341 lines
11 KiB
C
341 lines
11 KiB
C
|
|
#include "Frozen_Lake.h"
|
|
|
|
char * action_name ="DLRU";
|
|
char * content_name ="ESCBF";
|
|
|
|
struct qlearning_params * create_qlearning_params(
|
|
double learning_rate,
|
|
double discount_factor,
|
|
double exploration_factor
|
|
){
|
|
struct qlearning_params * q_param = malloc(sizeof(struct qlearning_params));
|
|
q_param->learning_rate = learning_rate;
|
|
q_param->discount_factor = discount_factor;
|
|
q_param->exploration_factor = exploration_factor;
|
|
|
|
return q_param;
|
|
}
|
|
struct game_params * create_game_params (
|
|
size_t limit_game_number,
|
|
dimension *dim,
|
|
size_t limit_FOX_number,
|
|
size_t limit_BLOCK_number,
|
|
size_t limit_CARROT_number,
|
|
size_t limit_EPISODES_number,
|
|
size_t limit_MOVE_number
|
|
){
|
|
struct game_params * gm_param = malloc(sizeof(struct game_params));
|
|
gm_param->limit_game_number = limit_game_number;
|
|
gm_param->dim = dim;
|
|
gm_param->limit_FOX_number = limit_FOX_number;
|
|
gm_param->limit_BLOCK_number = limit_BLOCK_number;
|
|
gm_param->limit_CARROT_number = limit_CARROT_number;
|
|
gm_param->limit_EPISODES_number = limit_EPISODES_number;
|
|
gm_param->limit_MOVE_number = limit_MOVE_number;
|
|
|
|
gm_param->add_move = malloc(ACTION_COUNT * sizeof(long int));
|
|
/*
|
|
gm_param->add_move[DOWN] = signedLineFromCoord((long int []){0,-1}, dim);
|
|
gm_param->add_move[LEFT] = signedLineFromCoord((long int []){-1,0}, dim);
|
|
gm_param->add_move[RIGHT] = signedLineFromCoord((long int []){1,0}, dim);
|
|
gm_param->add_move[UP] = signedLineFromCoord((long int []){0,1}, dim);
|
|
*/
|
|
gm_param->add_move[DOWN] = dim->perm[0];
|
|
gm_param->add_move[LEFT] = -1;
|
|
gm_param->add_move[RIGHT] = 1;
|
|
gm_param->add_move[UP] = -1*dim->perm[0];
|
|
return gm_param;
|
|
}
|
|
|
|
struct cell * create_game_cells(dimension *dim){
|
|
struct cell * cells = malloc((dim)->rank * sizeof(struct cell));
|
|
for(long int i=0; i<(dim)->rank; ++i)
|
|
(cells[i]).Q = malloc(ACTION_COUNT * sizeof(double));
|
|
|
|
return cells;
|
|
}
|
|
|
|
struct delay_params * create_delay_params (
|
|
size_t delay_episodes,
|
|
size_t delay_game
|
|
){
|
|
struct delay_params *delay = malloc(sizeof(struct delay_params));
|
|
delay->delay_between_episodes = delay_episodes;
|
|
delay->delay_between_games = delay_game;
|
|
return delay;
|
|
}
|
|
|
|
void reset_game_status(struct game_status * status){
|
|
status->rabbitOldRankPosition = status->startRankPosition ;
|
|
status->rabbitRankPosition = status->startRankPosition ;
|
|
status->endGame = false;
|
|
status->count_MOVES = 0;
|
|
|
|
status->final_reward = 0;
|
|
|
|
}
|
|
|
|
struct game_status * create_game_status(){
|
|
struct game_status *gm_stat = malloc(sizeof(struct game_status));
|
|
reset_game_status(gm_stat);
|
|
return gm_stat;
|
|
}
|
|
|
|
#define GENERATE_RANDOM_PLACES(CONTENT) \
|
|
int CONTENT##_Number = xrand() % (params->limit_##CONTENT##_number) + 1;\
|
|
for(int i=0; i< CONTENT##_Number; ++i) { \
|
|
do{\
|
|
random = xrand() % (dim->rank);\
|
|
}while((gm->cells[random]).content != EMPTY);\
|
|
(gm->cells[random]).content = CONTENT;\
|
|
}
|
|
|
|
long int generate_game(struct game *gm){
|
|
struct game_params *params = gm->params;
|
|
struct game_status *status = gm->status;
|
|
dimension *dim = params->dim;
|
|
for(long int i=0; i<(params->dim)->rank; ++i){
|
|
(gm->cells[i]).rankPosition = i;
|
|
(gm->cells[i]).content = EMPTY; //0
|
|
for(long int j=0; j < ACTION_COUNT; ++j)
|
|
(gm->cells[i]).Q[j] = 0;
|
|
}
|
|
//srand(time(NULL));
|
|
int random;
|
|
GENERATE_RANDOM_PLACES(FOX)
|
|
GENERATE_RANDOM_PLACES(CARROT)
|
|
GENERATE_RANDOM_PLACES(BLOCK)
|
|
do{
|
|
random = xrand() % (dim->rank);
|
|
}while((gm->cells[random]).content != EMPTY);
|
|
(gm->cells[random]).content = START;
|
|
|
|
(status)->startRankPosition = random;
|
|
(status)->rabbitOldRankPosition = random;
|
|
(status)->rabbitRankPosition = random;
|
|
|
|
return (status->startRankPosition);
|
|
}
|
|
|
|
struct game * create_game(
|
|
struct game_params *params,
|
|
struct qlearning_params *qlearnParams,
|
|
struct delay_params * delay
|
|
){
|
|
struct game *gm = malloc(sizeof(struct game));
|
|
gm->params = params;
|
|
gm->qlearnParams = qlearnParams;
|
|
gm->cells = create_game_cells(params->dim);
|
|
gm->delay = delay; // create_delay_params(10000000, 2000000);
|
|
gm->status = create_game_status();
|
|
|
|
|
|
generate_game(gm);
|
|
|
|
return gm;
|
|
}
|
|
|
|
|
|
//void free_qlearning_params(struct qlearning_params * qlearn_params);
|
|
//void free_game_params(struct game_params *game_prms);
|
|
//void free_cells(struct cell * cells);
|
|
//void free_game_status(struct game_status *status);
|
|
void free_game(struct game *gm){
|
|
for(long int i = 0; i < ((gm->params)->dim)->rank; ++i){
|
|
free((gm->cells[i]).Q);
|
|
}
|
|
free(gm->cells);
|
|
free_dimension((gm->params)->dim);
|
|
free((gm->params)->add_move);
|
|
free(gm->params);
|
|
free(gm->qlearnParams);
|
|
free(gm->delay);
|
|
free(gm->status);
|
|
free(gm);
|
|
}
|
|
|
|
long int function_add_move(struct game *gm, enum Action action){
|
|
struct game_status * status = gm->status;
|
|
dimension *dim = (gm->params)->dim;
|
|
|
|
if(action == UP && (status->rabbitRankPosition / dim->perm[0]) == 0){
|
|
return dim->rank * -1;
|
|
}
|
|
if(action == DOWN && (status->rabbitRankPosition / dim->perm[0]) == dim->perm[1]-1){
|
|
return dim->rank ;
|
|
}
|
|
if(action == LEFT && (status->rabbitRankPosition % dim->perm[0]) == 0){
|
|
return dim->rank * -1;
|
|
}
|
|
if(action == RIGHT && (status->rabbitRankPosition % dim->perm[0]) == dim->perm[1]-1){
|
|
return dim->rank;
|
|
}
|
|
|
|
return (gm->params)->add_move[action];
|
|
|
|
}
|
|
|
|
void move_game(struct game *gm, enum Action action){
|
|
struct game_params *params = gm->params;
|
|
struct game_status *status = gm->status;
|
|
struct cell * cells = gm->cells;
|
|
|
|
++(status->count_MOVES);
|
|
long int newRankPosition = status->rabbitRankPosition + function_add_move(gm, action);
|
|
|
|
if (newRankPosition < 0 || newRankPosition >= (params->dim)->rank) {
|
|
status->reward = REWARD_OUT;
|
|
}
|
|
else if (status->count_MOVES > params->limit_MOVE_number) {
|
|
status->rabbitRankPosition = newRankPosition;
|
|
status->endGame = true;
|
|
status->reward = REWARD_MOVES_OUT;
|
|
}
|
|
else if (cells[newRankPosition].content == BLOCK) {
|
|
status->reward = REWARD_BLOCK;
|
|
}
|
|
else if (cells[newRankPosition].content == FOX){
|
|
status->rabbitRankPosition = newRankPosition;
|
|
status->endGame = true;
|
|
status->reward = REWARD_FOX;
|
|
}
|
|
else if (cells[newRankPosition].content == CARROT) {
|
|
status->rabbitRankPosition = newRankPosition;
|
|
status->endGame = true;
|
|
status->reward = REWARD_CARROT;
|
|
}
|
|
else {
|
|
status->rabbitRankPosition = newRankPosition;
|
|
status->reward = REWARD_EMPTY;
|
|
}
|
|
}
|
|
|
|
|
|
void printLine(char c, int l, bool prec) {
|
|
//printf("\033[00m");
|
|
printf("\033[%d;0%dm", 0, 1); // noir // vide
|
|
if (prec) printf("%*c\n", 10, ' ');
|
|
for (int i = 0; i < l; i++) printf("%c", c);
|
|
printf("%*c\n", 10, ' ');
|
|
}
|
|
void printLinec(char c, int l) {
|
|
printLine(c, l, true);
|
|
}
|
|
|
|
void code2dCoul(dimension *color, enum Content content) {
|
|
if( color && color->size > 1 ){
|
|
if (content) {
|
|
color->perm[0] = content + 31; color->perm[1] = 1;
|
|
}
|
|
else {
|
|
color->perm[0] = 0; color->perm[1] = 1;
|
|
}
|
|
}
|
|
}
|
|
|
|
void gotoxy(int x, int y)
|
|
{
|
|
printf("%c[%d;%df", 0x1B, y, x);
|
|
}
|
|
|
|
void print_game_dim2(struct game *gm) {
|
|
struct cell *cells = gm->cells;
|
|
dimension *dim = ((gm->params)->dim);
|
|
|
|
if(dim->size == 2){
|
|
dimension *coul = create_dim(2);
|
|
long int *coord = malloc(2*sizeof(long int));
|
|
int mult = dim->perm[0] * 14;
|
|
char sep = '-';
|
|
gotoxy(0, 10);
|
|
for (long int j = 0; j < dim->perm[1]; j++) {
|
|
for (long int i = 0; i < dim->perm[0]; i++) {
|
|
long int cur = i + j * dim->perm[0];
|
|
code2dCoul(coul, cells[cur].content);
|
|
printf("\033[%ld;0%ldm", coul->perm[0], coul->perm[0]);
|
|
//printf("s:%2d,(%d,%d),%2c|", cells[cur].rankPosition, cells[cur].pos.x, cells[cur].pos.y, cont_name[cells[cur].c]);
|
|
printf("s:%2ld,(%ld,%ld),%2c|", cur, i , j , content_name[cells[cur].content]);
|
|
}
|
|
printLinec(sep, mult);
|
|
for (long int k = 0; k < ACTION_COUNT; k++) {
|
|
for (long int i = 0; i < dim->perm[0]; i++) {
|
|
long int cur = i + j * dim->perm[0];
|
|
code2dCoul(coul, cells[cur].content);
|
|
printf("\033[%ld;%ldm", coul->perm[0], coul->perm[1]);
|
|
if ( (gm->status)->rabbitRankPosition == cur ){ /*lapin_pos.x == i && lapin_pos.y == j */
|
|
//printf("\033[47;01m");
|
|
printf("\033[106m");
|
|
}else{
|
|
printf("\033[49m");
|
|
}
|
|
printf("%2c: %8.4f |", action_name[k], cells[cur].Q[k]);
|
|
}
|
|
printf("%*c\n", 10, ' ');
|
|
}
|
|
printLine(sep, mult, false);
|
|
}
|
|
free_dimension(coul);
|
|
free(coord);
|
|
}
|
|
}
|
|
|
|
|
|
void mainQlearning_game(struct game *gm){
|
|
struct main_list_TYPE_L_INT * list_final_rewards = create_var_list_TYPE_L_INT();
|
|
|
|
struct game_params *params = gm->params;
|
|
struct game_status *status = gm->status;
|
|
struct qlearning_params * qlearnParams = gm->qlearnParams;
|
|
struct cell * cells = gm->cells;
|
|
enum Action action;
|
|
int random;
|
|
long int NUMBER_EPISODE2 = (params->limit_EPISODES_number) * (params->limit_EPISODES_number);
|
|
double proba_explor;
|
|
//srand(time(NULL));
|
|
|
|
for(long int k=0 ; k < params->limit_game_number; ++k){
|
|
generate_game(gm);
|
|
for(long int episode = 0; episode < params->limit_EPISODES_number; ++episode){
|
|
reset_game_status(status);
|
|
|
|
while(!(status->endGame)){
|
|
random = xrand() % NUMBER_EPISODE2;
|
|
proba_explor = (double)random / NUMBER_EPISODE2;
|
|
if( proba_explor < qlearnParams->exploration_factor * (1. / ((episode / 10 ) + 1))){
|
|
action = xrand() % ACTION_COUNT;
|
|
printf("exploration action ");
|
|
}
|
|
else{
|
|
action = ARG_MAX_ARRAY_TYPE_DOUBLE( cells[status->rabbitRankPosition].Q, ACTION_COUNT );
|
|
printf("greedy action ");
|
|
}
|
|
|
|
status->rabbitOldRankPosition = status->rabbitRankPosition;
|
|
move_game(gm, action);
|
|
|
|
printf("ik=%ld, episode = %ld , action = %d rbPos = %ld, rwds = %ld, final_rwrd=%ld\n",k,episode,action, status->rabbitRankPosition,status->reward, status->final_reward);
|
|
// update Q array on the action of the state
|
|
cells[status->rabbitOldRankPosition].Q[action] = cells[status->rabbitOldRankPosition].Q[action] + (qlearnParams->learning_rate * (status->reward + qlearnParams->discount_factor * MAX_ARRAY_TYPE_DOUBLE(cells[status->rabbitRankPosition].Q, ACTION_COUNT) - cells[status->rabbitOldRankPosition].Q[action]));
|
|
//cells[status->rabbitRankPosition].Q[action] = cells[status->rabbitRankPosition].Q[action] + (qlearnParams->learning_rate * (status->reward + qlearnParams->discount_factor * MAX_ARRAY_TYPE_DOUBLE(cells[status->rabbitRankPosition].Q, ACTION_COUNT) - cells[status->rabbitRankPosition].Q[action]));
|
|
//cells[status->rabbitRankPosition].Q[action] = (1 - qlearnParams->learning_rate) * cells[status->rabbitRankPosition].Q[action] + (qlearnParams->learning_rate * (status->reward + qlearnParams->discount_factor * MAX_ARRAY_TYPE_DOUBLE(cells[status->rabbitRankPosition].Q, ACTION_COUNT )));
|
|
|
|
status->final_reward += status->reward;
|
|
|
|
print_game_dim2(gm);
|
|
|
|
usleep((gm->delay)->delay_between_episodes);
|
|
|
|
}
|
|
push_back_list_TYPE_L_INT(list_final_rewards, status->final_reward);
|
|
}
|
|
|
|
FOR_LIST_FORM_BEGIN(TYPE_L_INT, list_final_rewards){
|
|
printf(" %ld ",(list_final_rewards->current_list)->value);
|
|
}
|
|
remove_all_list_in_TYPE_L_INT(list_final_rewards);
|
|
usleep((gm->delay)->delay_between_games);
|
|
}
|
|
free(list_final_rewards);
|
|
}
|