fanasina/y_project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

debug : nan, it's the learning rate too high,s try to find optimum: 0.001

Browse Source
This commit is contained in:
fanasina
2025-12-13 05:00:23 +01:00
parent 32207170f6
commit 3d54639d4c
18 changed files with 834 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files
deepQlearn_0/src/deepQlearning/learn_to_drive.c
+74 -27
View File
@@ -2,17 +2,26 @@
char *action_name[8] = {"LEFT", "CENTER", "RIGHT"};
#define THRESHOLD_UP 10
#define UPPER_THRESHOLD 10
#define DIVIDER__ 1
#define USE_THRESHOLD 0
float reLU(float x){
if(x>THRESHOLD_UP) return THRESHOLD_UP;
if(x>0) return x;
if(x!=x){// nan
printf("nan relu ");
}
#if USE_THRESHOLD
if(x>UPPER_THRESHOLD) return UPPER_THRESHOLD;
#endif
if(x>0) return x/DIVIDER__;
return 0;
}
float d_reLU(float x){
if (x>THRESHOLD_UP) return 0;
if (x>0) return 1;
#if USE_THRESHOLD
if (x>UPPER_THRESHOLD) return 0;
#endif
if (x>0) return 1/DIVIDER__;
return 0;
}
@@ -37,7 +46,16 @@ float id(float x){ return x;}
float constOne(float x){return 1;}
struct networks_qlearning * create_nework_qlearning(
void tensorProdTHR_TYPE_FLOAT(tensor_TYPE_FLOAT **MM, tensor_TYPE_FLOAT *M0, tensor_TYPE_FLOAT *M1, size_t nbthread){
return tensorProd_TYPE_FLOAT(MM,M0,M1);
}
void tensorContractnProdTHR_TYPE_FLOAT(tensor_TYPE_FLOAT **MM, tensor_TYPE_FLOAT *M0, tensor_TYPE_FLOAT *M1, size_t contractionNumber, size_t nbthread) {
return tensorContractnProd_TYPE_FLOAT(MM,M0,M1,contractionNumber);
}
struct networks_qlearning * create_network_qlearning(
struct config_layers * config,
bool randomize, float minR, float maxR, int randomRange,
size_t nb_prod_thread,
@@ -55,10 +73,13 @@ struct networks_qlearning * create_nework_qlearning(
copy_weight_in_networks_from_main_to_best(qnets);
setup_all_layers_functions_TYPE_FLOAT(qnets->main_net, tensorContractnProdThread_TYPE_FLOAT, tensorProdThread_TYPE_FLOAT, D_L2, L2, reLU, d_reLU);
//setup_all_layers_functions_TYPE_FLOAT(qnets->main_net, tensorContractnProdTHR_TYPE_FLOAT, tensorProdTHR_TYPE_FLOAT, D_L2, L2, reLU, d_reLU);
setup_all_layers_params_TYPE_FLOAT(qnets->main_net, nb_prod_thread, nb_calc_thread, learning_rate);
setup_all_layers_functions_TYPE_FLOAT(qnets->target_net, tensorContractnProdThread_TYPE_FLOAT, tensorProdThread_TYPE_FLOAT, D_L2, L2, reLU, d_reLU);
//setup_all_layers_functions_TYPE_FLOAT(qnets->target_net, tensorContractnProdTHR_TYPE_FLOAT, tensorProdTHR_TYPE_FLOAT, D_L2, L2, reLU, d_reLU);
setup_all_layers_params_TYPE_FLOAT(qnets->target_net, nb_prod_thread, nb_calc_thread, learning_rate);
setup_all_layers_functions_TYPE_FLOAT(qnets->best_net, tensorContractnProdThread_TYPE_FLOAT, tensorProdThread_TYPE_FLOAT, D_L2, L2, reLU, d_reLU);
//setup_all_layers_functions_TYPE_FLOAT(qnets->best_net, tensorContractnProdTHR_TYPE_FLOAT, tensorProdTHR_TYPE_FLOAT, D_L2, L2, reLU, d_reLU);
setup_all_layers_params_TYPE_FLOAT(qnets->best_net, nb_prod_thread, nb_calc_thread, learning_rate);
// ne pas mettre fonction d'activation à la sortie , i.e: fonction identité : f(x) = x:
@@ -67,12 +88,12 @@ struct networks_qlearning * create_nework_qlearning(
neurons_TYPE_FLOAT *tmpBest = qnets->best_net;
while(tmpMain){
if(tmpMain->next_layer == NULL){
tmpMain->f_act = id;
tmpMain->d_f_act = constOne;
tmpTarget->f_act = id;
tmpTarget->d_f_act = constOne;
tmpBest->f_act = id;
tmpBest->d_f_act = constOne;
tmpMain->f_act = id_TYPE_FLOAT;//id;
tmpMain->d_f_act = d_id_TYPE_FLOAT; //constOne;
tmpTarget->f_act = id_TYPE_FLOAT;//id;
tmpTarget->d_f_act = d_id_TYPE_FLOAT;//constOne;
tmpBest->f_act = id_TYPE_FLOAT;//id;
tmpBest->d_f_act = d_id_TYPE_FLOAT;// constOne;
}
tmpMain = tmpMain->next_layer;
tmpTarget= tmpTarget->next_layer;
@@ -156,18 +177,18 @@ struct qlearning_params * create_qlearning_params (
){
struct qlearning_params * qparams = malloc(sizeof(struct qlearning_params));
qparams->gamma = gamma;
qparams->learning_rate = learning_rate ;
qparams->discount_factor = discount_factor ;
qparams->gamma = gamma; /* taux d'actualisation (discount rate): default : 0.95 */
qparams->learning_rate = learning_rate ; /* default : 0.001 */
qparams->discount_factor = discount_factor ; /* */
qparams->exploration_factor = exploration_factor ;
qparams->nb_training_before_update_weight_in_target = nb_training_before_update_weight_in_target;
qparams->number_episodes = number_episodes;
qparams->factor_update_learning_rate = 0.995;
qparams->minimum_threshold_learning_rate = 0.0001 ;
qparams->factor_update_exploration_factor = 0.9995 /*0.995*/;
qparams->minimum_threshold_exploration_factor = 0.01;
qparams->minimum_threshold_learning_rate = 0.00001 ;
qparams->factor_update_exploration_factor = 0.995 /*0.995*/;
qparams->minimum_threshold_exploration_factor = 0.0001;
// qparams->threshold_number_same_action = 500;
@@ -250,6 +271,9 @@ void free_RL_agent(struct RL_agent *rlAgent){
#define ACCEPTABLE_REWARD 1000
#define UPDATE_PARAMS 1
#define UPDATE_EXPLOR_FAC 1
void train_qlearning(struct RL_agent * rlAgent,
int action //, long reward
){
@@ -263,8 +287,11 @@ void train_qlearning(struct RL_agent * rlAgent,
calculate_output_by_network_neurons_TYPE_FLOAT(net_target, new_state, &next_action_value);
tensor_TYPE_FLOAT * experimental_values = CREATE_TENSOR_FROM_CPY_DIM_TYPE_FLOAT(action_value->dim);
// print_neurons_msg_TYPE_FLOAT(net_main, " net_main "); getchar();
struct game_status * car_status = rlAgent->car->status;
#if UPDATE_PARAMS
struct qlearning_params * qlParams = rlAgent->qlearnParams;
#endif
copy_tensor_TYPE_FLOAT(experimental_values, action_value) ;
//copy_tensor_TYPE_FLOAT(experimental_values, next_action_value) ;
// experimental_values === Q-tab learning
@@ -277,23 +304,33 @@ void train_qlearning(struct RL_agent * rlAgent,
copy_tensor_TYPE_FLOAT(ttmp->target, experimental_values);
while(ttmp != net_main){
calc_delta_neurons_TYPE_FLOAT(ttmp);
update_weight_neurons_TYPE_FLOAT(ttmp);
//update_weight_neurons_TYPE_FLOAT(ttmp);
ttmp = ttmp->prev_layer;
}
ttmp=net_main->next_layer;
while(ttmp){
update_weight_neurons_TYPE_FLOAT(ttmp);
ttmp = ttmp->next_layer;
}
// ***
if(car_status->cumulative_reward > ACCEPTABLE_REWARD){
#if UPDATE_PARAMS
if((car_status->cumulative_reward > ACCEPTABLE_REWARD) || (rlAgent->status->nb_episodes % 100 == 0) ){
float new_value = ( (net_main->learning_rate < qlParams->minimum_threshold_learning_rate /*0.0001*/) ? net_main->learning_rate :(net_main->learning_rate ) * qlParams->factor_update_learning_rate /*0.995*/ );
UPDATE_ATTRIBUTE_NEURONE_IN_ALL_LAYERS(TYPE_FLOAT, net_main, learning_rate, new_value);
qlParams->learning_rate = new_value;
#if UPDATE_EXPLOR_FAC
qlParams->exploration_factor = (qlParams->exploration_factor < qlParams->minimum_threshold_exploration_factor) ? qlParams->exploration_factor : qlParams->exploration_factor * qlParams->factor_update_exploration_factor ;
#endif
}
#endif
// free_tensor_TYPE_FLOAT(action_value);
// free_tensor_TYPE_FLOAT(next_action_value);
free_tensor_TYPE_FLOAT(experimental_values);
}
#define MAX_SUCCESSIVE_ACTION 200
#define SUCCESSIVE_ACTION_CHECK 1
#define MAX_SUCCESSIVE_ACTION 1000
int select_action(struct RL_agent * rlAgent){
//static size_t explore = 0;
int action;
@@ -313,15 +350,19 @@ int select_action(struct RL_agent * rlAgent){
if(proba_explor > rlAgent->qlearnParams->exploration_factor ){
action = ARG_MAX_ARRAY_TYPE_FLOAT( action_value->x, action_value->dim->rank );
//printf(" STRATEGY : action : %d , factor : %f nb_episodes : %ld \n",action,rlAgent->qlearnParams->exploration_factor, rlAgent->status->nb_episodes);
if(rlAgent->networks->nb_successive_action[action]>MAX_SUCCESSIVE_ACTION){
#if SUCCESSIVE_ACTION_CHECK
if(rlAgent->networks->nb_successive_action[action]>MAX_SUCCESSIVE_ACTION){
rlAgent->networks->nb_successive_action[action]=0;
int recAction=action;
while(action==recAction){
action = xrand() % action_value->dim->rank ;
//printf("debug: action=%d recAction=%d\n",action, recAction);
}
write(1,"#",1);
struct qlearning_params * qlParams = rlAgent->qlearnParams;
write(1,"#",1);
qlParams->exploration_factor = (qlParams->exploration_factor < 1 ) ? qlParams->exploration_factor / qlParams->factor_update_exploration_factor : qlParams->exploration_factor ;
}
#endif
////else write(1,".",1);
//if(action == ARG_MIN_ARRAY_TYPE_FLOAT( action_value->x, action_value->dim->rank ))
//action = xrand() % action_value->dim->rank ;
@@ -333,12 +374,13 @@ int select_action(struct RL_agent * rlAgent){
//printf(" EXPLORE : action : %d , factor : %f nb_episodes : %ld \n",action,rlAgent->qlearnParams->exploration_factor, rlAgent->status->nb_episodes);
////write(1,"*",1);
}
#if SUCCESSIVE_ACTION_CHECK
for(int a=0;a<COUNT_ACTION;++a){
if(a!=action)
rlAgent->networks->nb_successive_action[a]=0;
}
(rlAgent->networks->nb_successive_action[action])++;
#endif
/*
if(rlAgent->status->last_action == action){
++(rlAgent->status->count_last_action);
@@ -450,7 +492,12 @@ void* learn_to_drive(void * lrnarg){
////pthread_create(&threadPrint, NULL, runPrint, (void*)rlAgent);
// while(true){
for(size_t index_episode = 0; (!is_ending(qlStatus)) && (index_episode < qlParams->number_episodes) ; ++index_episode){
for(size_t index_episode = 0;
(!is_ending(qlStatus))
//|| (car_status->cumulative_reward > 2 * ACCEPTABLE_REWARD)
//|| (index_episode < qlParams->number_episodes)
;
++index_episode){
reset(car);
qlStatus->nb_training_after_updated_weight_in_target = 0;
qlStatus->index_episode = index_episode;