452 lines
14 KiB
C++
452 lines
14 KiB
C++
#pragma once
|
|
|
|
#include <algorithm>
|
|
#include <cfloat>
|
|
#include <cstdlib>
|
|
|
|
#include "util.h"
|
|
#include "sync.h"
|
|
#include "rand.h"
|
|
|
|
using namespace sync;
|
|
using namespace std;
|
|
|
|
namespace genetic {
|
|
|
|
template <class T> struct Stats;
|
|
template <class T> struct Strategy;
|
|
struct CellTracker;
|
|
|
|
template <class T> T run(Strategy<T>);
|
|
|
|
template <class T> struct Strategy {
|
|
// Number of worker threads that will be evaluating cell fitness
|
|
int num_threads;
|
|
|
|
// Period of print statements (in seconds)
|
|
float stats_print_period_s;
|
|
|
|
// Size of the population pool per sim thread
|
|
int num_cells_per_thread;
|
|
|
|
// Number of times (epochs) to run the algorithm
|
|
int num_generations;
|
|
|
|
// Each thread will integrate the best globally performing cell
|
|
bool share_breakthroughs;
|
|
|
|
// How many generations to explore before resyncing with the global best
|
|
int share_breakthrough_gen_period;
|
|
|
|
bool test_all; // Sets whether or not every cell's fitness is evaluated every
|
|
// generation
|
|
float test_chance; // Chance to test any given cell's fitness. Relevant only
|
|
// if test_all is false.
|
|
bool enable_crossover; // Cells that score well in the evaluation stage
|
|
// produce children that replace low-scoring cells
|
|
int crossover_parent_num; // Number of unique high-scoring parents in a
|
|
// crossover call.
|
|
int crossover_parent_stride; // Number of parents to skip over when moving to
|
|
// the next set of parents. A stride of 1 would
|
|
// produce maximum overlap because the set of
|
|
// parents would only change by one every
|
|
// crossover.
|
|
int crossover_children_num; // Number of children to expect the user to
|
|
// produce in the crossover function.
|
|
bool enable_mutation; // Cells may be mutated
|
|
// before fitness evaluation
|
|
float mutation_chance; // Chance for any given cell to be mutated cells during
|
|
// the mutation
|
|
uint64_t rand_seed;
|
|
bool higher_fitness_is_better; // Sets whether or not to consider higher
|
|
// fitness values better or worse. Set this to
|
|
// false if fitness is an error function.
|
|
|
|
// User defined functions
|
|
T (*make_default_cell)();
|
|
void (*mutate)(T &cell_to_modify);
|
|
void (*crossover)(const Array<T *> parents, const Array<T *> out_children);
|
|
float (*fitness)(const T &cell);
|
|
};
|
|
|
|
template<class T> struct Stats {
|
|
DynArray<T> best_cells;
|
|
DynArray<float> best_cell_fitness;
|
|
int gen;
|
|
bool done;
|
|
DynArray<TimeSpan> gen_time;
|
|
DynArray<TimeSpan> crossover_time;
|
|
DynArray<TimeSpan> mutate_time;
|
|
DynArray<TimeSpan> fitness_time;
|
|
DynArray<TimeSpan> sorting_time;
|
|
Mutex m;
|
|
};
|
|
|
|
struct CellTracker {
|
|
float score;
|
|
int cellid;
|
|
};
|
|
|
|
template<class T>
|
|
struct WorkerThreadArgs {
|
|
Strategy<T> strat;
|
|
Array<T> cells;
|
|
Array<CellTracker> trackers;
|
|
Stats<T> *stats;
|
|
|
|
Mutex m;
|
|
float *best_global_score;
|
|
T* best_global_cell;
|
|
};
|
|
|
|
template<class T> T* _cellp(Array<T> cells, CellTracker tracker) { return &cells[tracker.cellid]; }
|
|
|
|
template <class T> DWORD worker(LPVOID args) {
|
|
// Unpack everything...
|
|
WorkerThreadArgs<T>* worker_args = static_cast<WorkerThreadArgs<T>*>(args);
|
|
Strategy<T> strat = worker_args->strat;
|
|
Array<T> cells = worker_args->cells;
|
|
Array<CellTracker> trackers = worker_args->trackers;
|
|
Stats<T> &stats = *worker_args->stats;
|
|
float* best_global_score = worker_args->best_global_score;
|
|
T* best_global_cell = worker_args->best_global_cell;
|
|
Mutex best_m = worker_args->m;
|
|
|
|
// Prepare crossover operations as these will be the same every time except
|
|
// for the exact cell pointers
|
|
int npar = strat.crossover_parent_num;
|
|
int nchild = strat.crossover_children_num;
|
|
Array<T*> parents = make_array<T*>(npar);
|
|
Array<T*> children = make_array<T*>(nchild);
|
|
|
|
bool gt = strat.higher_fitness_is_better; // Writing strat.higher... is annoying
|
|
|
|
// printf("Core: %d\n", get_affinity());
|
|
|
|
TimeSpan start, diff, gen_start;
|
|
while(stats.gen < strat.num_generations) {
|
|
gen_start = now();
|
|
|
|
// 0. Share/Integrate global breakthrough
|
|
if (strat.share_breakthroughs && (stats.gen + get_affinity()) % strat.share_breakthrough_gen_period) {
|
|
lock(best_m);
|
|
if (better(gt, front(trackers).score, *best_global_score) != *best_global_score) {
|
|
// Share
|
|
*best_global_cell = *_cellp(cells, trackers[0]);
|
|
*best_global_score = trackers[0].score;
|
|
|
|
} else {
|
|
// Integrate
|
|
*_cellp(cells, trackers[0]) = *best_global_cell;
|
|
trackers[0].score = *best_global_score;
|
|
}
|
|
unlock(best_m);
|
|
}
|
|
|
|
// 1. crossover
|
|
start = now();
|
|
if (strat.enable_crossover) {
|
|
int parent_end = npar;
|
|
int child_begin = trackers.len-nchild;
|
|
while (parent_end <= child_begin) {
|
|
|
|
// Get pointers to all the parent cells
|
|
for (int i = parent_end-npar; i < parent_end; i++) {
|
|
T* cell = _cellp(cells, trackers[i]);
|
|
assert(cell != NULL);
|
|
parents[i - (parent_end-npar)] = cell;
|
|
}
|
|
|
|
// Get pointers to all the child cells (these will be overwritten)
|
|
for (int i = child_begin; i < child_begin+nchild; i++) {
|
|
T* cell = _cellp(cells, trackers[i]);
|
|
assert(cell != NULL);
|
|
children[i-child_begin] = cell;
|
|
}
|
|
CrossoverJob<T> cj = {parents, children};
|
|
TaggedJob<T> job;
|
|
job.data.c=cj;
|
|
job.type=JobType::CROSSOVER;
|
|
q.jobs[q.write_i++] = job;
|
|
parent_end += strat.crossover_parent_stride;
|
|
child_begin -= nchild;
|
|
}
|
|
}
|
|
lock(stats.m);
|
|
append(stats.crossover_time, now() - start);
|
|
unlock(stats.m);
|
|
|
|
|
|
// 2. mutate
|
|
start = now();
|
|
for (int i = 0; i < trackers.len; i++) {
|
|
if (abs(norm_rand(strat.rand_seed)) < strat.mutation_chance) {
|
|
strat.mutate(cells[trackers[i].cellid]);
|
|
}
|
|
}
|
|
lock(stats.m);
|
|
append(stats.mutate_time, now() - start);
|
|
unlock(stats.m);
|
|
|
|
// 3. evaluate
|
|
start = now();
|
|
if (strat.test_all) {
|
|
for (int i = 0; i < trackers.len; i++) {
|
|
FitnessJob<T> fj = {&cells[trackers[i].cellid], &trackers[i]};
|
|
TaggedJob<T> job;
|
|
job.data.f=fj;
|
|
job.type=JobType::FITNESS;
|
|
if (i == trackers.len-1) lock(q.m);
|
|
q.jobs[q.write_i++] = job;
|
|
if (i == trackers.len-1) { q.done_writing = true; unlock(q.m); }
|
|
}
|
|
} else {
|
|
lock(q.m);
|
|
for (int i = 0; i < trackers.len; i++) {
|
|
if (abs(norm_rand(strat.rand_seed)) < strat.test_chance) {
|
|
FitnessJob<T> fj = {&cells[trackers[i].cellid], &trackers[i]};
|
|
TaggedJob<T> job;
|
|
job.data.f=fj;
|
|
job.type=JobType::FITNESS;
|
|
q.jobs[q.write_i++] = job;
|
|
}
|
|
}
|
|
q.done_writing = true;
|
|
unlock(q.m);
|
|
}
|
|
lock(stats.m);
|
|
append(stats.fitness_time, now() - start);
|
|
unlock(stats.m);
|
|
|
|
// 4. sort
|
|
start = now();
|
|
std::sort(&trackers[0], &trackers[trackers.len-1], [strat](CellTracker &a, CellTracker &b){ return better(strat.higher_fitness_is_better, a.score, b.score) == a.score; });
|
|
lock(stats.m);
|
|
append(stats.sorting_time, now() - start);
|
|
|
|
append(stats.best_cells, cells[trackers[0].cellid]);
|
|
append(stats.best_cell_fitness, trackers[0].score);
|
|
append(stats.gen_time, now() - gen_start);
|
|
stats.gen++;
|
|
unlock(stats.m);
|
|
}
|
|
stats.done = true;
|
|
return 0;
|
|
}
|
|
|
|
template <class T> T run(Strategy<T> strat) {
|
|
Array<Stats<T>> stats = make_array<Stats<T>>(strat.num_threads);
|
|
Array<Thread> threads = make_array<Thread>(strat.num_threads);
|
|
Array<WorkerThreadArgs<T>> args = make_array<WorkerThreadArgs<T>>(strat.num_threads);
|
|
|
|
float best_global_score = strat.higher_fitness_is_better ? FLT_MIN : FLT_MAX;
|
|
T best_global_cell;
|
|
|
|
allow_all_processors();
|
|
set_affinity(0);
|
|
|
|
for (int i = 0; i < strat.num_threads; i++) {
|
|
stats[i] = {
|
|
.best_cells=make_dynarray<T>(strat.num_generations),
|
|
.best_cell_fitness=make_dynarray<float>(strat.num_generations),
|
|
.gen_time=make_dynarray<TimeSpan>(strat.num_generations),
|
|
.crossover_time=make_dynarray<TimeSpan>(strat.num_generations),
|
|
.mutate_time=make_dynarray<TimeSpan>(strat.num_generations),
|
|
.fitness_time=make_dynarray<TimeSpan>(strat.num_generations),
|
|
.sorting_time=make_dynarray<TimeSpan>(strat.num_generations),
|
|
.m=make_mutex()
|
|
};
|
|
Array<T> cells = make_array<T>(strat.num_threads*strat.num_cells_per_thread);
|
|
Array<CellTracker> trackers = make_array<CellTracker>(strat.num_cells_per_thread);
|
|
for (int i = 0; i < strat.num_cells_per_thread; i++) {
|
|
cells[i] = strat.make_default_cell();
|
|
trackers[i] = {0, i};
|
|
}
|
|
|
|
args[i].strat=strat;
|
|
args[i].cells=cells;
|
|
args[i].trackers=trackers;
|
|
args[i].stats=&stats[i];
|
|
args[i].best_global_score=&best_global_score;
|
|
args[i].best_global_cell=&best_global_cell;
|
|
args[i].m = make_mutex();
|
|
|
|
threads[i] = make_thread(worker<T>, &args[i], i+1);
|
|
}
|
|
|
|
// We are the stats thread
|
|
bool complete = false;
|
|
while (!complete) {
|
|
sleep(from_s(strat.stats_print_period_s));
|
|
|
|
printf("**********************\n");
|
|
float g_avg_gen_time = 0;
|
|
float g_avg_crossover_time = 0;
|
|
float g_avg_mutate_time = 0;
|
|
float g_avg_fitness_time = 0;
|
|
float g_avg_sorting_time = 0;
|
|
float g_avg_overhead_time = 0;
|
|
float g_progress_per = 0;
|
|
float g_best_fitness = strat.higher_fitness_is_better ? FLT_MIN : FLT_MAX;
|
|
|
|
complete = true;
|
|
|
|
for (int i = 0; i < stats.len; i++) {
|
|
lock(stats[i].m);
|
|
complete &= stats[i].done;
|
|
|
|
int end = stats[i].gen_time.end-1;
|
|
|
|
float gen_time = to_s(stats[i].gen_time[end]);
|
|
float crossover_time = to_s(stats[i].crossover_time[end]);
|
|
float mutate_time = to_s(stats[i].mutate_time[end]);
|
|
float fitness_time = to_s(stats[i].fitness_time[end]);
|
|
float sorting_time = to_s(stats[i].sorting_time[end]);
|
|
float progress_per = static_cast<float>(stats[i].gen) / static_cast<float>(strat.num_generations) * 100;
|
|
float best_score = back(stats[i].best_cell_fitness);
|
|
|
|
float overhead = max(0, gen_time - (crossover_time + mutate_time + fitness_time + sorting_time));
|
|
|
|
float overhead_per = overhead / gen_time * 100;
|
|
|
|
g_avg_gen_time += gen_time;
|
|
g_avg_crossover_time += crossover_time;
|
|
g_avg_mutate_time += mutate_time;
|
|
g_avg_fitness_time += fitness_time;
|
|
g_avg_sorting_time += sorting_time;
|
|
g_progress_per += progress_per;
|
|
g_best_fitness = better(strat.higher_fitness_is_better, best_score, g_best_fitness);
|
|
|
|
g_avg_overhead_time += overhead;
|
|
|
|
printf("%d, Progress %d/%d, Top: %.5e, Overhead Per: %.4f%%, Gen: %.4f, Overhead: %.4f, Cross: %.4f (s), Mutate: %.4f (s), Fitness: %.4f (s), Sorting: %.4f (s)\n", i, stats[i].gen, strat.num_generations, best_score, overhead_per, gen_time, overhead, crossover_time, mutate_time, fitness_time, sorting_time);
|
|
unlock(stats[i].m);
|
|
}
|
|
|
|
g_avg_gen_time /= stats.len;
|
|
g_avg_crossover_time /= stats.len;
|
|
g_avg_mutate_time /= stats.len;
|
|
g_avg_fitness_time /= stats.len;
|
|
g_avg_sorting_time /= stats.len;
|
|
g_progress_per /= stats.len;
|
|
|
|
g_avg_overhead_time /= stats.len;
|
|
|
|
float g_avg_overhead_per = g_avg_overhead_time / g_avg_gen_time * 100;
|
|
|
|
printf("GLOBAL, Progress %.1f%%, Top: %.5e, Overhead Per: %.4f%%, Gen: %.4f, Overhead: %.4f, Cross: %.4f (s), Mutate: %.4f (s), Fitness: %.4f (s), Sorting: %.4f (s)\n", g_progress_per, g_best_fitness, g_avg_overhead_per, g_avg_gen_time, g_avg_overhead_time, g_avg_crossover_time, g_avg_mutate_time, g_avg_fitness_time, g_avg_sorting_time);
|
|
|
|
if (complete) break;
|
|
}
|
|
|
|
for (int i = 0; i < threads.len; i++) {
|
|
join(threads[i]);
|
|
}
|
|
|
|
T best_cell;
|
|
// TODO: bad
|
|
float best_score = strat.higher_fitness_is_better ? FLT_MIN : FLT_MAX;
|
|
for (int i = 0; i < stats.len; i++) {
|
|
float score = back(stats[i].best_cell_fitness);
|
|
if (strat.higher_fitness_is_better ? score > best_score : score < best_score) {
|
|
best_cell = back(stats[i].best_cells);
|
|
best_score = score;
|
|
}
|
|
}
|
|
|
|
return best_cell;
|
|
}
|
|
|
|
template<class T> WorkQueue<T> make_work_queue(int len, int batch_size) {
|
|
return {
|
|
.jobs=make_array<TaggedJob<T>>(len),
|
|
.read_i=0,
|
|
.write_i=0,
|
|
.batch_size=batch_size,
|
|
.done_writing=false,
|
|
.work_complete=false,
|
|
.m=make_mutex(),
|
|
.done=make_condition_var(),
|
|
.jobs_ready=make_condition_var()
|
|
};
|
|
}
|
|
|
|
template<class T> bool tryget_job_batch(WorkQueue<T> &q, Array<TaggedJob<T>>* out_batch, bool* out_batch_is_end) {
|
|
lock(q.m);
|
|
if (q.stop) {
|
|
unlock(q.m);
|
|
return false;
|
|
}
|
|
|
|
// Keep waiting till jobs are available
|
|
while (q.read_i >= q.write_i) {
|
|
wait(q.jobs_ready, q.m, infinite_ts);
|
|
|
|
if (q.stop) {
|
|
unlock(q.m);
|
|
return false;
|
|
}
|
|
}
|
|
|
|
// Yay! Let's grab some jobs to do
|
|
|
|
// If the batch we're about to grab moves read_i to write_i and the producer
|
|
// is done writing, we should let our callee know it's handling this gen's last
|
|
// batch know that way it sets work_complete and signals done.
|
|
*out_batch_is_end = q.done_writing && q.read_i + q.batch_size >= q.write_i;
|
|
|
|
out_batch->data = &q.jobs[q.read_i];
|
|
out_batch->len = min(q.batch_size, q.write_i - q.read_i);
|
|
q.read_i += q.batch_size;
|
|
unlock(q.m);
|
|
return true;
|
|
}
|
|
|
|
template<class T>
|
|
void work_batch(Array<TaggedJob<T>> batch, Strategy<T> &s) {
|
|
for (int i = 0; i < batch.len; i++) {
|
|
switch (batch[i].type) {
|
|
case JobType::MUTATE: {
|
|
MutateJob<T> mj = batch[i].data.m;
|
|
s.mutate(*mj.cell);
|
|
} break;
|
|
case JobType::CROSSOVER: {
|
|
CrossoverJob<T> cj = batch[i].data.c;
|
|
s.crossover(cj.parents, cj.children);
|
|
} break;
|
|
case JobType::FITNESS: {
|
|
FitnessJob<T> fj = batch[i].data.f;
|
|
fj.track->score = s.fitness(*fj.cell);
|
|
} break;
|
|
default: {
|
|
assert(false);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
template<class T>
|
|
DWORD worker(LPVOID args) {
|
|
WorkerThreadArgs<T>* wa = static_cast<WorkerThreadArgs<T>*>(args);
|
|
WorkQueue<T> &q = wa->q;
|
|
Strategy<T> &s = wa->s;
|
|
|
|
// These are written by tryget_job_batch
|
|
bool batch_is_end;
|
|
Array<TaggedJob<T>> batch;
|
|
|
|
while (tryget_job_batch(q, &batch, &batch_is_end)) {
|
|
work_batch(batch, s);
|
|
if (batch_is_end) {
|
|
lock(q.m);
|
|
q.work_complete = true;
|
|
wake_one(q.done);
|
|
unlock(q.m);
|
|
}
|
|
}
|
|
|
|
return NULL;
|
|
}
|
|
|
|
} // namespace genetic
|