#include "src/generator/PrismNextStateGenerator.h" #include #include "src/models/sparse/StateLabeling.h" #include "src/storage/expressions/SimpleValuation.h" #include "src/solver/SmtSolver.h" #include "src/utility/constants.h" #include "src/utility/macros.h" #include "src/exceptions/InvalidArgumentException.h" #include "src/exceptions/WrongFormatException.h" namespace storm { namespace generator { template PrismNextStateGenerator::PrismNextStateGenerator(storm::prism::Program const& program, NextStateGeneratorOptions const& options) : PrismNextStateGenerator(program.substituteConstants(), options, false) { // Intentionally left empty. } template PrismNextStateGenerator::PrismNextStateGenerator(storm::prism::Program const& program, NextStateGeneratorOptions const& options, bool flag) : NextStateGenerator(program.getManager(), options), program(program), rewardModels() { STORM_LOG_TRACE("Creating next-state generator for PRISM program: " << program); STORM_LOG_THROW(!this->program.specifiesSystemComposition(), storm::exceptions::WrongFormatException, "The explicit next-state generator currently does not support custom system compositions."); // Only after checking validity of the program, we initialize the variable information. this->checkValid(); this->variableInformation = VariableInformation(program); // Create a proper evalator. this->evaluator = std::make_unique>(program.getManager()); if (this->options.isBuildAllRewardModelsSet()) { for (auto const& rewardModel : this->program.getRewardModels()) { rewardModels.push_back(rewardModel); } } else { // Extract the reward models from the program based on the names we were given. for (auto const& rewardModelName : this->options.getRewardModelNames()) { if (this->program.hasRewardModel(rewardModelName)) { rewardModels.push_back(this->program.getRewardModel(rewardModelName)); } else { STORM_LOG_THROW(rewardModelName.empty(), storm::exceptions::InvalidArgumentException, "Cannot build unknown reward model '" << rewardModelName << "'."); STORM_LOG_THROW(this->program.getNumberOfRewardModels() == 1, storm::exceptions::InvalidArgumentException, "Reference to standard reward model is ambiguous."); } } // If no reward model was yet added, but there was one that was given in the options, we try to build the // standard reward model. if (rewardModels.empty() && !this->options.getRewardModelNames().empty()) { rewardModels.push_back(this->program.getRewardModel(0)); } } // Determine whether any reward model has state action rewards. for (auto const& rewardModel : rewardModels) { hasStateActionRewards |= rewardModel.get().hasStateActionRewards(); } // If there are terminal states we need to handle, we now need to translate all labels to expressions. if (this->options.hasTerminalStates()) { for (auto const& expressionOrLabelAndBool : this->options.getTerminalStates()) { if (expressionOrLabelAndBool.first.isExpression()) { this->terminalStates.push_back(std::make_pair(expressionOrLabelAndBool.first.getExpression(), expressionOrLabelAndBool.second)); } else { if (program.hasLabel(expressionOrLabelAndBool.first.getLabel())) { this->terminalStates.push_back(std::make_pair(this->program.getLabelExpression(expressionOrLabelAndBool.first.getLabel()), expressionOrLabelAndBool.second)); } else { // If the label is not present in the program and is not a special one, we raise an error. STORM_LOG_THROW(expressionOrLabelAndBool.first.getLabel() == "init" || expressionOrLabelAndBool.first.getLabel() == "deadlock", storm::exceptions::InvalidArgumentException, "Terminal states refer to illegal label '" << expressionOrLabelAndBool.first.getLabel() << "'."); } } } } } template void PrismNextStateGenerator::checkValid() const { // If the program still contains undefined constants and we are not in a parametric setting, assemble an appropriate error message. #ifdef STORM_HAVE_CARL if (!std::is_same::value && program.hasUndefinedConstants()) { #else if (program.hasUndefinedConstants()) { #endif std::vector> undefinedConstants = program.getUndefinedConstants(); std::stringstream stream; bool printComma = false; for (auto const& constant : undefinedConstants) { if (printComma) { stream << ", "; } else { printComma = true; } stream << constant.get().getName() << " (" << constant.get().getType() << ")"; } stream << "."; STORM_LOG_THROW(false, storm::exceptions::InvalidArgumentException, "Program still contains these undefined constants: " + stream.str()); } #ifdef STORM_HAVE_CARL else if (std::is_same::value && !program.undefinedConstantsAreGraphPreserving()) { STORM_LOG_THROW(false, storm::exceptions::InvalidArgumentException, "The program contains undefined constants that appear in some places other than update probabilities and reward value expressions, which is not admitted."); } #endif } template ModelType PrismNextStateGenerator::getModelType() const { switch (program.getModelType()) { case storm::prism::Program::ModelType::DTMC: return ModelType::DTMC; case storm::prism::Program::ModelType::CTMC: return ModelType::CTMC; case storm::prism::Program::ModelType::MDP: return ModelType::MDP; case storm::prism::Program::ModelType::MA: return ModelType::MA; default: STORM_LOG_THROW(false, storm::exceptions::WrongFormatException, "Invalid model type."); } } template bool PrismNextStateGenerator::isDeterministicModel() const { return program.isDeterministicModel(); } template bool PrismNextStateGenerator::isDiscreteTimeModel() const { return program.isDiscreteTimeModel(); } template std::vector PrismNextStateGenerator::getInitialStates(StateToIdCallback const& stateToIdCallback) { // Prepare an SMT solver to enumerate all initial states. storm::utility::solver::SmtSolverFactory factory; std::unique_ptr solver = factory.create(program.getManager()); std::vector rangeExpressions = program.getAllRangeExpressions(); for (auto const& expression : rangeExpressions) { solver->add(expression); } solver->add(program.getInitialStatesExpression()); // Proceed ss long as the solver can still enumerate initial states. std::vector initialStateIndices; while (solver->check() == storm::solver::SmtSolver::CheckResult::Sat) { // Create fresh state. CompressedState initialState(this->variableInformation.getTotalBitOffset()); // Read variable assignment from the solution of the solver. Also, create an expression we can use to // prevent the variable assignment from being enumerated again. storm::expressions::Expression blockingExpression; std::shared_ptr model = solver->getModel(); for (auto const& booleanVariable : this->variableInformation.booleanVariables) { bool variableValue = model->getBooleanValue(booleanVariable.variable); storm::expressions::Expression localBlockingExpression = variableValue ? !booleanVariable.variable : booleanVariable.variable; blockingExpression = blockingExpression.isInitialized() ? blockingExpression || localBlockingExpression : localBlockingExpression; initialState.set(booleanVariable.bitOffset, variableValue); } for (auto const& integerVariable : this->variableInformation.integerVariables) { int_fast64_t variableValue = model->getIntegerValue(integerVariable.variable); storm::expressions::Expression localBlockingExpression = integerVariable.variable != model->getManager().integer(variableValue); blockingExpression = blockingExpression.isInitialized() ? blockingExpression || localBlockingExpression : localBlockingExpression; initialState.setFromInt(integerVariable.bitOffset, integerVariable.bitWidth, static_cast(variableValue - integerVariable.lowerBound)); } // Register initial state and return it. StateType id = stateToIdCallback(initialState); initialStateIndices.push_back(id); // Block the current initial state to search for the next one. if (!blockingExpression.isInitialized()) { break; } solver->add(blockingExpression); } return initialStateIndices; } template StateBehavior PrismNextStateGenerator::expand(StateToIdCallback const& stateToIdCallback) { // Prepare the result, in case we return early. StateBehavior 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 : rewardModels) { ValueType stateRewardValue = storm::utility::zero(); if (rewardModel.get().hasStateRewards()) { for (auto const& stateReward : rewardModel.get().getStateRewards()) { if (this->evaluator->asBool(stateReward.getStatePredicateExpression())) { stateRewardValue += ValueType(this->evaluator->asRational(stateReward.getRewardValueExpression())); } } } result.addStateReward(stateRewardValue); } // If a terminal expression was set and we must not expand this state, return now. if (!this->terminalStates.empty()) { for (auto const& expressionBool : this->terminalStates) { if (this->evaluator->asBool(expressionBool.first) == expressionBool.second) { return result; } } } // Get all choices for the state. result.setExpanded(); std::vector> allChoices = getUnlabeledChoices(*this->state, stateToIdCallback); std::vector> 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 (this->isDeterministicModel() && totalNumberOfChoices > 1) { Choice 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(totalNumberOfChoices) : storm::utility::zero(); // Iterate over all choices and combine the probabilities/rates into one choice. for (auto const& choice : allChoices) { for (auto const& stateProbabilityPair : choice) { if (this->isDiscreteTimeModel()) { globalChoice.addProbability(stateProbabilityPair.first, stateProbabilityPair.second / totalNumberOfChoices); } else { globalChoice.addProbability(stateProbabilityPair.first, stateProbabilityPair.second); } } if (hasStateActionRewards && !this->isDiscreteTimeModel()) { totalExitRate += choice.getTotalMass(); } if (this->options.isBuildChoiceLabelsSet()) { globalChoice.addLabels(choice.getLabels()); } } // Now construct the state-action reward for all selected reward models. for (auto const& rewardModel : rewardModels) { ValueType stateActionRewardValue = storm::utility::zero(); if (rewardModel.get().hasStateActionRewards()) { for (auto const& stateActionReward : rewardModel.get().getStateActionRewards()) { for (auto const& choice : allChoices) { if (stateActionReward.getActionIndex() == choice.getActionIndex() && this->evaluator->asBool(stateActionReward.getStatePredicateExpression())) { stateActionRewardValue += ValueType(this->evaluator->asRational(stateActionReward.getRewardValueExpression())) * choice.getTotalMass() / totalExitRate; } } } } globalChoice.addReward(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)); } this->postprocess(result); return result; } template CompressedState PrismNextStateGenerator::applyUpdate(CompressedState const& state, storm::prism::Update const& update) { CompressedState newState(state); // NOTE: the following process assumes that the assignments of the update are ordered in such a way that the // assignments to boolean variables precede the assignments to all integer variables and that within the // types, the assignments to variables are ordered (in ascending order) by the expression variables. // This is guaranteed for PRISM models, by sorting the assignments as soon as an update is created. auto assignmentIt = update.getAssignments().begin(); auto assignmentIte = update.getAssignments().end(); // Iterate over all boolean assignments and carry them out. auto boolIt = this->variableInformation.booleanVariables.begin(); for (; assignmentIt != assignmentIte && assignmentIt->getExpression().hasBooleanType(); ++assignmentIt) { while (assignmentIt->getVariable() != boolIt->variable) { ++boolIt; } newState.set(boolIt->bitOffset, this->evaluator->asBool(assignmentIt->getExpression())); } // Iterate over all integer assignments and carry them out. auto integerIt = this->variableInformation.integerVariables.begin(); for (; assignmentIt != assignmentIte && assignmentIt->getExpression().hasIntegerType(); ++assignmentIt) { while (assignmentIt->getVariable() != integerIt->variable) { ++integerIt; } int_fast64_t assignedValue = this->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() << "'."); STORM_LOG_THROW(assignedValue >= integerIt->lowerBound, 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(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 boost::optional>>> PrismNextStateGenerator::getActiveCommandsByActionIndex(uint_fast64_t const& actionIndex) { boost::optional>>> result((std::vector>>())); // 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 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> 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 (this->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>>>(); } result.get().push_back(std::move(commands)); } return result; } template std::vector> PrismNextStateGenerator::getUnlabeledChoices(CompressedState const& state, StateToIdCallback stateToIdCallback) { std::vector> 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 (!this->evaluator->asBool(command.getGuardExpression())) { continue; } result.push_back(Choice(command.getActionIndex(), command.isMarkovian())); Choice& choice = result.back(); // Remember the command labels only if we were asked to. if (this->options.isBuildChoiceLabelsSet()) { choice.addLabel(command.getGlobalIndex()); } // Iterate over all updates of the current command. ValueType probabilitySum = storm::utility::zero(); 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 = this->evaluator->asRational(update.getLikelihoodExpression()); choice.addProbability(stateIndex, probability); probabilitySum += probability; } // Create the state-action reward for the newly created choice. for (auto const& rewardModel : rewardModels) { ValueType stateActionRewardValue = storm::utility::zero(); if (rewardModel.get().hasStateActionRewards()) { for (auto const& stateActionReward : rewardModel.get().getStateActionRewards()) { if (stateActionReward.getActionIndex() == choice.getActionIndex() && this->evaluator->asBool(stateActionReward.getStatePredicateExpression())) { stateActionRewardValue += ValueType(this->evaluator->asRational(stateActionReward.getRewardValueExpression())); } } } choice.addReward(stateActionRewardValue); } // Check that the resulting distribution is in fact a distribution. STORM_LOG_THROW(!program.isDiscreteTimeModel() || this->comparator.isOne(probabilitySum), storm::exceptions::WrongFormatException, "Probabilities do not sum to one for command '" << command << "' (actually sum to " << probabilitySum << ")."); } } return result; } template std::vector> PrismNextStateGenerator::getLabeledChoices(CompressedState const& state, StateToIdCallback stateToIdCallback) { std::vector> result; for (uint_fast64_t actionIndex : program.getSynchronizingActionIndices()) { boost::optional>>> optionalActiveCommandLists = getActiveCommandsByActionIndex(actionIndex); // Only process this action label, if there is at least one feasible solution. if (optionalActiveCommandLists) { std::vector>> const& activeCommandList = optionalActiveCommandLists.get(); std::vector>::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* currentTargetStates = new boost::container::flat_map(); boost::container::flat_map* newTargetStates = new boost::container::flat_map(); currentTargetStates->emplace(state, storm::utility::one()); 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); // If the new state was already found as a successor state, update the probability // and otherwise insert it. auto targetStateIt = newTargetStates->find(newTargetState); if (targetStateIt != newTargetStates->end()) { targetStateIt->second += stateProbabilityPair.second * this->evaluator->asRational(update.getLikelihoodExpression()); } else { newTargetStates->emplace(newTargetState, stateProbabilityPair.second * this->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(); } } // 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(actionIndex)); // Now create the actual distribution. Choice& choice = result.back(); // Remember the command labels only if we were asked to. if (this->options.isBuildChoiceLabelsSet()) { // Add the labels of all commands to this choice. for (uint_fast64_t i = 0; i < iteratorList.size(); ++i) { choice.addLabel(iteratorList[i]->get().getGlobalIndex()); } } // Add the probabilities/rates to the newly created choice. ValueType probabilitySum = storm::utility::zero(); 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() || !this->comparator.isConstant(probabilitySum) || this->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 : rewardModels) { ValueType stateActionRewardValue = storm::utility::zero(); if (rewardModel.get().hasStateActionRewards()) { for (auto const& stateActionReward : rewardModel.get().getStateActionRewards()) { if (stateActionReward.getActionIndex() == choice.getActionIndex() && this->evaluator->asBool(stateActionReward.getStatePredicateExpression())) { stateActionRewardValue += ValueType(this->evaluator->asRational(stateActionReward.getRewardValueExpression())); } } } choice.addReward(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; !movedIterator && 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 storm::models::sparse::StateLabeling PrismNextStateGenerator::label(storm::storage::BitVectorHashMap const& states, std::vector const& initialStateIndices, std::vector const& deadlockStateIndices) { // Gather a vector of labels and their expressions. std::vector> labels; if (this->options.isBuildAllLabelsSet()) { for (auto const& label : program.getLabels()) { labels.push_back(std::make_pair(label.getName(), label.getStatePredicateExpression())); } } else { for (auto const& labelName : this->options.getLabelNames()) { if (program.hasLabel(labelName)) { labels.push_back(std::make_pair(labelName, program.getLabelExpression(labelName))); } else { STORM_LOG_THROW(labelName == "init" || labelName == "deadlock", storm::exceptions::InvalidArgumentException, "Cannot build labeling for unknown label '" << labelName << "'."); } } } return NextStateGenerator::label(states, initialStateIndices, deadlockStateIndices, labels); } template std::size_t PrismNextStateGenerator::getNumberOfRewardModels() const { return rewardModels.size(); } template RewardModelInformation PrismNextStateGenerator::getRewardModelInformation(uint64_t const& index) const { storm::prism::RewardModel const& rewardModel = rewardModels[index].get(); return RewardModelInformation(rewardModel.getName(), rewardModel.hasStateRewards(), rewardModel.hasStateActionRewards(), rewardModel.hasTransitionRewards()); } template class PrismNextStateGenerator; #ifdef STORM_HAVE_CARL template class PrismNextStateGenerator; template class PrismNextStateGenerator; #endif } }