memega 0.1.0

genetic algorithm library
Documentation 
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
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190

use derive_more::Display;
use eyre::{eyre, Result};

use crate::cfg::{Cfg, Crossover, Duplicates, Mutation, Replacement, Selection, Survival};
use crate::eval::{Evaluator, Genome, Mem};
use crate::gen::species::SpeciesId;
use crate::gen::unevaluated::UnevaluatedGen;
use crate::ops::mutation::{mutate_lognorm, mutate_normal, mutate_rate};
use crate::ops::sampling::{multi_rws, rws, sus};
use crate::run::runner::RandGenome;
#[derive(Display, Clone, PartialOrd, PartialEq)]
#[display(fmt = "pop: {}, best: {}", "mems.len()", "self.mems[0]")]
pub struct EvaluatedGen<G: Genome> {
    pub mems: Vec<Mem<G>>,
}

impl<G: Genome> EvaluatedGen<G> {
    #[must_use]
    pub fn new(mut mems: Vec<Mem<G>>) -> Self {
        // Sort by base fitness. Selection should happen using selection
        // fitness. Generate survivors using base fitness, to make sure we keep
        // the top individuals.
        mems.sort_unstable_by(|a, b| b.base_fitness.partial_cmp(&a.base_fitness).unwrap());
        Self { mems }
    }

    #[must_use]
    pub fn mems(&self) -> &[Mem<G>] {
        &self.mems
    }

    #[must_use]
    pub fn species_mems(&self, n: SpeciesId) -> Vec<Mem<G>> {
        self.mems.iter().filter(|v| v.species == n).cloned().collect()
    }

    // Get list of species.
    #[must_use]
    pub fn species(&self) -> Vec<SpeciesId> {
        // Relies on species index assignment to be contigous from zero.
        let mut species: Vec<_> = self.mems.iter().map(|mem| mem.species).collect();
        species.sort_unstable();
        species.dedup();
        species
    }

    fn survivors(&self, survival: Survival, cfg: &Cfg) -> Vec<Mem<G>> {
        match survival {
            Survival::TopProportion(prop) => {
                // Ceiling so we don't miss keeping things for small sizes.
                // Use the target population size rather than the size of the
                // current generation so small generations don't have a smaller
                // number of survivors selected from them. This is useful for
                // with a small number of individuals.
                let num = (cfg.pop_size as f64 * prop).ceil() as usize;
                self.mems.iter().take(num).cloned().collect()
            }
            Survival::SpeciesTopProportion(prop) => {
                let mut survivors = Vec::new();
                let species = self.species();
                let num = (cfg.pop_size as f64 * prop / species.len() as f64).ceil() as usize;
                for id in species {
                    survivors.extend(self.species_mems(id).into_iter().take(num));
                }
                survivors
            }
        }
    }

    fn selection(&self, selection: Selection) -> [Mem<G>; 2] {
        let fitnesses = self.mems.iter().map(|v| v.selection_fitness).collect::<Vec<_>>();
        let idxs = match selection {
            Selection::Sus => sus(&fitnesses, 2),
            Selection::Roulette => multi_rws(&fitnesses, 2),
        };
        [self.mems[idxs[0]].clone(), self.mems[idxs[1]].clone()]
    }

    fn check_weights(weights: &[f64], l: usize) -> Result<()> {
        if weights.len() != l {
            return Err(eyre!("number of fixed weights {} doesn't match {}", weights.len(), l));
        }
        for &v in weights.iter() {
            if v < 0.0 {
                return Err(eyre!("weights must all be non-negative: {}", v));
            }
        }
        Ok(())
    }

    fn crossover<E: Evaluator<Genome = G>>(
        &self,
        crossover: &Crossover,
        eval: &E,
        s1: &mut Mem<G>,
        s2: &mut Mem<G>,
    ) -> Result<()> {
        match crossover {
            Crossover::Fixed(rates) => {
                s1.params.crossover = rates.clone();
                s2.params.crossover = rates.clone();
            }
            Crossover::Adaptive => {
                let lrate = 1.0 / (self.mems.len() as f64).sqrt();
                mutate_rate(&mut s1.params.crossover, 1.0, |v| mutate_normal(v, lrate).max(0.0));
                mutate_rate(&mut s2.params.crossover, 1.0, |v| mutate_normal(v, lrate).max(0.0));
            }
        };
        Self::check_weights(&s1.params.crossover, E::NUM_CROSSOVER)?;
        Self::check_weights(&s2.params.crossover, E::NUM_CROSSOVER)?;
        let idx = rws(&s1.params.crossover).unwrap();
        eval.crossover(&mut s1.genome, &mut s2.genome, idx);
        Ok(())
    }

    fn mutation<E: Evaluator<Genome = G>>(
        &self,
        mutation: &Mutation,
        eval: &E,
        s: &mut Mem<G>,
    ) -> Result<()> {
        match mutation {
            Mutation::Fixed(rates) => {
                s.params.mutation = rates.clone();
            }
            Mutation::Adaptive => {
                // Apply every mutation with the given rate.
                // c' = c * e^(learning rate * N(0, 1))
                let lrate = 1.0 / (self.mems.len() as f64).sqrt();
                mutate_rate(&mut s.params.mutation, 1.0, |v| {
                    mutate_lognorm(v, lrate).clamp(0.0, 1.0)
                });
            }
        };
        Self::check_weights(&s.params.mutation, E::NUM_MUTATION)?;
        for (idx, &rate) in s.params.mutation.iter().enumerate() {
            eval.mutate(&mut s.genome, rate, idx);
        }
        Ok(())
    }


    pub fn next_gen<E: Evaluator<Genome = G>>(
        &self,
        genfn: &mut (dyn RandGenome<G> + '_),
        stagnant: bool,
        cfg: &Cfg,
        eval: &E,
    ) -> Result<UnevaluatedGen<G>> {
        // Pick survivors:
        let mut new_mems = self.survivors(cfg.survival, cfg);
        // Min here to avoid underflow - can happen if we produce too many parents.
        new_mems.reserve(cfg.pop_size);

        // If stagnant, fill with random individuals.
        if stagnant {
            let num = match cfg.replacement {
                Replacement::ReplaceChildren(prop) => {
                    let remaining = cfg.pop_size as f64 - new_mems.len() as f64;
                    (prop * remaining).ceil().max(0.0) as usize
                }
            };
            for _ in 0..num {
                new_mems.push(Mem::new::<E>((*genfn)(), cfg));
            }
        }

        // If DisallowDuplicates on, try up to NUM_TRIES times
        // to fill the population up.
        const NUM_TRIES: usize = 3;
        for _ in 0..NUM_TRIES {
            // Reproduce.
            while new_mems.len() < cfg.pop_size {
                let [mut s1, mut s2] = self.selection(cfg.selection);
                self.crossover(&cfg.crossover, eval, &mut s1, &mut s2).unwrap();
                self.mutation(&cfg.mutation, eval, &mut s1).unwrap();
                self.mutation(&cfg.mutation, eval, &mut s2).unwrap();
                new_mems.push(s1);
                new_mems.push(s2);
            }

            // Remove duplicates if we need to.
            if cfg.duplicates == Duplicates::DisallowDuplicates {
                new_mems.sort_unstable_by(|a, b| a.genome.partial_cmp(&b.genome).unwrap());
                new_mems.dedup_by(|a, b| a.genome.eq(&b.genome));
            }
        }
        Ok(UnevaluatedGen::new(new_mems))
    }
}