1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152

use std::sync::Arc;

use cge::Network;

use cmaes_utils::*;
use threads::update_generation;
use generation::initialize_generation;
use mutation::mutate;
use options::EANT2Options;
use select::select;
use fitness::NNFitnessFunction;

/// Returns a neural network with as small a fitness as possible, along with the network's fitness.
/// See the library level documentation for examples.
pub fn eant_loop<T>(object: &T, options: EANT2Options) -> (Network, f64)
	where T: 'static + NNFitnessFunction + Clone + Send + Sync
{
    // TODO: Allow user to pass a neural network as an argument to generate the initial population
    
    // Allow user to pass a neural network as an argument to generate the initial population

    // Extract end conditions from the options, using the CMAESOptions methods to take advantage of
    // the checks they make

    let cmaes_options = get_cmaes_options(options.cmaes_end_conditions).end_conditions;

    // Extract other options
    let threads = options.threads as usize;
    let fitness_threshold = options.fitness_threshold;
    let threshold = options.similar_fitness;
    let max_generations = options.max_generations;
    let inputs = options.inputs;
    let outputs = options.outputs;
    let cmaes_runs = options.cmaes_runs;
    let population_size = options.population_size;
    let offspring_count = options.offspring_count;
    let print = options.print_option;
    let weights = options.weights;
    let transfer_function = options.transfer_function;

    // Wrap object in an Arc for sharing across threads
    // The object is a user-defined type that implements the FitnessFunction trait
    // It can use the &self argument to factor addition information into the fitness calculation
    let object = Arc::new(object.clone());

    // Loop counter
    let mut g = 0;

    // Initialize generation
    // One neuron per output, each connected to approximately 50% of inputs (exact number is
    // random)

    let mut generation = initialize_generation(population_size,
                                               offspring_count,
                                               inputs,
                                               outputs,
											   transfer_function,
                                               object);

    // Everything from here on is part of a loop
    // EANT2 repeatedly runs these operations until the end conditions are met (minimum fitness or
    // maximum generations)

    loop {
        // Get fitness of network topologies by optimizing the parameters of each individual with
        // CMA-ES to get their maximum potential
        // This stage makes up for nearly all the running time of the algorithm, sometimes taking
        // hours or days
        // The good news is that EANT2 will only need 20 generations or so at most to find a solution,
        // even on complicated tasks

        if print {
            println!("Beginning EANT2 generation {}", g + 1);
        }

        update_generation(&mut generation,
                          &cmaes_options,
                          cmaes_runs,
                          threads);

        // Select individuals to go on to the next generation
        // This is necessary because there will be too many individuals in the population after
        // generating new ones

        generation = select(population_size, generation, threshold);
        generation.sort_by(|a, b| a.fitness.partial_cmp(&b.fitness).expect("12"));

        // Test end conditions, and if they are met, break the loop

        let best = generation[0].fitness;

        if best <= fitness_threshold || g + 1 >= max_generations {
            let solution = generation[0].clone();

            if print {
                println!("EANT2 terminated in {} generations", g + 1);
                println!("Solution found with size {} and {} fitness", solution.network.size + 1, solution.fitness);
            }

            return (solution.network, best);
        }

        // Add new individuals to the population by mutating the existing ones

        let mut new_individuals = Vec::new();

        for individual in &generation {
            for _ in 0..offspring_count {
                let mut new = individual.clone();

                mutate(&mut new, &weights);

                new_individuals.push(new);
            }
        }

        //use this for testing
        //individuals should stay at a small size as long as population size and offspring count
        //are reasonable
        //for (a, i) in generation.iter().enumerate() {
            //println!("size of individual {}: {}", a, i.network.size + 1);
        //}

        //for i in &generation {
            //let compact = i.network.genome.iter().map(|i| &i.variant).collect::<Vec<_>>();
            //println!("{:?}", compact);
        //}
        //println!("");

        // if finished, polish the code and docs up, then multithread it
        // finish todo in cmaes and optimization todo in cge
        // then publish

        generation.extend_from_slice(&new_individuals);

        // Increment gene ages

        for i in &mut generation {
            for age in &mut i.ages {
                *age += 1;
            }
        }

        // Increment loop counter

        g += 1;

        if print {
            println!("Current best fitness: {}", generation[0].fitness);
        }
    }
}