418 lines
25 KiB

#include "src/generator/PrismNextStateGenerator.h"
#include <boost/container/flat_map.hpp>
#include "src/utility/constants.h"
#include "src/utility/macros.h"
#include "src/exceptions/WrongFormatException.h"
namespace storm {
namespace generator {
template<typename ValueType, typename StateType>
PrismNextStateGenerator<ValueType, StateType>::PrismNextStateGenerator(storm::prism::Program const& program, VariableInformation const& variableInformation, bool buildChoiceLabeling) : program(program), selectedRewardModels(), buildChoiceLabeling(buildChoiceLabeling), variableInformation(variableInformation), evaluator(program.getManager()), state(nullptr), comparator() {
// Intentionally left empty.
}
template<typename ValueType, typename StateType>
void PrismNextStateGenerator<ValueType, StateType>::addRewardModel(storm::prism::RewardModel const& rewardModel) {
selectedRewardModels.push_back(rewardModel);
hasStateActionRewards |= rewardModel.hasStateActionRewards();
}
template<typename ValueType, typename StateType>
void PrismNextStateGenerator<ValueType, StateType>::setTerminalExpression(storm::expressions::Expression const& terminalExpression) {
this->terminalExpression = terminalExpression;
}
template<typename ValueType, typename StateType>
bool PrismNextStateGenerator<ValueType, StateType>::isDeterministicModel() const {
return program.isDeterministicModel();
}
template<typename ValueType, typename StateType>
std::vector<StateType> PrismNextStateGenerator<ValueType, StateType>::getInitialStates(StateToIdCallback const& stateToIdCallback) {
// FIXME: This only works for models with exactly one initial state. We should make this more general.
CompressedState initialState(variableInformation.getTotalBitOffset());
// We need to initialize the values of the variables to their initial value.
for (auto const& booleanVariable : variableInformation.booleanVariables) {
initialState.set(booleanVariable.bitOffset, booleanVariable.initialValue);
}
for (auto const& integerVariable : variableInformation.integerVariables) {
initialState.setFromInt(integerVariable.bitOffset, integerVariable.bitWidth, static_cast<uint_fast64_t>(integerVariable.initialValue - integerVariable.lowerBound));
}
// Register initial state and return it.
StateType id = stateToIdCallback(initialState);
return {id};
}
template<typename ValueType, typename StateType>
void PrismNextStateGenerator<ValueType, StateType>::load(CompressedState const& state) {
// Since almost all subsequent operations are based on the evaluator, we load the state into it now.
unpackStateIntoEvaluator(state, variableInformation, evaluator);
// Also, we need to store a pointer to the state itself, because we need to be able to access it when expanding it.
this->state = &state;
}
template<typename ValueType, typename StateType>
bool PrismNextStateGenerator<ValueType, StateType>::satisfies(storm::expressions::Expression const& expression) const {
if (expression.isTrue()) {
return true;
}
return evaluator.asBool(expression);
}
template<typename ValueType, typename StateType>
StateBehavior<ValueType, StateType> PrismNextStateGenerator<ValueType, StateType>::expand(StateToIdCallback const& stateToIdCallback) {
// Prepare the result, in case we return early.
StateBehavior<ValueType, StateType> result;
// First, construct the state rewards, as we may return early if there are no choices later and we already
// need the state rewards then.
for (auto const& rewardModel : selectedRewardModels) {
ValueType stateRewardValue = storm::utility::zero<ValueType>();
if (rewardModel.get().hasStateRewards()) {
for (auto const& stateReward : rewardModel.get().getStateRewards()) {
if (evaluator.asBool(stateReward.getStatePredicateExpression())) {
stateRewardValue += ValueType(evaluator.asRational(stateReward.getRewardValueExpression()));
}
}
}
result.addStateReward(stateRewardValue);
}
// If a terminal expression was set and we must not expand this state, return now.
if (terminalExpression && evaluator.asBool(terminalExpression.get())) {
return result;
}
// Get all choices for the state.
std::vector<Choice<ValueType>> allChoices = getUnlabeledChoices(*this->state, stateToIdCallback);
std::vector<Choice<ValueType>> allLabeledChoices = getLabeledChoices(*this->state, stateToIdCallback);
for (auto& choice : allLabeledChoices) {
allChoices.push_back(std::move(choice));
}
std::size_t totalNumberOfChoices = allChoices.size();
// If there is not a single choice, we return immediately, because the state has no behavior (other than
// the state reward).
if (totalNumberOfChoices == 0) {
return result;
}
// If the model is a deterministic model, we need to fuse the choices into one.
if (program.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 = program.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) {
if (program.isDiscreteTimeModel()) {
globalChoice.addProbability(stateProbabilityPair.first, stateProbabilityPair.second / totalNumberOfChoices);
} else {
globalChoice.addProbability(stateProbabilityPair.first, stateProbabilityPair.second);
}
}
if (hasStateActionRewards && !program.isDiscreteTimeModel()) {
totalExitRate += choice.getTotalMass();
}
if (buildChoiceLabeling) {
globalChoice.addChoiceLabels(choice.getChoiceLabels());
}
}
// Now construct the state-action reward for all selected reward models.
for (auto const& rewardModel : selectedRewardModels) {
ValueType stateActionRewardValue = storm::utility::zero<ValueType>();
if (rewardModel.get().hasStateActionRewards()) {
for (auto const& stateActionReward : rewardModel.get().getStateActionRewards()) {
for (auto const& choice : allChoices) {
if (stateActionReward.getActionIndex() == choice.getActionIndex() && evaluator.asBool(stateActionReward.getStatePredicateExpression())) {
stateActionRewardValue += ValueType(evaluator.asRational(stateActionReward.getRewardValueExpression())) * choice.getTotalMass() / totalExitRate;
}
}
}
}
globalChoice.addChoiceReward(stateActionRewardValue);
}
// Move the newly fused choice in place.
allChoices.clear();
allChoices.push_back(std::move(globalChoice));
}
// Move all remaining choices in place.
for (auto& choice : allChoices) {
result.addChoice(std::move(choice));
}
result.setExpanded();
return result;
}
template<typename ValueType, typename StateType>
CompressedState PrismNextStateGenerator<ValueType, StateType>::applyUpdate(CompressedState const& state, storm::prism::Update const& update) {
CompressedState newState(state);
auto assignmentIt = update.getAssignments().begin();
auto assignmentIte = update.getAssignments().end();
// Iterate over all boolean assignments and carry them out.
auto boolIt = variableInformation.booleanVariables.begin();
for (; assignmentIt != assignmentIte && assignmentIt->getExpression().hasBooleanType(); ++assignmentIt) {
while (assignmentIt->getVariable() != boolIt->variable) {
++boolIt;
}
newState.set(boolIt->bitOffset, evaluator.asBool(assignmentIt->getExpression()));
}
// Iterate over all integer assignments and carry them out.
auto integerIt = variableInformation.integerVariables.begin();
for (; assignmentIt != assignmentIte && assignmentIt->getExpression().hasIntegerType(); ++assignmentIt) {
while (assignmentIt->getVariable() != integerIt->variable) {
++integerIt;
}
int_fast64_t assignedValue = evaluator.asInt(assignmentIt->getExpression());
STORM_LOG_THROW(assignedValue <= integerIt->upperBound, storm::exceptions::WrongFormatException, "The update " << update << " leads to an out-of-bounds value (" << assignedValue << ") for the variable '" << assignmentIt->getVariableName() << "'.");
newState.setFromInt(integerIt->bitOffset, integerIt->bitWidth, assignedValue - integerIt->lowerBound);
STORM_LOG_ASSERT(static_cast<int_fast64_t>(newState.getAsInt(integerIt->bitOffset, integerIt->bitWidth)) + integerIt->lowerBound == assignedValue, "Writing to the bit vector bucket failed (read " << newState.getAsInt(integerIt->bitOffset, integerIt->bitWidth) << " but wrote " << assignedValue << ").");
}
// Check that we processed all assignments.
STORM_LOG_ASSERT(assignmentIt == assignmentIte, "Not all assignments were consumed.");
return newState;
}
template<typename ValueType, typename StateType>
boost::optional<std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>>> PrismNextStateGenerator<ValueType, StateType>::getActiveCommandsByActionIndex(uint_fast64_t const& actionIndex) {
boost::optional<std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>>> result((std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>>()));
// Iterate over all modules.
for (uint_fast64_t i = 0; i < program.getNumberOfModules(); ++i) {
storm::prism::Module const& module = program.getModule(i);
// If the module has no command labeled with the given action, we can skip this module.
if (!module.hasActionIndex(actionIndex)) {
continue;
}
std::set<uint_fast64_t> const& commandIndices = module.getCommandIndicesByActionIndex(actionIndex);
// If the module contains the action, but there is no command in the module that is labeled with
// this action, we don't have any feasible command combinations.
if (commandIndices.empty()) {
return boost::optional<std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>>>();
}
std::vector<std::reference_wrapper<storm::prism::Command const>> commands;
// Look up commands by their indices and add them if the guard evaluates to true in the given state.
for (uint_fast64_t commandIndex : commandIndices) {
storm::prism::Command const& command = module.getCommand(commandIndex);
if (evaluator.asBool(command.getGuardExpression())) {
commands.push_back(command);
}
}
// If there was no enabled command although the module has some command with the required action label,
// we must not return anything.
if (commands.size() == 0) {
return boost::optional<std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>>>();
}
result.get().push_back(std::move(commands));
}
return result;
}
template<typename ValueType, typename StateType>
std::vector<Choice<ValueType>> PrismNextStateGenerator<ValueType, StateType>::getUnlabeledChoices(CompressedState const& state, StateToIdCallback stateToIdCallback) {
std::vector<Choice<ValueType>> result;
// Iterate over all modules.
for (uint_fast64_t i = 0; i < program.getNumberOfModules(); ++i) {
storm::prism::Module const& module = program.getModule(i);
// Iterate over all commands.
for (uint_fast64_t j = 0; j < module.getNumberOfCommands(); ++j) {
storm::prism::Command const& command = module.getCommand(j);
// Only consider unlabeled commands.
if (command.isLabeled()) continue;
// Skip the command, if it is not enabled.
if (!evaluator.asBool(command.getGuardExpression())) {
continue;
}
result.push_back(Choice<ValueType>(command.getActionIndex()));
Choice<ValueType>& choice = result.back();
// Remember the command labels only if we were asked to.
if (buildChoiceLabeling) {
choice.addChoiceLabel(command.getGlobalIndex());
}
// Iterate over all updates of the current command.
ValueType probabilitySum = storm::utility::zero<ValueType>();
for (uint_fast64_t k = 0; k < command.getNumberOfUpdates(); ++k) {
storm::prism::Update const& update = command.getUpdate(k);
// Obtain target state index and add it to the list of known states. If it has not yet been
// seen, we also add it to the set of states that have yet to be explored.
StateType stateIndex = stateToIdCallback(applyUpdate(state, update));
// Update the choice by adding the probability/target state to it.
ValueType probability = evaluator.asRational(update.getLikelihoodExpression());
choice.addProbability(stateIndex, probability);
probabilitySum += probability;
}
// Create the state-action reward for the newly created choice.
for (auto const& rewardModel : selectedRewardModels) {
ValueType stateActionRewardValue = storm::utility::zero<ValueType>();
if (rewardModel.get().hasStateActionRewards()) {
for (auto const& stateActionReward : rewardModel.get().getStateActionRewards()) {
if (stateActionReward.getActionIndex() == choice.getActionIndex() && evaluator.asBool(stateActionReward.getStatePredicateExpression())) {
stateActionRewardValue += ValueType(evaluator.asRational(stateActionReward.getRewardValueExpression())) * choice.getTotalMass();
}
}
}
choice.addChoiceReward(stateActionRewardValue);
}
// Check that the resulting distribution is in fact a distribution.
STORM_LOG_THROW(!program.isDiscreteTimeModel() || comparator.isOne(probabilitySum), storm::exceptions::WrongFormatException, "Probabilities do not sum to one for command '" << command << "' (actually sum to " << probabilitySum << ").");
}
}
return result;
}
template<typename ValueType, typename StateType>
std::vector<Choice<ValueType>> PrismNextStateGenerator<ValueType, StateType>::getLabeledChoices(CompressedState const& state, StateToIdCallback stateToIdCallback) {
std::vector<Choice<ValueType>> result;
for (uint_fast64_t actionIndex : program.getSynchronizingActionIndices()) {
boost::optional<std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>>> optionalActiveCommandLists = getActiveCommandsByActionIndex(actionIndex);
// Only process this action label, if there is at least one feasible solution.
if (optionalActiveCommandLists) {
std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>> const& activeCommandList = optionalActiveCommandLists.get();
std::vector<std::vector<std::reference_wrapper<storm::prism::Command const>>::const_iterator> iteratorList(activeCommandList.size());
// Initialize the list of iterators.
for (size_t i = 0; i < activeCommandList.size(); ++i) {
iteratorList[i] = activeCommandList[i].cbegin();
}
// As long as there is one feasible combination of commands, keep on expanding it.
bool done = false;
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>();
currentTargetStates->emplace(state, storm::utility::one<ValueType>());
for (uint_fast64_t i = 0; i < iteratorList.size(); ++i) {
storm::prism::Command const& command = *iteratorList[i];
for (uint_fast64_t j = 0; j < command.getNumberOfUpdates(); ++j) {
storm::prism::Update const& update = command.getUpdate(j);
for (auto const& stateProbabilityPair : *currentTargetStates) {
// Compute the new state under the current update and add it to the set of new target states.
CompressedState newTargetState = applyUpdate(stateProbabilityPair.first, update);
newTargetStates->emplace(newTargetState, stateProbabilityPair.second * evaluator.asRational(update.getLikelihoodExpression()));
}
}
// If there is one more command to come, shift the target states one time step back.
if (i < iteratorList.size() - 1) {
delete currentTargetStates;
currentTargetStates = newTargetStates;
newTargetStates = new boost::container::flat_map<CompressedState, ValueType>();
}
}
// At this point, we applied all commands of the current command combination and newTargetStates
// contains all target states and their respective probabilities. That means we are now ready to
// add the choice to the list of transitions.
result.push_back(Choice<ValueType>(actionIndex));
// Now create the actual distribution.
Choice<ValueType>& choice = result.back();
// Remember the command labels only if we were asked to.
if (buildChoiceLabeling) {
// 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());
}
}
// Add the probabilities/rates to the newly created choice.
ValueType probabilitySum = storm::utility::zero<ValueType>();
for (auto const& stateProbabilityPair : *newTargetStates) {
StateType actualIndex = stateToIdCallback(stateProbabilityPair.first);
choice.addProbability(actualIndex, stateProbabilityPair.second);
probabilitySum += stateProbabilityPair.second;
}
// Check that the resulting distribution is in fact a distribution.
STORM_LOG_THROW(!program.isDiscreteTimeModel() || !comparator.isConstant(probabilitySum) || comparator.isOne(probabilitySum), storm::exceptions::WrongFormatException, "Sum of update probabilities do not some to one for some command (actually sum to " << probabilitySum << ").");
// Create the state-action reward for the newly created choice.
for (auto const& rewardModel : selectedRewardModels) {
ValueType stateActionRewardValue = storm::utility::zero<ValueType>();
if (rewardModel.get().hasStateActionRewards()) {
for (auto const& stateActionReward : rewardModel.get().getStateActionRewards()) {
if (stateActionReward.getActionIndex() == choice.getActionIndex() && evaluator.asBool(stateActionReward.getStatePredicateExpression())) {
stateActionRewardValue += ValueType(evaluator.asRational(stateActionReward.getRewardValueExpression())) * choice.getTotalMass();
}
}
}
choice.addChoiceReward(stateActionRewardValue);
}
// Dispose of the temporary maps.
delete currentTargetStates;
delete newTargetStates;
// Now, check whether there is one more command combination to consider.
bool movedIterator = false;
for (int_fast64_t j = iteratorList.size() - 1; j >= 0; --j) {
++iteratorList[j];
if (iteratorList[j] != activeCommandList[j].end()) {
movedIterator = true;
} else {
// Reset the iterator to the beginning of the list.
iteratorList[j] = activeCommandList[j].begin();
}
}
done = !movedIterator;
}
}
}
return result;
}
template class PrismNextStateGenerator<double>;
template class PrismNextStateGenerator<storm::RationalFunction>;
}
}