sp
/
tempest
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
Browse Source

added proper action reward handling to JANI next-state generator 

Former-commit-id: cd554d6e12 [formerly 47dfb5a796]
Former-commit-id: 67a31637c5
tempestpy_adaptions
dehnert 8 years ago
parent
99badd02c5
commit
ce5ca9d1ce
6 changed files with 72 additions and 34 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 4
      src/builder/ExplicitModelBuilder.cpp
  2. 35
      src/generator/Choice.cpp
  3. 19
      src/generator/Choice.h
  4. 33
      src/generator/JaniNextStateGenerator.cpp
  5. 3
      src/generator/JaniNextStateGenerator.h
  6. 12
      src/generator/PrismNextStateGenerator.cpp

4
src/builder/ExplicitModelBuilder.cpp
View File

@ -274,7 +274,7 @@ namespace storm {
for (auto const& choice : behavior) {
// Add command labels if requested.
if (generator->getOptions().isBuildChoiceLabelsSet()) {
choiceLabels.get().push_back(choice.getChoiceLabels());
choiceLabels.get().push_back(choice.getLabels());
}
// If we keep track of the Markovian choices, store whether the current one is Markovian.
@ -289,7 +289,7 @@ namespace storm {
}
// Add the rewards to the reward models.
auto choiceRewardIt = choice.getChoiceRewards().begin();
auto choiceRewardIt = choice.getRewards().begin();
for (auto& rewardModelBuilder : rewardModelBuilders) {
if (rewardModelBuilder.hasStateActionRewards()) {
rewardModelBuilder.addStateActionReward(*choiceRewardIt);

35
src/generator/Choice.cpp
View File

@ -8,7 +8,7 @@ namespace storm {
namespace generator {
template<typename ValueType, typename StateType>
Choice<ValueType, StateType>::Choice(uint_fast64_t actionIndex, bool markovian) : markovian(markovian), actionIndex(actionIndex), distribution(), totalMass(storm::utility::zero<ValueType>()), choiceRewards() {
Choice<ValueType, StateType>::Choice(uint_fast64_t actionIndex, bool markovian) : markovian(markovian), actionIndex(actionIndex), distribution(), totalMass(storm::utility::zero<ValueType>()), rewards(), labels() {
// Intentionally left empty.
}
@ -33,24 +33,24 @@ namespace storm {
}
template<typename ValueType, typename StateType>
void Choice<ValueType, StateType>::addChoiceLabel(uint_fast64_t label) {
if (!choiceLabels) {
choiceLabels = LabelSet();
void Choice<ValueType, StateType>::addLabel(uint_fast64_t label) {
if (!labels) {
labels = LabelSet();
}
choiceLabels->insert(label);
labels->insert(label);
}
template<typename ValueType, typename StateType>
void Choice<ValueType, StateType>::addChoiceLabels(LabelSet const& labelSet) {
if (!choiceLabels) {
choiceLabels = LabelSet();
void Choice<ValueType, StateType>::addLabels(LabelSet const& labelSet) {
if (!labels) {
labels = LabelSet();
}
choiceLabels->insert(labelSet.begin(), labelSet.end());
labels->insert(labelSet.begin(), labelSet.end());
}
template<typename ValueType, typename StateType>
boost::container::flat_set<uint_fast64_t> const& Choice<ValueType, StateType>::getChoiceLabels() const {
return *choiceLabels;
boost::container::flat_set<uint_fast64_t> const& Choice<ValueType, StateType>::getLabels() const {
return *labels;
}
template<typename ValueType, typename StateType>
@ -70,13 +70,18 @@ namespace storm {
}
template<typename ValueType, typename StateType>
void Choice<ValueType, StateType>::addChoiceReward(ValueType const& value) {
choiceRewards.push_back(value);
void Choice<ValueType, StateType>::addReward(ValueType const& value) {
rewards.push_back(value);
}
template<typename ValueType, typename StateType>
std::vector<ValueType> const& Choice<ValueType, StateType>::getChoiceRewards() const {
return choiceRewards;
void Choice<ValueType, StateType>::addRewards(std::vector<ValueType>&& values) {
this->rewards = std::move(values);
}
template<typename ValueType, typename StateType>
std::vector<ValueType> const& Choice<ValueType, StateType>::getRewards() const {
return rewards;
}
template<typename ValueType, typename StateType>

19
src/generator/Choice.h
View File

@ -66,21 +66,21 @@ namespace storm {
*
* @param label The label to associate with this choice.
*/
void addChoiceLabel(uint_fast64_t label);
void addLabel(uint_fast64_t label);
/*!
* Adds the given label set to the labels associated with this choice.
*
* @param labelSet The label set to associate with this choice.
*/
void addChoiceLabels(LabelSet const& labelSet);
void addLabels(LabelSet const& labelSet);
/*!
* Retrieves the set of labels associated with this choice.
*
* @return The set of labels associated with this choice.
*/
LabelSet const& getChoiceLabels() const;
LabelSet const& getLabels() const;
/*!
* Retrieves the index of the action of this choice.
@ -104,12 +104,17 @@ namespace storm {
/*!
* Adds the given value to the reward associated with this choice.
*/
void addChoiceReward(ValueType const& value);
void addReward(ValueType const& value);
/*!
* Adds the given choices rewards to this choice.
*/
void addRewards(std::vector<ValueType>&& values);
/*!
* Retrieves the rewards for this choice under selected reward models.
*/
std::vector<ValueType> const& getChoiceRewards() const;
std::vector<ValueType> const& getRewards() const;
/*!
* Retrieves whether the choice is Markovian.
@ -135,10 +140,10 @@ namespace storm {
ValueType totalMass;
// The reward values associated with this choice.
std::vector<ValueType> choiceRewards;
std::vector<ValueType> rewards;
// The labels that are associated with this choice.
boost::optional<LabelSet> choiceLabels;
boost::optional<LabelSet> labels;
};
template<typename ValueType, typename StateType>

33
src/generator/JaniNextStateGenerator.cpp
View File

@ -22,7 +22,7 @@ namespace storm {
}
template<typename ValueType, typename StateType>
JaniNextStateGenerator<ValueType, StateType>::JaniNextStateGenerator(storm::jani::Model const& model, NextStateGeneratorOptions const& options, bool flag) : NextStateGenerator<ValueType, StateType>(model.getExpressionManager(), VariableInformation(model), options), model(model), rewardVariables() {
JaniNextStateGenerator<ValueType, StateType>::JaniNextStateGenerator(storm::jani::Model const& model, NextStateGeneratorOptions const& options, bool flag) : NextStateGenerator<ValueType, StateType>(model.getExpressionManager(), VariableInformation(model), options), model(model), rewardVariables(), hasStateActionRewards(false) {
STORM_LOG_THROW(model.hasDefaultComposition(), storm::exceptions::WrongFormatException, "The explicit next-state generator currently does not support custom system compositions.");
STORM_LOG_THROW(!model.hasNonGlobalTransientVariable(), storm::exceptions::InvalidSettingsException, "The explicit next-state generator currently does not support automata-local transient variables.");
STORM_LOG_THROW(!this->options.isBuildChoiceLabelsSet(), storm::exceptions::InvalidSettingsException, "JANI next-state generator cannot generate choice labels.");
@ -282,6 +282,10 @@ namespace storm {
if (this->isDeterministicModel() && totalNumberOfChoices > 1) {
Choice<ValueType> globalChoice;
// For CTMCs, we need to keep track of the total exit rate to scale the action rewards later. For DTMCs
// this is equal to the number of choices, which is why we initialize it like this here.
ValueType totalExitRate = this->isDiscreteTimeModel() ? static_cast<ValueType>(totalNumberOfChoices) : storm::utility::zero<ValueType>();
// Iterate over all choices and combine the probabilities/rates into one choice.
for (auto const& choice : allChoices) {
for (auto const& stateProbabilityPair : choice) {
@ -292,11 +296,23 @@ namespace storm {
}
}
if (hasStateActionRewards && !this->isDiscreteTimeModel()) {
totalExitRate += choice.getTotalMass();
}
if (this->options.isBuildChoiceLabelsSet()) {
globalChoice.addChoiceLabels(choice.getChoiceLabels());
globalChoice.addLabels(choice.getLabels());
}
}
std::vector<ValueType> stateActionRewards(rewardVariables.size(), storm::utility::zero<ValueType>());
for (auto const& choice : allChoices) {
for (uint_fast64_t rewardVariableIndex = 0; rewardVariableIndex < rewardVariables.size(); ++rewardVariableIndex) {
stateActionRewards[rewardVariableIndex] += choice.getRewards()[rewardVariableIndex] * choice.getTotalMass() / totalExitRate;
}
}
globalChoice.addRewards(std::move(stateActionRewards));
// Move the newly fused choice in place.
allChoices.clear();
allChoices.push_back(std::move(globalChoice));
@ -349,7 +365,7 @@ namespace storm {
}
// Create the state-action reward for the newly created choice.
performTransientAssignments(edge.getAssignments().getTransientAssignments(), [&choice] (ValueType const& value) { choice.addChoiceReward(value); } );
performTransientAssignments(edge.getAssignments().getTransientAssignments(), [&choice] (ValueType const& value) { choice.addReward(value); } );
// Check that the resulting distribution is in fact a distribution.
STORM_LOG_THROW(!this->isDiscreteTimeModel() || this->comparator.isOne(probabilitySum), storm::exceptions::WrongFormatException, "Probabilities do not sum to one for edge (actually sum to " << probabilitySum << ").");
@ -385,6 +401,7 @@ namespace storm {
while (!done) {
boost::container::flat_map<CompressedState, ValueType>* currentTargetStates = new boost::container::flat_map<CompressedState, ValueType>();
boost::container::flat_map<CompressedState, ValueType>* newTargetStates = new boost::container::flat_map<CompressedState, ValueType>();
std::vector<ValueType> stateActionRewards(rewardVariables.size(), storm::utility::zero<ValueType>());
currentTargetStates->emplace(state, storm::utility::one<ValueType>());
@ -405,6 +422,10 @@ namespace storm {
newTargetStates->emplace(newTargetState, stateProbabilityPair.second * this->evaluator.asRational(destination.getProbability()));
}
}
// Create the state-action reward for the newly created choice.
auto valueIt = stateActionRewards.begin();
performTransientAssignments(edge.getAssignments().getTransientAssignments(), [&valueIt] (ValueType const& value) { *valueIt += value; ++valueIt; } );
}
// If there is one more command to come, shift the target states one time step back.
@ -423,6 +444,9 @@ namespace storm {
// Now create the actual distribution.
Choice<ValueType>& choice = result.back();
// Add the rewards to the choice.
choice.addRewards(std::move(stateActionRewards));
// Add the probabilities/rates to the newly created choice.
ValueType probabilitySum = storm::utility::zero<ValueType>();
for (auto const& stateProbabilityPair : *newTargetStates) {
@ -601,6 +625,7 @@ namespace storm {
}
if (*rewardVariableIt == assignment.getExpressionVariable()) {
rewardModelInformation[std::distance(rewardVariables.begin(), rewardVariableIt)].setHasStateActionRewards();
hasStateActionRewards = true;
++rewardVariableIt;
}
}

3
src/generator/JaniNextStateGenerator.h
View File

@ -108,6 +108,9 @@ namespace storm {
/// A vector storing information about the corresponding reward models (variables).
std::vector<RewardModelInformation> rewardModelInformation;
// A flag that stores whether at least one of the selected reward models has state-action rewards.
bool hasStateActionRewards;
};
}

12
src/generator/PrismNextStateGenerator.cpp
View File

@ -242,7 +242,7 @@ namespace storm {
}
if (this->options.isBuildChoiceLabelsSet()) {
globalChoice.addChoiceLabels(choice.getChoiceLabels());
globalChoice.addLabels(choice.getLabels());
}
}
@ -259,7 +259,7 @@ namespace storm {
}
}
globalChoice.addChoiceReward(stateActionRewardValue);
globalChoice.addReward(stateActionRewardValue);
}
// Move the newly fused choice in place.
@ -382,7 +382,7 @@ namespace storm {
// Remember the command labels only if we were asked to.
if (this->options.isBuildChoiceLabelsSet()) {
choice.addChoiceLabel(command.getGlobalIndex());
choice.addLabel(command.getGlobalIndex());
}
// Iterate over all updates of the current command.
@ -410,7 +410,7 @@ namespace storm {
}
}
}
choice.addChoiceReward(stateActionRewardValue);
choice.addReward(stateActionRewardValue);
}
// Check that the resulting distribution is in fact a distribution.
@ -486,7 +486,7 @@ namespace storm {
if (this->options.isBuildChoiceLabelsSet()) {
// Add the labels of all commands to this choice.
for (uint_fast64_t i = 0; i < iteratorList.size(); ++i) {
choice.addChoiceLabel(iteratorList[i]->get().getGlobalIndex());
choice.addLabel(iteratorList[i]->get().getGlobalIndex());
}
}
@ -511,7 +511,7 @@ namespace storm {
}
}
}
choice.addChoiceReward(stateActionRewardValue);
choice.addReward(stateActionRewardValue);
}
// Dispose of the temporary maps.

Write Preview
Loading…
Cancel
Save