diff --git a/trainer/include/TrainInfos.hpp b/trainer/include/TrainInfos.hpp
index 830d0ef6f84b505c3748818978b7540df85e669c..fb0de9727180bbf47ff2e76159479ef2f4b1a71e 100644
--- a/trainer/include/TrainInfos.hpp
+++ b/trainer/include/TrainInfos.hpp
@@ -39,6 +39,10 @@ class TrainInfos
public :
+ std::map<std::string, bool> lastActionWasPredicted;
+
+ public :
+
TrainInfos();
void addTrainLoss(const std::string & classifier, float loss);
void addDevLoss(const std::string & classifier, float loss);
diff --git a/trainer/src/Trainer.cpp b/trainer/src/Trainer.cpp
index 8890b80fb512cd93620c64a3a93192e1895223d4..2bf245e0264affbcab249c036edf56c238720fa5 100644
--- a/trainer/src/Trainer.cpp
+++ b/trainer/src/Trainer.cpp
@@ -304,19 +304,28 @@ void Trainer::doStepTrain()
if (TI.getEpoch() >= k && choiceWithProbability(ProgramParameters::dynamicProbability))
{
actionName = pAction;
+ TI.lastActionWasPredicted[tm.getCurrentClassifier()->name] = true;
}
else
{
if (pActionIsZeroCost)
+ {
actionName = pAction;
+ TI.lastActionWasPredicted[tm.getCurrentClassifier()->name] = true;
+ }
else
+ {
actionName = oAction;
+ TI.lastActionWasPredicted[tm.getCurrentClassifier()->name] = false;
+ }
+
}
if (ProgramParameters::debug)
{
fprintf(stderr, "Speed : %s actions/s\n", int2humanStr((int)currentSpeed).c_str());
trainConfig.printForDebug(stderr);
+ tm.getCurrentClassifier()->printWeightedActions(stderr, weightedActions, 10);
fprintf(stderr, "pAction=<%s> oAction=<%s> action=<%s>\n", pAction.c_str(), oAction.c_str(), actionName.c_str());
}
@@ -356,7 +365,7 @@ void Trainer::doStepTrain()
}
actionName = pAction;
- if (choiceWithProbability(0.2) && TI.getEpoch() >= ProgramParameters::dynamicEpoch)
+ if (choiceWithProbability(0.6) && TI.getEpoch() >= ProgramParameters::dynamicEpoch)
actionName = oAction;
char buffer[1024];
@@ -370,12 +379,15 @@ void Trainer::doStepTrain()
auto & normalHistory = trainConfig.getActionsHistory(normalClassifierName);
// If a BACK just happened
- if (normalHistory.size() > 1 && trainConfig.getCurrentStateHistory().size() > 0 && trainConfig.getCurrentStateHistory().top() != "EPSILON" && TI.getEpoch() >= ProgramParameters::dynamicEpoch)
+ if (normalHistory.size() > 1 && trainConfig.getCurrentStateHistory().size() > 0 && split(trainConfig.getCurrentStateHistory().top(), ' ')[0] == "BACK" && TI.getEpoch() >= ProgramParameters::dynamicEpoch)
{
- auto & lastAction = normalHistory[normalHistory.size()-2];
+ fprintf(stderr, "Current classifier : <%s>\n", tm.getCurrentClassifier()->name.c_str());
+ fprintf(stderr, "Current state history : <%s>\n", trainConfig.getCurrentStateHistory().top().c_str());
+
+ auto & lastAction = trainConfig.lastUndoneAction;
auto & newAction = normalHistory[normalHistory.size()-1];
auto & lastActionName = lastAction.first;
- auto & newActionName = lastAction.first;
+ auto & newActionName = newAction.first;
auto lastCost = lastAction.second;
auto newCost = newAction.second;
@@ -384,16 +396,24 @@ void Trainer::doStepTrain()
fprintf(stderr, "<%s>(%d) -> <%s>(%d)\n", lastActionName.c_str(), lastCost, newActionName.c_str(), newCost);
}
- if (newCost >= lastCost)
+ if (TI.lastActionWasPredicted[normalClassifierName])
{
- loss = tm.getCurrentClassifier()->trainOnExample(pendingFD[tm.getCurrentClassifier()->name], tm.getCurrentClassifier()->getActionIndex("EPSILON"));
- }
- else
- {
- loss = tm.getCurrentClassifier()->trainOnExample(pendingFD[tm.getCurrentClassifier()->name], tm.getCurrentClassifier()->getActionIndex(trainConfig.getCurrentStateHistory().top()));
+ if (ProgramParameters::debug)
+ {
+ fprintf(stderr, "Updating neural network \'%s\'\n", tm.getCurrentClassifier()->name.c_str());
+ }
+ if (newCost >= lastCost)
+ {
+ loss = tm.getCurrentClassifier()->trainOnExample(pendingFD[tm.getCurrentClassifier()->name], tm.getCurrentClassifier()->getActionIndex("EPSILON"));
+ }
+ else
+ {
+ loss = tm.getCurrentClassifier()->trainOnExample(pendingFD[tm.getCurrentClassifier()->name], tm.getCurrentClassifier()->getActionIndex(trainConfig.getCurrentStateHistory().top()));
+ }
+
+ TI.addTrainLoss(tm.getCurrentClassifier()->name, loss);
}
- TI.addTrainLoss(tm.getCurrentClassifier()->name, loss);
}
if (ProgramParameters::debug)
@@ -414,7 +434,7 @@ void Trainer::doStepTrain()
TransitionMachine::Transition * transition = tm.getTransition(actionName);
action->setInfos(transition->headMvt, tm.getCurrentState());
- trainConfig.addToActionsHistory(tm.getCurrentClassifier()->name, actionName, tm.getCurrentClassifier()->getActionCost(trainConfig, actionName));
+ trainConfig.addToActionsHistory(trainConfig.getCurrentStateName(), actionName, tm.getCurrentClassifier()->getActionCost(trainConfig, actionName));
action->apply(trainConfig);
tm.takeTransition(transition);
diff --git a/transition_machine/include/Config.hpp b/transition_machine/include/Config.hpp
index 0d43fe15a2032f3987aa4df269bba3737c7b0d4f..5b99d95b71e25b75addbd7051862d1d6feb4d30e 100644
--- a/transition_machine/include/Config.hpp
+++ b/transition_machine/include/Config.hpp
@@ -176,6 +176,8 @@ class Config
LimitedStack<std::size_t> hashHistory;
/// @brief The sequence of Actions that made that Config.
LimitedStack< std::pair<std::string, Action> > pastActions;
+ /// @brief The last action that have been undone.
+ std::pair<std::string, int> lastUndoneAction;
public :
@@ -346,8 +348,8 @@ class Config
///
/// @param index Index of the column to print.
void printColumnInfos(unsigned int index);
- void addToActionsHistory(std::string & classifier, std::string & action, int cost);
- std::vector< std::pair<std::string, int> > & getActionsHistory(std::string & classifier);
+ void addToActionsHistory(std::string & state, std::string & action, int cost);
+ std::vector< std::pair<std::string, int> > & getActionsHistory(std::string & state);
};
#endif
diff --git a/transition_machine/src/Action.cpp b/transition_machine/src/Action.cpp
index b67e64d1cfbafd0bf3192b5ef22304d7c7915195..0dab75d52bfe917d350d28b23c33c342ff0b996a 100644
--- a/transition_machine/src/Action.cpp
+++ b/transition_machine/src/Action.cpp
@@ -41,7 +41,31 @@ void Action::undo(Config & config)
if (ProgramParameters::debug)
fprintf(stderr, "Undoing action <%s><%s>, state history size = %d past actions size = %d...", stateName.c_str(), name.c_str(), config.getStateHistory(stateName).size(), config.pastActions.size());
- config.getStateHistory(stateName).pop();
+ if (true)
+ {
+ std::string undoneName = config.getStateHistory(stateName).pop();
+
+ auto & history = config.getActionsHistory(stateName);
+ if (history.size() == 0)
+ {
+ fprintf(stderr, "ERROR (%s) : actionsHistory of \'%s\' is empty (undoneName=\'%s\'). Aborting.\n", ERRINFO, stateName.c_str(), undoneName.c_str());
+ exit(1);
+ }
+
+ for (int i = (int)history.size()-1; i >= 0; i--)
+ {
+ if (history[i].first == undoneName)
+ {
+ config.lastUndoneAction = history[i];
+ break;
+ }
+ else if (i == 0)
+ {
+ fprintf(stderr, "ERROR (%s) : could not find action \'%s\' in actionsHistory of state \'%s\'. Aborting.\n", ERRINFO, undoneName.c_str(), stateName.c_str());
+ exit(1);
+ }
+ }
+ }
if (ProgramParameters::debug)
fprintf(stderr, "done\n");
diff --git a/transition_machine/src/Config.cpp b/transition_machine/src/Config.cpp
index 54815fa4ed137bfc5f58d2715f2988739af50eef..a82342d71f9f5931e8e7aab5841bacddbf40309d 100644
--- a/transition_machine/src/Config.cpp
+++ b/transition_machine/src/Config.cpp
@@ -615,14 +615,14 @@ void Config::printColumnInfos(unsigned int index)
fprintf(stderr, "\n");
}
-void Config::addToActionsHistory(std::string & classifier, std::string & action, int cost)
+void Config::addToActionsHistory(std::string & state, std::string & action, int cost)
{
- actionsHistory[classifier+"_"+std::to_string(head)].emplace_back(action, cost);
+ actionsHistory[state+"_"+std::to_string(head)].emplace_back(action, cost);
}
-std::vector< std::pair<std::string, int> > & Config::getActionsHistory(std::string & classifier)
+std::vector< std::pair<std::string, int> > & Config::getActionsHistory(std::string & state)
{
- return actionsHistory[classifier+"_"+std::to_string(head)];
+ return actionsHistory[state+"_"+std::to_string(head)];
}
float Config::Tape::getScore()