#genetic-algorithm #evolutionary #neuroevolution

revonet

Rust implementation of real-coded genetic algorithm for solving optimization problems and training of neural networks. The latter is also known as neuroevolution.

3 unstable releases

Uses old Rust 2015

0.2.1 Aug 20, 2017
0.2.0 Aug 20, 2017
0.1.0 Jul 14, 2017

#1777 in Algorithms

40 downloads per month

GPL-3.0 license

105KB
1.5K SLoC

revonet

Rust implementation of real-coded genetic algorithm for solving optimization problems and training of neural networks. The latter is also known as neuroevolution.

Features:

  • real-coded Evolutionary Algorithm
  • NeuroEvolutionary tuning of weights of Neural Network with fixed structure
  • supports several feed-forward architectures

https://github.com/yurytsoy/revonet/blob/master/imgs/nn_arch.png

  • automatically computes statistics for single and multiple runs for EA and NE
  • EA settings and results can be saved to json
  • allows defining user-specified objective functions for EA and NE (see examples below)

Examples

Real-coded genetic algorithm

let pop_size = 20u32;       // population size.
let problem_dim = 10u32;    // number of optimization parameters.

let problem = RosenbrockProblem{};  // objective function.
let gen_count = 10u32;      // generations number.
let settings = GASettings::new(pop_size, gen_count, problem_dim);
let mut ga: GA<RosenbrockProblem> = GA::new(settings, &problem);   // init GA.
let res = ga.run(settings).expect("Error during GA run");  // run and fetch the results.

// get and print results of the current run.
println!("\n\nGA results: {:?}", res);

// make multiple runs and get combined results.
let res = ga.run_multiple(settings, 10 as u32).expect("Error during multiple GA runs");
println!("\n\nResults of multple GA runs: {:?}", res);

Run evolution of NN weights to solve regression problem

let (pop_size, gen_count, param_count) = (20, 20, 100); // gene_count does not matter here as NN structure is defined by a problem.
let settings = EASettings::new(pop_size, gen_count, param_count);
let problem = SymbolicRegressionProblem::new_f();

let mut ne: NE<SymbolicRegressionProblem> = NE::new(&problem);
let res = ne.run(settings).expect("Error: NE result is empty");
println!("result: {:?}", res);
println!("\nbest individual: {:?}", res.best);

Creating multilayered neural network with 2 hidden layers with sigmoid activation and with linear output nodes.

const INPUT_SIZE: usize = 20;
const OUTPUT_SIZE: usize = 2;

let mut rng = rand::thread_rng();   // needed for weights initialization when NN is built.
let mut net: MultilayeredNetwork = MultilayeredNetwork::new(INPUT_SIZE, OUTPUT_SIZE);
net.add_hidden_layer(30 as usize, ActivationFunctionType::Sigmoid)
     .add_hidden_layer(20 as usize, ActivationFunctionType::Sigmoid)
     .build(&mut rng, NeuralArchitecture::Multilayered);       // `build` finishes creation of neural network.

let (ws, bs) = net.get_weights();   // `ws` and `bs` are `Vec` arrays containing weights and biases for each layer.
assert!(ws.len() == 3);		// number of elements equals to number of hidden layers + 1 output layer
assert!(bs.len() == 3);		// number of elements equals to number of hidden layers + 1 output layer

Creating custom optimization problem for GA

// Dummy problem returning random fitness.
pub struct DummyProblem;
impl Problem for DummyProblem {
    // Function to evaluate a specific individual.
    fn compute<T: Individual>(&self, ind: &mut T) -> f32 {
        // use `to_vec` to get real-coded representation of an individual.
        let v = ind.to_vec().unwrap();

        let mut rng: StdRng = StdRng::from_seed(&[0]);
        rng.gen::<f32>()
    }
}

Creating custom problem for NN evolution

// Dummy problem returning random fitness.
struct RandomNEProblem {}
impl RandomNEProblem {
    fn new() -> RandomNEProblem {
        RandomNEProblem{}
    }
}
impl NeuroProblem for RandomNEProblem {
    // return number of NN inputs.
    fn get_inputs_num(&self) -> usize {1}
    // return number of NN outputs.
    fn get_outputs_num(&self) -> usize {1}
    // return NN with random weights and a fixed structure. For now the structure should be the same all the time to make sure that crossover is possible. Likely to change in the future.
    fn get_default_net(&self) -> MultilayeredNetwork {
        let mut rng = rand::thread_rng();
        let mut net: MultilayeredNetwork = MultilayeredNetwork::new(self.get_inputs_num(), self.get_outputs_num());
        net.add_hidden_layer(5 as usize, ActivationFunctionType::Sigmoid)
            .build(&mut rng, NeuralArchitecture::Multilayered);
        net
    }
    // Function to evaluate performance of a given NN.
    fn compute_with_net<T: NeuralNetwork>(&self, nn: &mut T) -> f32 {
        let mut rng: StdRng = StdRng::from_seed(&[0]);
        let mut input = (0..self.get_inputs_num())
                            .map(|_| rng.gen::<f32>())
                            .collect::<Vec<f32>>();
        // compute NN output using random input.
        let mut output = nn.compute(&input);
        output[0]
    }
}


Dependencies

~1–2MB
~41K SLoC