SwiftGenetics as follows:
- Build a
Population
. - Call the
epoch()
method on your population. - Set the fitness for each
Organism
in the population'sorganisms
array. - Repeat steps 2 and 3 ad infinitum (or until you're happy with a solution).
Oftentimes you might have long-running fitness calculations that you want to run concurrently, in which case you can use EvolutionWrapper
types as the entry point into SwiftGenetics. ConcurrentSynchronousEvaluationGA
and ConcurrentAsynchronousEvaluationGA
perform synchronous and asynchronous, respectively, fitness evaluations. These wrappers make adding a GA trivial, all you need are an initial population and a type that conforms to EvolutionLoggingDelegate
.
One cool feature that ConcurrentAsynchronousEvaluationGA
has is the ability to detect and efficiently handle duplicate genomes within a generation (based on their hash value from a conformance to Hashable
).
An end-to-end example of evolution wrapper use can be found in Tests/SwiftGeneticsTests/Integration/GeneticAlgorithmIntegrationTests.swift
. Specifically, the testSortingGA
integration test shows how SwiftGenetics can be used to evolve vectors of continuous values.
Every state object in SwiftGenetics conforms to the Swift Codable
protocol, allowing serialization and deserialization. The CheckpointManager
provides a lovely abstraction for this that allows writing and reading checkpoints in one line:
typealias G = ... // A concrete Genome type.
let population: Population<G> = ...
let checkpointFile: URL = ...
// Save the population to a checkpoint
try CheckpointManager<G>.save(population, to: checkpointFile)
// Build a new population from the checkpoint.
let reconstitutedPopulation = try CheckpointManager<G>.population(from: checkpointFile)
// Example coming soon!
Genetic algorithms, as a black box optimizer, aim find the global maximum, so your fitness function must be formulated so that larger values are better. Types can conform to the SynchronousFitnessEvaluator
or AsynchronousFitnessEvaluator
protocols as a way to provide fitness values.
struct ExampleFitnessEvaluator: SynchronousFitnessEvaluator {
typealias G = MyGenomeType
func fitnessFor(organism: Organism<G>, solutionCallback: (G, Double) -> ()) -> Double {
// Example coming soon!
}
}
There are no external dependencies, yeehaw! On macOS, you can get off the ground running. Linux is supported out-of-the-box too!
If you use SwiftGenetics in a publication, you can use the following BibTeX entry to cite this project:
@misc{swiftgenetics,
Title = {SwiftGenetics},
Author = {Santiago Gonzalez},
howpublished = {\url{https://github.com/sgonzalez/SwiftGenetics}}
}
- More flexibility.
- More extensibility.
- More concrete implementations.
- Support for dominance relations, like in NSGA-II.
- Generalize
coefficient
inLivingTreeGene
. - More tests (unit, integration, performance).
- Test app that visualizes the evolution process.