Skip to content

garro95/oxigen

 
 

Repository files navigation

oxigen (Fork)

Build Status Current Crates.io Version

This is a fork of the oxigen crate. The link is to the original crate. This fork is released under a different but compatible license.

Oxigen is a parallel genetic algorithm library implemented in Rust. The name comes from the merge of OXIdación (Rust translated to Spanish) and GENetic.

Oxigen provides the following features:

  • Fast and parallel genetic algorithm implementation (it solves the N Queens problem for N=255 in few seconds). For benchmarks view benchmarks section of this file.
  • Customizable mutation and selection rates with constant, linear and cuadratic functions according to generations built-in (you can implement your own functions via the MutationRate and SelectionRate traits).
  • Customizable age unfitness of individuals, with no unfitness, linear and cuadratic unfitness with threshold according to generations of the individual built-in (you can implement your own age functions via the Age trait).
  • Accumulated Roulette, Tournaments and Cup built-in selection functions (you can implement your own selection functions via the Selection trait).
  • SingleCrossPoint built-in crossover function (you can implement your own crossover function via the Crossover trait).
  • Worst built-in survival pressure function (the worst individuals are killed until reaching the original population size). You can implement your own survival pressure functions via the SurvivalPressure trait.
  • SolutionFound, Generation and Progress built-in stop criteria (you can implement your own stop criteria via the StopCriterion trait).
  • Genotype trait to define the genotype of your genetic algorithm. Whatever struct can implement the Genotype trait under the following restrictions:
    • It has a iter function that returns a use std::slice::Iter over its genes.
    • It has a into_iter function that consumes the individual and returns a use std::vec::IntoIter over its genes.
    • It implements FromIterator over its genes type, Display, Clone, Send and Sync.
    • It has functions to generate a random individual, to mutate an individual, to get the fitness of an individual and to know if and individual is_solution of the problem.
  • Individual's fitness is cached to not do unnecessary recomputations (this can be disabled with .cache_fitness(false) if your fitness function is stochastic and so you need to recompute fitness in each generation).
  • Progress statistics can be configured to be printed every certain number of generations to a file.
  • Population individuals with their fitnesses can be configured to be printed every certain number of generations to a file.
  • Specific initial individuals can be inserted in the genetic algorithm execution.
  • Genetic executions can be resumed using the population of the last generation as initial population.
  • Coevolution is possible executing little genetic algorithm re-executions inside the fitness function.

Usage

In your Cargo.toml file add the oxigen dependency:

[dependencies]
oxigen = "^1.4.1"

To use oxigen use oxigen::prelude::* and call the run method over a GeneticExecution instance overwriting the default hyperparameters and functions folllowing your needs:

let n_queens: u8 = std::env::args()
    .nth(1)
    .expect("Enter a number between 4 and 255 as argument")
    .parse()
    .expect("Enter a number between 4 and 255 as argument");

let progress_log = File::create("progress.csv").expect("Error creating progress log file");
let population_log = File::create("population.txt").expect("Error creating population log file");
let log2 = (f64::from(n_queens) * 4_f64).log2().ceil();

let population_size = 2_i32.pow(log2 as u32) as usize;

let (solutions, generation, progress) = GeneticExecution::<u8, QueensBoard>::new()
    .population_size(population_size)
    .genotype_size(n_queens as u8)
    .mutation_rate(Box::new(MutationRates::Linear(SlopeParams {
        start: f64::from(n_queens) / (8_f64 + 2_f64 * log2) / 100_f64,
        bound: 0.005,
        coefficient: -0.0002,
    })))
    .selection_rate(Box::new(SelectionRates::Linear(SlopeParams {
        start: log2 - 2_f64,
        bound: log2 / 1.5,
        coefficient: -0.0005,
    })))
    .select_function(Box::new(SelectionFunctions::Cup))
    .age_function(Box::new(AgeFunctions::Cuadratic(
        AgeThreshold(50),
        AgeSlope(1_f64),
    )))
    .progress_log(20, progress_log)
    .population_log(2000, population_log)
    .run();

For a full example visit the nqueens-oxigen example.

For more information visit the documentation.

Resuming a previous execution

Since version 1.1.0, genetic algorithm executions return the population of the last generation and new genetic executions accept a initial population. This permits to resuming previous executions and it also enables coevolution, since little genetic algorithm re-executions can be launched in the fitness function.

In the following example a execution with 10000 generations is launched and after it is resumed until finding a solution with different rates.

let n_queens: u8 = std::env::args()
    .nth(1)
    .expect("Enter a number between 4 and 255 as argument")
    .parse()
    .expect("Enter a number between 4 and 255 as argument");

let progress_log = File::create("progress.csv").expect("Error creating progress log file");
let population_log = File::create("population.txt").expect("Error creating population log file");
let log2 = (f64::from(n_queens) * 4_f64).log2().ceil();

let population_size = 2_i32.pow(log2 as u32) as usize;

let (_solutions, _generation, _progress, population) = GeneticExecution::<u8, QueensBoard>::new()
    .population_size(population_size)
    .genotype_size(n_queens as u8)
    .mutation_rate(Box::new(MutationRates::Linear(SlopeParams {
        start: f64::from(n_queens) / (8_f64 + 2_f64 * log2) / 100_f64,
        bound: 0.005,
        coefficient: -0.0002,
    })))
    .selection_rate(Box::new(SelectionRates::Linear(SlopeParams {
        start: log2 - 2_f64,
        bound: log2 / 1.5,
        coefficient: -0.0005,
    })))
    .select_function(Box::new(SelectionFunctions::Cup))
    .age_function(Box::new(AgeFunctions::Cuadratic(
        AgeThreshold(50),
        AgeSlope(1_f64),
    )))
    .stop_criterion(Box::new(StopCriteria::Generation(10000)))
    .run();

let (solutions, generation, progress, _population) = GeneticExecution::<u8, QueensBoard>::new()
    .population_size(population_size)
    .genotype_size(n_queens as u8)
    .mutation_rate(Box::new(MutationRates::Linear(SlopeParams {
        start: f64::from(n_queens) / (8_f64 + 4_f64 * log2) / 100_f64,
        bound: 0.005,
        coefficient: -0.0002,
    })))
    .selection_rate(Box::new(SelectionRates::Linear(SlopeParams {
        start: log2 - 4_f64,
        bound: log2 / 1.5,
        coefficient: -0.0005,
    })))
    .select_function(Box::new(SelectionFunctions::Cup))
    .age_function(Box::new(AgeFunctions::Cuadratic(
        AgeThreshold(50),
        AgeSlope(1_f64),
    )))
    .population(population)
    .progress_log(20, progress_log)
    .population_log(2000, population_log)
    .run();

Building

To build oxigen, use cargo like for any Rust project:

  • cargo build to build in debug mode.
  • cargo build --release to build with optimizations.

To run benchmarks, you will need a nightly Rust compiler. Uncomment the lines // #![feature(test)] and // mod benchmarks; from lib.rs and then bechmarks can be run using cargo bench.

Benchmarks

The following benchmarks have been created to measure the genetic algorithm functions performance:

running 14 tests
test benchmarks::bench_cross                  ... bench:      31,740 ns/iter (+/- 2,106)
test benchmarks::bench_cup                    ... bench:     319,214 ns/iter (+/- 31,373)
test benchmarks::bench_fitness                ... bench:     542,184 ns/iter (+/- 31,025)
test benchmarks::bench_fitness_age            ... bench:      44,545 ns/iter (+/- 2,616)
test benchmarks::bench_get_fitnesses          ... bench:      16,936 ns/iter (+/- 703)
test benchmarks::bench_get_solutions          ... bench:      34,221 ns/iter (+/- 1,332)
test benchmarks::bench_mutation               ... bench:           7 ns/iter (+/- 0)
test benchmarks::bench_not_cached_fitness     ... bench:     531,058 ns/iter (+/- 2,958)
test benchmarks::bench_not_cached_fitness_age ... bench:     531,087 ns/iter (+/- 26,893)
test benchmarks::bench_roulette               ... bench:     276,643 ns/iter (+/- 1,858)
test benchmarks::bench_sort_population        ... bench:       1,483 ns/iter (+/- 6)
test benchmarks::bench_survival_pressure      ... bench:          20 ns/iter (+/- 0)
test benchmarks::bench_tournaments            ... bench:     964,101 ns/iter (+/- 45,469)
test benchmarks::bench_update_progress        ... bench:       5,678 ns/iter (+/- 172)

test result: ok. 0 passed; 0 failed; 0 ignored; 14 measured; 0 filtered out

The difference of performance among the different fitness benchmarks have the following explanations:

  • bench_fitness measures the performance of a cached execution cleaning the fitnesses after each bench iteration. This cleaning is the reason of being a bit slower than not cached benchmarks.
  • bench_fitness_age measures the performance with fitness cached in all bench iterations, so it is very much faster.
  • Not cached benchmarks measure the performance of not cached executions, with 1 generation individuals in the last case, so the performance is similar but a bit slower for the benchmark that must apply age unfitness.

The bench_tournaments is slower than cup and roulette because it has been configured with population_size / 2 tournaments of the same size of individuals. In practice, a configuration like this is only used in the last generations of the algorithm where the selection rate is high, being very much faster in the previous generations.

Contributing

Contributions are absolutely, positively welcome and encouraged! Contributions come in many forms. You could:

  1. Submit a feature request or bug report as an issue.
  2. Ask for improved documentation as an issue.
  3. Comment on issues that require feedback.
  4. Contribute code via pull requests, don't forget to run cargo fmt before submitting your PR!

We aim to keep Rocket's code quality at the highest level. This means that any code you contribute must be:

  • Commented: Public items must be commented.
  • Documented: Exposed items must have rustdoc comments with examples, if applicable.
  • Styled: Your code should be rustfmt'd when possible.
  • Simple: Your code should accomplish its task as simply and idiomatically as possible.
  • Tested: You should add (and pass) convincing tests for any functionality you add when it is possible.
  • Focused: Your code should do what it's supposed to do and nothing more.

Note that unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this fork of oxigen by you shall be licensed under a license compatible with the GNU GPL version 3 or later.

License

This fork of oxigen is released under the GNU GPL version 3 or, at your opinion, any later version.

About

Fast, paralell, extensible and adaptable genetic algorithm library written in Rust

Resources

License

Code of conduct

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Rust 100.0%