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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
use crate::{Config, ANN};
use rand::{thread_rng, Rng};
use rand_distr::{Distribution, Normal, Uniform};
pub(crate) const DEFAULT_FIT: f64 = std::f64::MIN;
/// stores all sub-populations and their fitness values
pub struct Population {
config: Config,
network_topology: ANN,
// stores an n x m matrix of chromosome value and corresponding fitness
// NOTE: indexing is reverse of what the paper refers to:
// sub_populations[j] is a complete network
// sub_populations[j][i] is the ith weight of the jth network
sub_populations: Vec<Vec<(f64, f64)>>,
// number of species (components, sub-populations) or just weights and biases of the network topology
n: usize,
// m sub-genotypes or population size,
// so for each weight (or bias) in the nn topology there are m different weight (or bias) variations
m: usize,
current_generation: usize,
}
impl Population {
/// Set the mutation probability of the population
/// panics in debug mode if mp < 0.0 || mp > 1.0
pub fn set_mutation_prob(&mut self, mp: f64) {
self.config.mutation_prob = mp;
}
/// Set the mutation strength of the population
/// panics in debug mode if mp < 0.0 || mp > 1.0
pub fn set_mutation_strength(&mut self, ms: f64) {
self.config.mutation_strength = ms;
}
/// Set the perturb probability of the population
/// panics in debug mode if pp < 0.0 || pp > 1.0
pub fn set_perturb_prob(&mut self, pp: f64) {
self.config.perturb_prob = pp;
}
/// Return a neural network at index j
pub fn get_network(&self, j: usize) -> ANN {
let genes: Vec<f64> = self.sub_populations[j].iter().map(|(v, _f)| *v).collect();
let mut nn = self.network_topology.clone();
nn.set_genes(&genes);
nn
}
/// Return the fitness of a given network
pub fn get_network_fitness(&self, j: usize) -> f64 {
self.sub_populations[j].iter().map(|(_, f)| *f).sum()
}
/// Create a new population with a given config and network
pub fn new(config: Config, nn: &ANN) -> Population {
let n: usize = nn.num_genes();
let m: usize = config.pop_size;
let mut sub_populations: Vec<Vec<(f64, f64)>> = Vec::new();
for _i in 0..config.pop_size {
let genes: Vec<f64> = nn.randomize().genes();
let mut chromosome: Vec<(f64, f64)> = Vec::with_capacity(n);
for v in &genes {
chromosome.push((*v, DEFAULT_FIT));
}
sub_populations.push(chromosome);
}
return Population {
config,
network_topology: nn.clone(),
sub_populations,
n,
m,
current_generation: 0,
};
}
/// Perform a single generational evolutionary step in a given environment
/// assumes all network have been evaluated and their fitness updated
pub fn evolve(&mut self) {
let offspring = self.spawn_offspring();
self.replace_and_permute(&offspring);
}
/// Update existing chromosome fitnesses with the new network fits
pub fn update_fitnesses(&mut self, new_fits: &Vec<f64>) {
let g: f64 = self.current_generation as f64;
for (j, new_fit) in new_fits.iter().enumerate() {
self.sub_populations[j].iter_mut().for_each(|(_, old_fit)| {
*old_fit *= g; // undo the mean
*old_fit += new_fit; // add new fitness
*old_fit /= g + 1.0; // redo the mean
});
}
self.current_generation += 1;
}
/// Create offspring population from top n% of population
fn spawn_offspring(&mut self) -> Vec<Vec<(f64, f64)>> {
// find parents with highest mean fitness based on elite_threshold in config
// compute mean fitness of each column (network mean)
let mut mean_fits: Vec<(usize, f64)> = Vec::with_capacity(self.m);
for j in 0..self.m {
let mean_fit: f64 =
self.sub_populations[j].iter().map(|(_, f)| *f).sum::<f64>() / self.n as f64;
mean_fits.push((j, mean_fit));
}
// sort mean_fits by fitness, lower indices will have lower fitness
mean_fits.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap());
let elite_threshold: usize =
(self.m as f64 * self.config.top_ratio_to_recombine).floor() as usize;
let mut o: Vec<Vec<(f64, f64)>> = Vec::with_capacity(
(self.config.top_ratio_to_recombine * self.m as f64).ceil() as usize,
);
for (j, _f) in mean_fits.iter().take(elite_threshold) {
o.push(self.sub_populations[*j].clone());
}
self.crossover(&mut o);
self.mutate(&mut o);
o
}
/// Perform crossover mutation operator on offspring population,
fn crossover(&mut self, o: &mut Vec<Vec<(f64, f64)>>) {
// create vec of deranged indices, not sure if actually good but should be good enough
let deranged = random_derangement(o.len());
let mut rng = thread_rng();
let d = Normal::new(self.n as f64 / 2.0, self.n as f64 * 0.33).unwrap();
for (p1, p2) in (0..o.len()).zip(&deranged) {
// TODO: different user defined crossover methods
let cross_p: f64 = d.sample(&mut rng);
// clip to min and max
let crossover_point: usize = if cross_p < 0.0 {
0
} else if cross_p > self.n as f64 {
self.n
} else {
cross_p.round() as usize
};
// perform single point crossover
for i in 0..crossover_point {
// switch chromosomes and fitness between parent 1 and 2
let old_p1_vals = self.sub_populations[p1][i];
self.sub_populations[p1][i] = self.sub_populations[*p2][i];
self.sub_populations[*p2][i] = old_p1_vals;
}
}
}
/// Perform a mutation operator on offspring population,
/// by either perturbing or completely replacing values
fn mutate(&self, o: &mut Vec<Vec<(f64, f64)>>) {
// TODO: user defined mutation distribution in case of pertubation
let d = Normal::new(0.0, 0.4).unwrap();
let mut rng = rand::thread_rng();
o.iter_mut().flatten().for_each(|(v, f)| {
if rng.gen::<f64>() < self.config.mutation_prob {
if rng.gen::<f64>() < self.config.perturb_prob {
*v += rng.sample(d) * self.config.mutation_strength;
} else {
*v = (rng.gen::<f64>() * 2.0 - 1.0) * self.config.mutation_strength;
}
*f = DEFAULT_FIT;
}
});
}
/// Replace the least fit chromosome in each sub-population with newly created offspring
/// Also permute the left over original chromosomes among each other in the sub-population
fn replace_and_permute(&mut self, o: &Vec<Vec<(f64, f64)>>) {
let mut rng = thread_rng();
for i in 0..self.n {
// sort the sub-population
let mut genes: Vec<(f64, f64)> = Vec::with_capacity(self.m);
for j in 0..self.m {
genes.push(self.sub_populations[j][i]);
}
genes.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap());
let fit_threshold: f64 =
genes[(self.config.top_ratio_to_recombine * genes.len() as f64).floor() as usize].1;
let mut offspring_idx: usize = 0;
for j in 0..self.m {
if self.sub_populations[j][i].1 < fit_threshold {
// replace least fit
self.sub_populations[j][i] = o[offspring_idx][i];
offspring_idx += 1;
}
}
// assign permutation probability of left over (original) chromosomes in
// sub-population
let sub_pop_fits: Vec<f64> =
(0..self.m).map(|j| self.sub_populations[j][i].1).collect();
let mut marked: Vec<usize> = vec![];
for j in 0..self.m {
let prob: f64 = self
.config
.permutation_prob_f
.get_probability(&sub_pop_fits, self.sub_populations[j][i].1);
if rng.gen::<f64>() < prob {
// mark for permutation
marked.push(j);
}
}
if marked.len() == 0 {
return;
}
// permute marked by shifting among them
let mut temp = self.sub_populations[marked[0]][i];
for marked_idx in marked.iter().skip(1) {
// swap
let old_val = self.sub_populations[*marked_idx][i];
self.sub_populations[*marked_idx][i] = temp;
temp = old_val;
}
self.sub_populations[marked[0]][i] = temp;
}
}
}
/// Create random permutations without fixed points a.k.a. derangement
fn random_derangement(length: usize) -> Vec<usize> {
let mut rng = thread_rng();
'l: loop {
let mut v: Vec<usize> = (0..length).collect();
for j in (1..length).rev() {
let d = Uniform::new(0, j);
let p = rng.sample(d);
if v[j] == p {
continue 'l;
} else {
// swap
let old = v[j];
v[j] = v[p];
v[p] = old;
}
}
return v;
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_random_derangement() {
let length: usize = 10;
let d = random_derangement(length);
println!("d: {:?}", d);
assert_eq!(d.len(), length);
assert!(!d.iter().zip(0..length).any(|(d, i)| *d == i));
}
}