#neural-network #neat #genome #traits #input #activation #oxineat

oxineat-nn

A neural-network implementation of the OxiNEAT crate's Genome trait

2 releases

0.1.1 Nov 1, 2021
0.1.0 Oct 26, 2021

#487 in Machine learning

37 downloads per month
Used in oxineat

MIT license

195KB
2.5K SLoC

OxiNEAT-NN

An neural network-based implementation of the OxiNEAT crate's Genome trait.

Provides a GenomeNN type usable in OxiNEAT Populations, as well as two neural network implementations which can be generated from a GenomeNN:

  • RealTimeNetwork: best suited for real-time control tasks, with new inputs set for each activation, and multiple time-steps involved.
  • FunctionApproximatorNetwork: best suited for more instantaneous single-output-per-input function approximation tasks.

Example usage: evolution of XOR function approximator

use oxineat::{Population, PopulationConfig};
use oxineat_nn::{
    genomics::{ActivationType, GeneticConfig, NNGenome},
    networks::FunctionApproximatorNetwork,
};
use serde_json;
use std::num::NonZeroUsize;

// Allowed error margin for neural net answers.
const ERROR_MARGIN: f32 = 0.3;

fn evaluate_xor(genome: &NNGenome) -> f32 {
    let mut network = FunctionApproximatorNetwork::from::<1>(genome);

    let values = [
        ([1.0, 0.0, 0.0], 0.0),
        ([1.0, 0.0, 1.0], 1.0),
        ([1.0, 1.0, 0.0], 1.0),
        ([1.0, 1.0, 1.0], 0.0),
    ];

    let mut errors = [0.0, 0.0, 0.0, 0.0];
    for (i, (input, output)) in values.iter().enumerate() {
        errors[i] = (network.evaluate_at(input)[0] - output).abs();
        if errors[i] < ERROR_MARGIN {
            errors[i] = 0.0;
        }
    }

    (4.0 - errors.iter().copied().sum::<f32>()).powf(2.0)
}

fn main() {
    let genetic_config = GeneticConfig {
        input_count: NonZeroUsize::new(3).unwrap(),
        output_count: NonZeroUsize::new(1).unwrap(),
        activation_types: vec![ActivationType::Sigmoid],
        output_activation_types: vec![ActivationType::Sigmoid],
        child_mutation_chance: 0.65,
        mate_by_averaging_chance: 0.4,
        suppression_reset_chance: 1.0,
        initial_expression_chance: 1.0,
        weight_bound: 5.0,
        weight_reset_chance: 0.2,
        weight_nudge_chance: 0.9,
        weight_mutation_power: 2.5,
        node_addition_mutation_chance: 0.03,
        gene_addition_mutation_chance: 0.05,
        max_gene_addition_mutation_attempts: 20,
        recursion_chance: 0.0,
        excess_gene_factor: 1.0,
        disjoint_gene_factor: 1.0,
        common_weight_factor: 0.4,
        ..GeneticConfig::zero()
    };

    let population_config = PopulationConfig {
        size: NonZeroUsize::new(150).unwrap(),
        distance_threshold: 3.0,
        elitism: 1,
        survival_threshold: 0.2,
        sexual_reproduction_chance: 0.6,
        adoption_rate: 1.0,
        interspecies_mating_chance: 0.001,
        stagnation_threshold: NonZeroUsize::new(15).unwrap(),
        stagnation_penalty: 1.0,
    };

    let mut population = Population::new(population_config, genetic_config);
    for _ in 0..100 {
        population.evaluate_fitness(evaluate_xor);
        if (population.champion().fitness() - 16.0).abs() < f32::EPSILON {
            println!("Solution found!: {}", serde_json::to_string(&population.champion()).unwrap());
            break;
        }
        if let Err(e) = population.evolve() {
            eprintln!("{}", e);
            break;
        }
    }
}

License

Licensed under the MIT license.

Dependencies

~0.8–1.6MB
~34K SLoC