You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 

298 lines
25 KiB

#include "SparseMdpParameterLiftingModelChecker.h"
#include "storm/adapters/CarlAdapter.h"
#include "storm/modelchecker/propositional/SparsePropositionalModelChecker.h"
#include "storm/modelchecker/results/ExplicitQualitativeCheckResult.h"
#include "storm/modelchecker/results/ExplicitQuantitativeCheckResult.h"
#include "storm/models/sparse/Mdp.h"
#include "storm/models/sparse/StandardRewardModel.h"
#include "storm/utility/vector.h"
#include "storm/utility/graph.h"
#include "storm/solver/GameSolver.h"
#include "storm/logic/FragmentSpecification.h"
#include "storm/exceptions/InvalidArgumentException.h"
#include "storm/exceptions/InvalidPropertyException.h"
#include "storm/exceptions/NotSupportedException.h"
namespace storm {
namespace modelchecker {
namespace parametric {
template <typename SparseModelType, typename ConstantType>
SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::SparseMdpParameterLiftingModelChecker(SparseModelType const& parametricModel) : SparseParameterLiftingModelChecker<SparseModelType, ConstantType>(parametricModel), solverFactory(std::make_unique<storm::utility::solver::GameSolverFactory<ConstantType>>()) {
}
template <typename SparseModelType, typename ConstantType>
SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::SparseMdpParameterLiftingModelChecker(SparseModelType const& parametricModel, std::unique_ptr<storm::utility::solver::GameSolverFactory<ConstantType>>&& solverFactory) : SparseParameterLiftingModelChecker<SparseModelType, ConstantType>(parametricModel), solverFactory(std::move(solverFactory)) {
}
template <typename SparseModelType, typename ConstantType>
bool SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::canHandle(CheckTask<storm::logic::Formula, typename SparseModelType::ValueType> const& checkTask) const {
return checkTask.getFormula().isInFragment(storm::logic::reachability().setRewardOperatorsAllowed(true).setReachabilityRewardFormulasAllowed(true).setBoundedUntilFormulasAllowed(true).setCumulativeRewardFormulasAllowed(true));
}
template <typename SparseModelType, typename ConstantType>
void SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::specifyBoundedUntilFormula(CheckTask<storm::logic::BoundedUntilFormula, ConstantType> const& checkTask) {
// get the step bound
STORM_LOG_THROW(!checkTask.getFormula().hasLowerBound(), storm::exceptions::NotSupportedException, "Lower step bounds are not supported.");
STORM_LOG_THROW(checkTask.getFormula().hasUpperBound(), storm::exceptions::NotSupportedException, "Expected a bounded until formula with an upper bound.");
STORM_LOG_THROW(checkTask.getFormula().isStepBounded(), storm::exceptions::NotSupportedException, "Expected a bounded until formula with step bounds.");
stepBound = checkTask.getFormula().getUpperBound().evaluateAsInt();
STORM_LOG_THROW(*stepBound > 0, storm::exceptions::NotSupportedException, "Can not apply parameter lifting on step bounded formula: The step bound has to be positive.");
if (checkTask.getFormula().isUpperBoundStrict()) {
STORM_LOG_THROW(*stepBound > 0, storm::exceptions::NotSupportedException, "Expected a strict upper step bound that is greater than zero.");
--(*stepBound);
}
STORM_LOG_THROW(*stepBound > 0, storm::exceptions::NotSupportedException, "Can not apply parameter lifting on step bounded formula: The step bound has to be positive.");
// get the results for the subformulas
storm::modelchecker::SparsePropositionalModelChecker<SparseModelType> propositionalChecker(this->parametricModel);
STORM_LOG_THROW(propositionalChecker.canHandle(checkTask.getFormula().getLeftSubformula()) && propositionalChecker.canHandle(checkTask.getFormula().getRightSubformula()), storm::exceptions::NotSupportedException, "Parameter lifting with non-propositional subformulas is not supported");
storm::storage::BitVector phiStates = std::move(propositionalChecker.check(checkTask.getFormula().getLeftSubformula())->asExplicitQualitativeCheckResult().getTruthValuesVector());
storm::storage::BitVector psiStates = std::move(propositionalChecker.check(checkTask.getFormula().getRightSubformula())->asExplicitQualitativeCheckResult().getTruthValuesVector());
// get the maybeStates
maybeStates = storm::solver::minimize(checkTask.getOptimizationDirection()) ?
storm::utility::graph::performProbGreater0A(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), phiStates, psiStates, true, *stepBound) :
storm::utility::graph::performProbGreater0E(this->parametricModel.getBackwardTransitions(), phiStates, psiStates, true, *stepBound);
maybeStates &= ~psiStates;
// set the result for all non-maybe states
resultsForNonMaybeStates = std::vector<ConstantType>(this->parametricModel.getNumberOfStates(), storm::utility::zero<ConstantType>());
storm::utility::vector::setVectorValues(resultsForNonMaybeStates, psiStates, storm::utility::one<ConstantType>());
// if there are maybestates, create the parameterLifter
if (!maybeStates.empty()) {
// Create the vector of one-step probabilities to go to target states.
std::vector<typename SparseModelType::ValueType> b = this->parametricModel.getTransitionMatrix().getConstrainedRowSumVector(storm::storage::BitVector(this->parametricModel.getTransitionMatrix().getRowCount(), true), psiStates);
parameterLifter = std::make_unique<storm::transformer::ParameterLifter<typename SparseModelType::ValueType, ConstantType>>(this->parametricModel.getTransitionMatrix(), b, this->parametricModel.getTransitionMatrix().getRowIndicesOfRowGroups(maybeStates), maybeStates);
computePlayer1Matrix();
applyPreviousResultAsHint = false;
}
// We know some bounds for the results
lowerResultBound = storm::utility::zero<ConstantType>();
upperResultBound = storm::utility::one<ConstantType>();
}
template <typename SparseModelType, typename ConstantType>
void SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::specifyUntilFormula(CheckTask<storm::logic::UntilFormula, ConstantType> const& checkTask) {
// get the results for the subformulas
storm::modelchecker::SparsePropositionalModelChecker<SparseModelType> propositionalChecker(this->parametricModel);
STORM_LOG_THROW(propositionalChecker.canHandle(checkTask.getFormula().getLeftSubformula()) && propositionalChecker.canHandle(checkTask.getFormula().getRightSubformula()), storm::exceptions::NotSupportedException, "Parameter lifting with non-propositional subformulas is not supported");
storm::storage::BitVector phiStates = std::move(propositionalChecker.check(checkTask.getFormula().getLeftSubformula())->asExplicitQualitativeCheckResult().getTruthValuesVector());
storm::storage::BitVector psiStates = std::move(propositionalChecker.check(checkTask.getFormula().getRightSubformula())->asExplicitQualitativeCheckResult().getTruthValuesVector());
// get the maybeStates
std::pair<storm::storage::BitVector, storm::storage::BitVector> statesWithProbability01 = storm::solver::minimize(checkTask.getOptimizationDirection()) ?
storm::utility::graph::performProb01Min(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), phiStates, psiStates) :
storm::utility::graph::performProb01Max(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), phiStates, psiStates);
maybeStates = ~(statesWithProbability01.first | statesWithProbability01.second);
// set the result for all non-maybe states
resultsForNonMaybeStates = std::vector<ConstantType>(this->parametricModel.getNumberOfStates(), storm::utility::zero<ConstantType>());
storm::utility::vector::setVectorValues(resultsForNonMaybeStates, statesWithProbability01.second, storm::utility::one<ConstantType>());
// if there are maybestates, create the parameterLifter
if (!maybeStates.empty()) {
// Create the vector of one-step probabilities to go to target states.
std::vector<typename SparseModelType::ValueType> b = this->parametricModel.getTransitionMatrix().getConstrainedRowSumVector(storm::storage::BitVector(this->parametricModel.getTransitionMatrix().getRowCount(), true), statesWithProbability01.second);
parameterLifter = std::make_unique<storm::transformer::ParameterLifter<typename SparseModelType::ValueType, ConstantType>>(this->parametricModel.getTransitionMatrix(), b, this->parametricModel.getTransitionMatrix().getRowIndicesOfRowGroups(maybeStates), maybeStates);
computePlayer1Matrix();
// Check whether there is an EC consisting of maybestates
applyPreviousResultAsHint = storm::solver::minimize(checkTask.getOptimizationDirection()) || // when minimizing, there can not be an EC within the maybestates
storm::utility::graph::performProb1A(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), maybeStates, ~maybeStates).full();
}
// We know some bounds for the results
lowerResultBound = storm::utility::zero<ConstantType>();
upperResultBound = storm::utility::one<ConstantType>();
}
template <typename SparseModelType, typename ConstantType>
void SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::specifyReachabilityRewardFormula(CheckTask<storm::logic::EventuallyFormula, ConstantType> const& checkTask) {
// get the results for the subformula
storm::modelchecker::SparsePropositionalModelChecker<SparseModelType> propositionalChecker(this->parametricModel);
STORM_LOG_THROW(propositionalChecker.canHandle(checkTask.getFormula().getSubformula()), storm::exceptions::NotSupportedException, "Parameter lifting with non-propositional subformulas is not supported");
storm::storage::BitVector targetStates = std::move(propositionalChecker.check(checkTask.getFormula().getSubformula())->asExplicitQualitativeCheckResult().getTruthValuesVector());
// get the maybeStates
storm::storage::BitVector infinityStates = storm::solver::minimize(checkTask.getOptimizationDirection()) ?
storm::utility::graph::performProb1E(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), storm::storage::BitVector(this->parametricModel.getNumberOfStates(), true), targetStates) :
storm::utility::graph::performProb1A(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), storm::storage::BitVector(this->parametricModel.getNumberOfStates(), true), targetStates);
infinityStates.complement();
maybeStates = ~(targetStates | infinityStates);
// set the result for all the non-maybe states
resultsForNonMaybeStates = std::vector<ConstantType>(this->parametricModel.getNumberOfStates(), storm::utility::zero<ConstantType>());
storm::utility::vector::setVectorValues(resultsForNonMaybeStates, infinityStates, storm::utility::infinity<ConstantType>());
// if there are maybestates, create the parameterLifter
if (!maybeStates.empty()) {
// Create the reward vector
STORM_LOG_THROW((checkTask.isRewardModelSet() && this->parametricModel.hasRewardModel(checkTask.getRewardModel())) || (!checkTask.isRewardModelSet() && this->parametricModel.hasUniqueRewardModel()), storm::exceptions::InvalidPropertyException, "The reward model specified by the CheckTask is not available in the given model.");
typename SparseModelType::RewardModelType const& rewardModel = checkTask.isRewardModelSet() ? this->parametricModel.getRewardModel(checkTask.getRewardModel()) : this->parametricModel.getUniqueRewardModel();
std::vector<typename SparseModelType::ValueType> b = rewardModel.getTotalRewardVector(this->parametricModel.getTransitionMatrix());
// We need to handle choices that lead to an infinity state.
// As a maybeState does not have reward infinity, such a choice will never be picked. Hence, we can unselect the corresponding rows
storm::storage::BitVector selectedRows = this->parametricModel.getTransitionMatrix().getRowIndicesOfRowGroups(maybeStates);
for (uint_fast64_t row : selectedRows) {
for (auto const& element : this->parametricModel.getTransitionMatrix().getRow(row)) {
if (infinityStates.get(element.getColumn())) {
selectedRows.set(row, false);
break;
}
}
}
parameterLifter = std::make_unique<storm::transformer::ParameterLifter<typename SparseModelType::ValueType, ConstantType>>(this->parametricModel.getTransitionMatrix(), b, selectedRows, maybeStates);
computePlayer1Matrix();
// Check whether there is an EC consisting of maybestates
applyPreviousResultAsHint = !storm::solver::minimize(checkTask.getOptimizationDirection()) || // when maximizing, there can not be an EC within the maybestates
storm::utility::graph::performProb1A(this->parametricModel.getTransitionMatrix(), this->parametricModel.getTransitionMatrix().getRowGroupIndices(), this->parametricModel.getBackwardTransitions(), maybeStates, ~maybeStates).full();
}
// We only know a lower bound for the result
lowerResultBound = storm::utility::zero<ConstantType>();
}
template <typename SparseModelType, typename ConstantType>
void SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::specifyCumulativeRewardFormula(CheckTask<storm::logic::CumulativeRewardFormula, ConstantType> const& checkTask) {
// Obtain the stepBound
stepBound = checkTask.getFormula().getBound().evaluateAsInt();
if (checkTask.getFormula().isBoundStrict()) {
STORM_LOG_THROW(*stepBound > 0, storm::exceptions::NotSupportedException, "Expected a strict upper step bound that is greater than zero.");
--(*stepBound);
}
STORM_LOG_THROW(*stepBound > 0, storm::exceptions::NotSupportedException, "Can not apply parameter lifting on step bounded formula: The step bound has to be positive.");
//Every state is a maybeState
maybeStates = storm::storage::BitVector(this->parametricModel.getTransitionMatrix().getColumnCount(), true);
resultsForNonMaybeStates = std::vector<ConstantType>(this->parametricModel.getNumberOfStates());
// Create the reward vector
STORM_LOG_THROW((checkTask.isRewardModelSet() && this->parametricModel.hasRewardModel(checkTask.getRewardModel())) || (!checkTask.isRewardModelSet() && this->parametricModel.hasUniqueRewardModel()), storm::exceptions::InvalidPropertyException, "The reward model specified by the CheckTask is not available in the given model.");
typename SparseModelType::RewardModelType const& rewardModel = checkTask.isRewardModelSet() ? this->parametricModel.getRewardModel(checkTask.getRewardModel()) : this->parametricModel.getUniqueRewardModel();
std::vector<typename SparseModelType::ValueType> b = rewardModel.getTotalRewardVector(this->parametricModel.getTransitionMatrix());
parameterLifter = std::make_unique<storm::transformer::ParameterLifter<typename SparseModelType::ValueType, ConstantType>>(this->parametricModel.getTransitionMatrix(), b, maybeStates, maybeStates);
applyPreviousResultAsHint = false;
// We only know a lower bound for the result
lowerResultBound = storm::utility::zero<ConstantType>();
}
template <typename SparseModelType, typename ConstantType>
std::unique_ptr<CheckResult> SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::computeQuantitativeValues(storm::storage::ParameterRegion<typename SparseModelType::ValueType> const& region, storm::solver::OptimizationDirection const& dirForParameters) {
if(maybeStates.empty()) {
return std::make_unique<storm::modelchecker::ExplicitQuantitativeCheckResult<ConstantType>>(resultsForNonMaybeStates);
}
parameterLifter->specifyRegion(region, dirForParameters);
// Set up the solver
auto solver = solverFactory->create(player1Matrix, parameterLifter->getMatrix());
if(lowerResultBound) solver->setLowerBound(lowerResultBound.get());
if(upperResultBound) solver->setUpperBound(upperResultBound.get());
if(applyPreviousResultAsHint) {
solver->setTrackScheduler(true);
x.resize(maybeStates.getNumberOfSetBits(), storm::utility::zero<ConstantType>());
if(storm::solver::minimize(dirForParameters) && minSched && player1Sched) solver->setSchedulerHint(std::move(player1Sched.get()), std::move(minSched.get()));
if(storm::solver::maximize(dirForParameters) && maxSched && player1Sched) solver->setSchedulerHint(std::move(player1Sched.get()), std::move(maxSched.get()));
} else {
x.assign(maybeStates.getNumberOfSetBits(), storm::utility::zero<ConstantType>());
}
if (this->currentCheckTask->isBoundSet() && this->currentCheckTask->getOptimizationDirection() == dirForParameters && solver->hasSchedulerHints()) {
// If we reach this point, we know that after applying the hints, the x-values can only become larger (if we maximize) or smaller (if we minimize).
std::unique_ptr<storm::solver::TerminationCondition<ConstantType>> termCond;
storm::storage::BitVector relevantStatesInSubsystem = this->currentCheckTask->isOnlyInitialStatesRelevantSet() ? this->parametricModel.getInitialStates() % maybeStates : storm::storage::BitVector(maybeStates.getNumberOfSetBits(), true);
if (storm::solver::minimize(dirForParameters)) {
// Terminate if the value for ALL relevant states is already below the threshold
termCond = std::make_unique<storm::solver::TerminateIfFilteredExtremumBelowThreshold<ConstantType>> (relevantStatesInSubsystem, this->currentCheckTask->getBoundThreshold(), true, false);
} else {
// Terminate if the value for ALL relevant states is already above the threshold
termCond = std::make_unique<storm::solver::TerminateIfFilteredExtremumExceedsThreshold<ConstantType>> (relevantStatesInSubsystem, this->currentCheckTask->getBoundThreshold(), true, true);
}
solver->setTerminationCondition(std::move(termCond));
}
// Invoke the solver
if(stepBound) {
assert(*stepBound > 0);
solver->repeatedMultiply(this->currentCheckTask->getOptimizationDirection(), dirForParameters, x, &parameterLifter->getVector(), *stepBound);
} else {
solver->solveGame(this->currentCheckTask->getOptimizationDirection(), dirForParameters, x, parameterLifter->getVector());
if(applyPreviousResultAsHint) {
if(storm::solver::minimize(dirForParameters)) {
minSched = solver->getPlayer2Scheduler();
} else {
maxSched = solver->getPlayer2Scheduler();
}
player1Sched = solver->getPlayer1Scheduler();
}
}
// Get the result for the complete model (including maybestates)
std::vector<ConstantType> result = resultsForNonMaybeStates;
auto maybeStateResIt = x.begin();
for(auto const& maybeState : maybeStates) {
result[maybeState] = *maybeStateResIt;
++maybeStateResIt;
}
return std::make_unique<storm::modelchecker::ExplicitQuantitativeCheckResult<ConstantType>>(std::move(result));
}
template <typename SparseModelType, typename ConstantType>
void SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::computePlayer1Matrix() {
uint_fast64_t n = 0;
for(auto const& maybeState : maybeStates) {
n += this->parametricModel.getTransitionMatrix().getRowGroupSize(maybeState);
}
// The player 1 matrix is the identity matrix of size n with the row groups as given by the original matrix
storm::storage::SparseMatrixBuilder<storm::storage::sparse::state_type> matrixBuilder(n, n, n, true, true, maybeStates.getNumberOfSetBits());
uint_fast64_t p1MatrixRow = 0;
for (auto maybeState : maybeStates){
matrixBuilder.newRowGroup(p1MatrixRow);
for (uint_fast64_t row = this->parametricModel.getTransitionMatrix().getRowGroupIndices()[maybeState]; row < this->parametricModel.getTransitionMatrix().getRowGroupIndices()[maybeState + 1]; ++row){
matrixBuilder.addNextValue(p1MatrixRow, p1MatrixRow, storm::utility::one<storm::storage::sparse::state_type>());
++p1MatrixRow;
}
}
player1Matrix = matrixBuilder.build();
}
template <typename SparseModelType, typename ConstantType>
void SparseMdpParameterLiftingModelChecker<SparseModelType, ConstantType>::reset() {
maybeStates.resize(0);
resultsForNonMaybeStates.clear();
stepBound = boost::none;
player1Matrix = storm::storage::SparseMatrix<storm::storage::sparse::state_type>();
parameterLifter = nullptr;
minSched = boost::none;
maxSched = boost::none;
x.clear();
lowerResultBound = boost::none;
upperResultBound = boost::none;
applyPreviousResultAsHint = false;
}
template class SparseMdpParameterLiftingModelChecker<storm::models::sparse::Mdp<storm::RationalFunction>, double>;
}
}
}