v/examples/flappylearning/modules/neuroevolution/neuronevolution.v

288 lines
5.3 KiB
V

module neuroevolution
import rand
import math
fn random_clamped() f64 {
return rand.f64() * 2 - 1
}
pub fn activation(a f64) f64 {
ap := (-a) / 1
return 1 / (1 + math.exp(ap))
}
fn round(a int, b f64) int {
return int(math.round(f64(a) * b))
}
struct Neuron {
mut:
value f64
weights []f64
}
fn (mut n Neuron) populate(nb int) {
for _ in 0 .. nb {
n.weights << random_clamped()
}
}
struct Layer {
id int
mut:
neurons []Neuron
}
fn (mut l Layer) populate(nb_neurons int, nb_inputs int) {
for _ in 0 .. nb_neurons {
mut n := Neuron{}
n.populate(nb_inputs)
l.neurons << n
}
}
pub struct Network {
mut:
layers []Layer
}
fn (mut n Network) populate(network []int) {
assert network.len >= 2
input := network[0]
hiddens := network[1..network.len - 1]
output := network[network.len - 1]
mut index := 0
mut previous_neurons := 0
mut input_layer := Layer{
id: index
}
input_layer.populate(input, previous_neurons)
n.layers << input_layer
previous_neurons = input
index++
for hidden in hiddens {
mut hidden_layer := Layer{
id: index
}
hidden_layer.populate(hidden, previous_neurons)
previous_neurons = hidden
n.layers << hidden_layer
index++
}
mut output_layer := Layer{
id: index
}
output_layer.populate(output, previous_neurons)
n.layers << output_layer
}
fn (n Network) get_save() Save {
mut save := Save{}
for layer in n.layers {
save.neurons << layer.neurons.len
for neuron in layer.neurons {
for weight in neuron.weights {
save.weights << weight
}
}
}
return save
}
fn (mut n Network) set_save(save Save) {
mut previous_neurons := 0
mut index := 0
mut index_weights := 0
n.layers = []
for save_neuron in save.neurons {
mut layer := Layer{
id: index
}
layer.populate(save_neuron, previous_neurons)
for mut neuron in layer.neurons {
for i in 0 .. neuron.weights.len {
neuron.weights[i] = save.weights[index_weights]
index_weights++
}
}
previous_neurons = save_neuron
index++
n.layers << layer
}
}
pub fn (mut n Network) compute(inputs []f64) []f64 {
assert n.layers.len > 0
assert inputs.len == n.layers[0].neurons.len
for i, input in inputs {
n.layers[0].neurons[i].value = input
}
mut prev_layer := n.layers[0]
for i in 1 .. n.layers.len {
for j, neuron in n.layers[i].neurons {
mut sum := f64(0)
for k, prev_layer_neuron in prev_layer.neurons {
sum += prev_layer_neuron.value * neuron.weights[k]
}
n.layers[i].neurons[j].value = activation(sum)
}
prev_layer = n.layers[i]
}
mut outputs := []f64{}
mut last_layer := n.layers[n.layers.len - 1]
for neuron in last_layer.neurons {
outputs << neuron.value
}
return outputs
}
struct Save {
mut:
neurons []int
weights []f64
}
fn (s Save) clone() Save {
mut save := Save{}
save.neurons << s.neurons
save.weights << s.weights
return save
}
struct Genome {
score int
network Save
}
struct Generation {
mut:
genomes []Genome
}
fn (mut g Generation) add_genome(genome Genome) {
mut i := 0
for gg in g.genomes {
if genome.score > gg.score {
break
}
i++
}
g.genomes.insert(i, genome)
}
fn (g1 Genome) breed(g2 Genome, nb_child int) []Save {
mut datas := []Save{}
for _ in 0 .. nb_child {
mut data := g1.network.clone()
for i, weight in g2.network.weights {
if rand.f64() <= 0.5 {
data.weights[i] = weight
}
}
for i, _ in data.weights {
if rand.f64() <= 0.1 {
data.weights[i] += (rand.f64() * 2 - 1) * 0.5
}
}
datas << data
}
return datas
}
fn (g Generation) next(population int) []Save {
mut nexts := []Save{}
if population == 0 {
return nexts
}
keep := round(population, 0.2)
for i in 0 .. keep {
if nexts.len < population {
nexts << g.genomes[i].network.clone()
}
}
random := round(population, 0.2)
for _ in 0 .. random {
if nexts.len < population {
mut n := g.genomes[0].network.clone()
for k, _ in n.weights {
n.weights[k] = random_clamped()
}
nexts << n
}
}
mut max := 0
out: for {
for i in 0 .. max {
mut childs := g.genomes[i].breed(g.genomes[max], 1)
for c in childs {
nexts << c
if nexts.len >= population {
break out
}
}
}
max++
if max >= g.genomes.len - 1 {
max = 0
}
}
return nexts
}
pub struct Generations {
pub:
population int
network []int
mut:
generations []Generation
}
fn (mut gs Generations) first() []Save {
mut out := []Save{}
for _ in 0 .. gs.population {
mut nn := Network{}
nn.populate(gs.network)
out << nn.get_save()
}
gs.generations << Generation{}
return out
}
fn (mut gs Generations) next() []Save {
assert gs.generations.len > 0
gen := gs.generations[gs.generations.len - 1].next(gs.population)
gs.generations << Generation{}
return gen
}
fn (mut gs Generations) add_genome(genome Genome) {
assert gs.generations.len > 0
gs.generations[gs.generations.len - 1].add_genome(genome)
}
fn (mut gs Generations) restart() {
gs.generations = []
}
pub fn (mut gs Generations) generate() []Network {
saves := if gs.generations.len == 0 { gs.first() } else { gs.next() }
mut nns := []Network{}
for save in saves {
mut nn := Network{}
nn.set_save(save)
nns << nn
}
if gs.generations.len >= 2 {
gs.generations.delete(0)
}
return nns
}
pub fn (mut gs Generations) network_score(network Network, score int) {
gs.add_genome(Genome{
score: score
network: network.get_save()
})
}