KScalar stats_inbreeding_depression (KConfig K)
// Compute inbreeding depression (1 - self/outcross) from
// the load class data for selfed and outcrossed progeny.
{
const char * thisfunction = "stats_inbreeding_depression";
KInt i, j, g;
KArray a, via_fgam, via_mgam;
KVector1 thismgam, thisfgam, allmgam, allfgam;
KScalar w_self, w_out, ibd, thisibd;
if (K->genotypes != 1) {
not_implemented(thisfunction, "genotypes != 1");
}
ibd = 0.0;
apply_gametes_full(K, allmgam, allfgam, K->x);
for (i=0; i <= K->MI; i++) {
for (j=0; j <= K->MJ; j++) {
for (g=0; g < K->genotypes; g++) {
/* compute fitness of selfed and outcrossed
** progeny of this load class by creating dummy
** arrays representing the progeny resulting
** from matings involving just these load classes.
**
** Selfing is straightforward.
**
** For outcrossing, matings involving both female
** and male gametes must be considered. This is
** done by mating female gametes produced by this
** load class with the population male gametes,
** and the male gametes produced by this load
** class with the population female gametes.
*/
if (K->x[i][j][g] == 0.0)
continue;
/* fill_KArray(K, a1, 0.0);
** a1[i][j][g] = K->x[i][j][g];
** apply_self_progeny(K, a2, a1);
*/
apply_self_progeny_stats(K, a, K->x[i][j][g], i, j, g);
w_self = cumulative_fitness(K, a);
apply_gametes_stats(K, thismgam, thisfgam, K->x[i][j][g], i, j, g);
/* apply_gametes(K, allmgam, allfgam, K->x); */
apply_zygotes(K, via_fgam, allmgam, thisfgam);
apply_zygotes(K, via_mgam, thismgam, allfgam);
w_out = cumulative_fitness(K, via_fgam) +
cumulative_fitness(K, via_mgam);
w_out *= 0.5;
thisibd = 1.0 - (w_self / w_out);
ibd += thisibd * K->x[i][j][g];
}
}
}
return ibd;
}
void GeneticAlgorithm<Model>::Run(
const GeneticFitnessFunction<Model>& fitness_function,
GeneticSearchStrategy<Model>& searcher,
size_t population_size,
size_t breeds_per_generation) {
// A lambda for sorting the population based on fitness via std::sort.
auto fitness_sort =
[](const member_t& a, const member_t& b) {
return a.fitness() > b.fitness();
};
// Randomly selects an index based off of a given probability distribution.
// The CDF does not need to be normalized, just monotonically increasing.
auto select_index =
[](const std::vector<double>& cdf) {
double p = (double)rand() / RAND_MAX * cdf.back();
int i = 0;
while (p > cdf[i]) i++;
return i;
};
// Given the population, construct a cumulative distribution function based
// on the fitness of the members. Despite the name, the CDF is not
// normalized.
auto compute_fitness_cdf =
[&fitness_sort](const population_t& population,
std::vector<double>& fitness_cdf) {
fitness_cdf.resize(population.size());
auto minmax_members = std::minmax_element(
std::begin(population), std::end(population), fitness_sort);
double min_fitness = minmax_members.second->fitness();
double max_fitness = minmax_members.first->fitness();
double range = max_fitness - min_fitness;
double offset = (range == 0)
? 1 - min_fitness : -min_fitness;
int index = 0;
for (const auto& tmp : population) {
fitness_cdf[index] = (index == 0) ? tmp.fitness() + offset
: fitness_cdf[index-1] + tmp.fitness() + offset;
index++;
}
};
// Initialize the population.
{
std::lock_guard<std::mutex> lock(generation_lock_);
population_.clear();
population_.reserve(population_size + breeds_per_generation);
for (size_t i = 0; i < population_size; i++) {
member_t member(std::move(searcher.Introduce(fitness_function)), 0);
fitness_function(member);
population_.push_back(std::move(member));
}
std::sort(std::begin(population_), std::end(population_), fitness_sort);
}
std::vector<double> cumulative_fitness(population_size, 0);
std::vector<double> new_cumulative_fitness(
population_size + breeds_per_generation, 0);
std::vector<size_t> selection_indices(population_size, 0);
generation_num_ = 0;
// Main loop of the genetic algorithm.
do {
std::lock_guard<std::mutex> lock(generation_lock_);
running_ = true;
generation_num_++;
compute_fitness_cdf(population_, cumulative_fitness);
// Construct the new population members from the existing ones.
for (size_t i = 0; i < breeds_per_generation; i++) {
int index1 = select_index(cumulative_fitness);
int index2 = index1;
while (index1 == index2) {
index2 = select_index(cumulative_fitness);
}
do {
member_t member(std::move(searcher.Crossover(population_[index1],
population_[index2])));
searcher.Mutate(member);
// Only add if the model is valid.
if (fitness_function(member)) {
population_.push_back(std::move(member));
break;
}
} while (true);
}
// If we extra members hanging around from the previous generation, select
// the best population member and fill out the population using the fitness
// to weight selection.
if (population_.size() > population_size) {
compute_fitness_cdf(population_, new_cumulative_fitness);
selection_indices[0] = 0;
for (size_t i = 1; i < population_size; i++) {
bool unique = false;
size_t index = 0;
while (!unique) {
index = select_index(new_cumulative_fitness);
unique = true;
for (int j = i - 1; j >= 0; j--) {
if (selection_indices[j] == index) {
unique = false;
break;
}
}
}
selection_indices[i] = index;
}
// Sort the selection indices in ascending order, so we can select them in
// the population in one pass.
std::sort(std::begin(selection_indices), std::end(selection_indices));
// Construct the new population based on the selections.
for (size_t i = 0; i < population_size; i++) {
if (i == selection_indices[i]) continue;
std::swap(population_[i], population_[selection_indices[i]]);
}
population_.erase(std::begin(population_) + population_size,
std::end(population_));
}
std::sort(std::begin(population_), std::end(population_), fitness_sort);
} while (searcher.ShouldContinue(population_, generation_num_));
running_ = false;
}
