338 lines
22 KiB

#include "storm/modelchecker/multiobjective/pcaa/SparseMdpRewardBoundedPcaaWeightVectorChecker.h"
#include "storm/adapters/RationalFunctionAdapter.h"
#include "storm/models/sparse/Mdp.h"
#include "storm/models/sparse/StandardRewardModel.h"
#include "storm/utility/macros.h"
#include "storm/utility/vector.h"
#include "storm/logic/Formulas.h"
#include "storm/solver/MinMaxLinearEquationSolver.h"
#include "storm/solver/LinearEquationSolver.h"
#include "storm/settings/SettingsManager.h"
#include "storm/utility/export.h"
#include "storm/settings/modules/IOSettings.h"
#include "storm/exceptions/InvalidPropertyException.h"
#include "storm/exceptions/InvalidOperationException.h"
#include "storm/exceptions/IllegalArgumentException.h"
#include "storm/exceptions/NotSupportedException.h"
#include "storm/exceptions/UnexpectedException.h"
#include "storm/exceptions/UncheckedRequirementException.h"
namespace storm {
namespace modelchecker {
namespace multiobjective {
template <class SparseMdpModelType>
SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::SparseMdpRewardBoundedPcaaWeightVectorChecker(SparseMultiObjectivePreprocessorResult<SparseMdpModelType> const& preprocessorResult) : PcaaWeightVectorChecker<SparseMdpModelType>(preprocessorResult.objectives), swAll(true), rewardUnfolding(*preprocessorResult.preprocessedModel, preprocessorResult.objectives) {
STORM_LOG_THROW(preprocessorResult.rewardFinitenessType == SparseMultiObjectivePreprocessorResult<SparseMdpModelType>::RewardFinitenessType::AllFinite, storm::exceptions::NotSupportedException, "There is a scheduler that yields infinite reward for one objective. This is not supported.");
STORM_LOG_THROW(preprocessorResult.preprocessedModel->getInitialStates().getNumberOfSetBits() == 1, storm::exceptions::NotSupportedException, "The model has multiple initial states.");
numSchedChanges = 0;
numCheckedEpochs = 0;
numChecks = 0;
}
template <class SparseMdpModelType>
void SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::check(std::vector<ValueType> const& weightVector) {
STORM_LOG_DEBUG("Analyzing weight vector " << storm::utility::vector::toString(weightVector));
++numChecks;
// In case we want to export the cdf, we will collect the corresponding data
std::vector<std::vector<ValueType>> cdfData;
auto initEpoch = rewardUnfolding.getStartEpoch();
auto epochOrder = rewardUnfolding.getEpochComputationOrder(initEpoch);
EpochCheckingData cachedData;
for (auto const& epoch : epochOrder) {
computeEpochSolution(epoch, weightVector, cachedData);
if (storm::settings::getModule<storm::settings::modules::IOSettings>().isExportCdfSet() && !rewardUnfolding.getEpochManager().hasBottomDimension(epoch)) {
std::vector<ValueType> cdfEntry;
for (uint64_t i = 0; i < rewardUnfolding.getEpochManager().getDimensionCount(); ++i) {
uint64_t offset = rewardUnfolding.getDimension(i).isUpperBounded ? 0 : 1;
cdfEntry.push_back(storm::utility::convertNumber<ValueType>(rewardUnfolding.getEpochManager().getDimensionOfEpoch(epoch, i) + offset) * rewardUnfolding.getDimension(i).scalingFactor);
}
auto const& solution = rewardUnfolding.getInitialStateResult(epoch);
auto solutionIt = solution.begin();
++solutionIt;
cdfEntry.insert(cdfEntry.end(), solutionIt, solution.end());
cdfData.push_back(std::move(cdfEntry));
}
}
if (storm::settings::getModule<storm::settings::modules::IOSettings>().isExportCdfSet()) {
std::vector<std::string> headers;
for (uint64_t i = 0; i < rewardUnfolding.getEpochManager().getDimensionCount(); ++i) {
headers.push_back("obj" + std::to_string(rewardUnfolding.getDimension(i).objectiveIndex) + ":" + rewardUnfolding.getDimension(i).formula->toString());
}
for (uint64_t i = 0; i < this->objectives.size(); ++i) {
headers.push_back("obj" + std::to_string(i));
}
storm::utility::exportDataToCSVFile<ValueType, ValueType, std::string>("cdf" + std::to_string(numChecks) + ".csv", cdfData, weightVector, headers);
}
auto solution = rewardUnfolding.getInitialStateResult(initEpoch);
// Todo: we currently assume precise results...
auto solutionIt = solution.begin();
++solutionIt;
underApproxResult = std::vector<ValueType>(solutionIt, solution.end());
overApproxResult = underApproxResult;
}
template <class SparseMdpModelType>
void SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::computeEpochSolution(typename MultiDimensionalRewardUnfolding<ValueType, false>::Epoch const& epoch, std::vector<ValueType> const& weightVector, EpochCheckingData& cachedData) {
++numCheckedEpochs;
swEpochModelBuild.start();
auto& epochModel = rewardUnfolding.setCurrentEpoch(epoch);
swEpochModelBuild.stop();
swEpochModelAnalysis.start();
std::vector<typename MultiDimensionalRewardUnfolding<ValueType, false>::SolutionType> result;
result.reserve(epochModel.epochInStates.getNumberOfSetBits());
uint64_t solutionSize = this->objectives.size() + 1;
// If the epoch matrix is empty we do not need to solve linear equation systems
if (epochModel.epochMatrix.getEntryCount() == 0) {
std::vector<ValueType> weights = weightVector;
for (uint64_t objIndex = 0; objIndex < this->objectives.size(); ++objIndex) {
if (storm::solver::minimize(this->objectives[objIndex].formula->getOptimalityType())) {
weights[objIndex] *= -storm::utility::one<ValueType>();
}
}
auto stepSolutionIt = epochModel.stepSolutions.begin();
auto stepChoiceIt = epochModel.stepChoices.begin();
for (auto const& state : epochModel.epochInStates) {
// Obtain the best choice for this state according to the weighted combination of objectives
ValueType bestValue;
uint64_t bestChoice = std::numeric_limits<uint64_t>::max();
auto bestChoiceStepSolutionIt = epochModel.stepSolutions.end();
uint64_t lastChoice = epochModel.epochMatrix.getRowGroupIndices()[state + 1];
bool firstChoice = true;
for (uint64_t choice = epochModel.epochMatrix.getRowGroupIndices()[state]; choice < lastChoice; ++choice) {
ValueType choiceValue = storm::utility::zero<ValueType>();
// Obtain the (weighted) objective rewards
for (uint64_t objIndex = 0; objIndex < this->objectives.size(); ++objIndex) {
if (epochModel.objectiveRewardFilter[objIndex].get(choice)) {
choiceValue += weights[objIndex] * epochModel.objectiveRewards[objIndex][choice];
}
}
// Obtain the step solution if this is a step choice
while (*stepChoiceIt < choice) {
++stepChoiceIt;
++stepSolutionIt;
}
if (*stepChoiceIt == choice) {
choiceValue += stepSolutionIt->front();
// Check if this choice is better
if (firstChoice || choiceValue > bestValue) {
bestValue = std::move(choiceValue);
bestChoice = choice;
bestChoiceStepSolutionIt = stepSolutionIt;
}
} else if (firstChoice || choiceValue > bestValue) {
bestValue = std::move(choiceValue);
bestChoice = choice;
bestChoiceStepSolutionIt = epochModel.stepSolutions.end();
}
firstChoice = false;
}
// Insert the solution w.r.t. this choice
result.emplace_back();
result.back().reserve(solutionSize);
result.back().push_back(std::move(bestValue));
if (bestChoiceStepSolutionIt != epochModel.stepSolutions.end()) {
for (uint64_t objIndex = 0; objIndex < this->objectives.size(); ++objIndex) {
if (epochModel.objectiveRewardFilter[objIndex].get(bestChoice)) {
result.back().push_back((epochModel.objectiveRewards[objIndex][bestChoice] + (*bestChoiceStepSolutionIt)[objIndex + 1]));
} else {
result.back().push_back((*bestChoiceStepSolutionIt)[objIndex + 1]);
}
}
} else {
for (uint64_t objIndex = 0; objIndex < this->objectives.size(); ++objIndex) {
if (epochModel.objectiveRewardFilter[objIndex].get(bestChoice)) {
result.back().push_back((epochModel.objectiveRewards[objIndex][bestChoice]));
} else {
result.back().push_back(storm::utility::zero<ValueType>());
}
}
}
}
} else {
updateCachedData(epochModel, cachedData, weightVector);
// Formulate a min-max equation system max(A*x+b)=x for the weighted sum of the objectives
assert(cachedData.bMinMax.capacity() >= epochModel.epochMatrix.getRowCount());
assert(cachedData.xMinMax.size() == epochModel.epochMatrix.getRowGroupCount());
cachedData.bMinMax.assign(epochModel.epochMatrix.getRowCount(), storm::utility::zero<ValueType>());
for (uint64_t objIndex = 0; objIndex < this->objectives.size(); ++objIndex) {
ValueType weight = storm::solver::minimize(this->objectives[objIndex].formula->getOptimalityType()) ? -weightVector[objIndex] : weightVector[objIndex];
if (!storm::utility::isZero(weight)) {
std::vector<ValueType> const& objectiveReward = epochModel.objectiveRewards[objIndex];
for (auto const& choice : epochModel.objectiveRewardFilter[objIndex]) {
cachedData.bMinMax[choice] += weight * objectiveReward[choice];
}
}
}
auto stepSolutionIt = epochModel.stepSolutions.begin();
for (auto const& choice : epochModel.stepChoices) {
cachedData.bMinMax[choice] += stepSolutionIt->front();
++stepSolutionIt;
}
// Invoke the min max solver
cachedData.minMaxSolver->solveEquations(cachedData.xMinMax, cachedData.bMinMax);
for (auto const& state : epochModel.epochInStates) {
result.emplace_back();
result.back().reserve(solutionSize);
result.back().push_back(cachedData.xMinMax[state]);
}
// Check whether the linear equation solver needs to be updated
auto const& choices = cachedData.minMaxSolver->getSchedulerChoices();
if (cachedData.schedulerChoices != choices) {
std::vector<uint64_t> choicesTmp = choices;
cachedData.minMaxSolver->setInitialScheduler(std::move(choicesTmp));
++numSchedChanges;
cachedData.schedulerChoices = choices;
storm::solver::GeneralLinearEquationSolverFactory<ValueType> linEqSolverFactory;
bool needEquationSystem = linEqSolverFactory.getEquationProblemFormat() == storm::solver::LinearEquationSolverProblemFormat::EquationSystem;
storm::storage::SparseMatrix<ValueType> subMatrix = epochModel.epochMatrix.selectRowsFromRowGroups(choices, needEquationSystem);
if (needEquationSystem) {
subMatrix.convertToEquationSystem();
}
cachedData.linEqSolver = linEqSolverFactory.create(std::move(subMatrix));
cachedData.linEqSolver->setCachingEnabled(true);
}
// Formulate for each objective the linear equation system induced by the performed choices
assert(cachedData.bLinEq.size() == choices.size());
for (uint64_t objIndex = 0; objIndex < this->objectives.size(); ++objIndex) {
auto const& obj = this->objectives[objIndex];
std::vector<ValueType> const& objectiveReward = epochModel.objectiveRewards[objIndex];
auto rowGroupIndexIt = epochModel.epochMatrix.getRowGroupIndices().begin();
auto choiceIt = choices.begin();
auto stepChoiceIt = epochModel.stepChoices.begin();
auto stepSolutionIt = epochModel.stepSolutions.begin();
std::vector<ValueType>& x = cachedData.xLinEq[objIndex];
auto xIt = x.begin();
for (auto& b_i : cachedData.bLinEq) {
uint64_t i = *rowGroupIndexIt + *choiceIt;
if (epochModel.objectiveRewardFilter[objIndex].get(i)) {
b_i = objectiveReward[i];
} else {
b_i = storm::utility::zero<ValueType>();
}
while (*stepChoiceIt < i) {
++stepChoiceIt;
++stepSolutionIt;
}
if (i == *stepChoiceIt) {
b_i += (*stepSolutionIt)[objIndex + 1];
++stepChoiceIt;
++stepSolutionIt;
}
// We can already set x_i correctly if row i is empty.
// Appearingly, some linear equation solvers struggle to converge otherwise.
if (epochModel.epochMatrix.getRow(i).getNumberOfEntries() == 0) {
*xIt = b_i;
}
++xIt;
++rowGroupIndexIt;
++choiceIt;
}
assert(x.size() == choices.size());
auto req = cachedData.linEqSolver->getRequirements();
cachedData.linEqSolver->clearBounds();
if (obj.lowerResultBound) {
req.clearLowerBounds();
cachedData.linEqSolver->setLowerBound(*obj.lowerResultBound);
}
if (obj.upperResultBound) {
cachedData.linEqSolver->setUpperBound(*obj.upperResultBound);
req.clearUpperBounds();
}
STORM_LOG_THROW(req.empty(), storm::exceptions::UncheckedRequirementException, "At least one requirement of the LinearEquationSolver was not met.");
cachedData.linEqSolver->solveEquations(x, cachedData.bLinEq);
auto resultIt = result.begin();
for (auto const& state : epochModel.epochInStates) {
resultIt->push_back(x[state]);
++resultIt;
}
}
}
rewardUnfolding.setSolutionForCurrentEpoch(std::move(result));
swEpochModelAnalysis.stop();
}
template <class SparseMdpModelType>
void SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::updateCachedData(typename MultiDimensionalRewardUnfolding<ValueType, false>::EpochModel const& epochModel, EpochCheckingData& cachedData, std::vector<ValueType> const& weightVector) {
if (epochModel.epochMatrixChanged) {
// Update the cached MinMaxSolver data
cachedData.bMinMax.resize(epochModel.epochMatrix.getRowCount());
cachedData.xMinMax.assign(epochModel.epochMatrix.getRowGroupCount(), storm::utility::zero<ValueType>());
storm::solver::GeneralMinMaxLinearEquationSolverFactory<ValueType> minMaxSolverFactory;
cachedData.minMaxSolver = minMaxSolverFactory.create(epochModel.epochMatrix);
cachedData.minMaxSolver->setTrackScheduler(true);
cachedData.minMaxSolver->setCachingEnabled(true);
auto req = cachedData.minMaxSolver->getRequirements(storm::solver::EquationSystemType::StochasticShortestPath);
req.clearNoEndComponents();
boost::optional<ValueType> lowerBound = this->computeWeightedResultBound(true, weightVector, storm::storage::BitVector(weightVector.size(), true));
if (lowerBound) {
cachedData.minMaxSolver->setLowerBound(lowerBound.get());
req.clearLowerBounds();
}
boost::optional<ValueType> upperBound = this->computeWeightedResultBound(false, weightVector, storm::storage::BitVector(weightVector.size(), true));
if (upperBound) {
cachedData.minMaxSolver->setUpperBound(upperBound.get());
req.clearUpperBounds();
}
STORM_LOG_THROW(req.empty(), storm::exceptions::UncheckedRequirementException, "At least one requirement of the MinMaxSolver was not met.");
cachedData.minMaxSolver->setRequirementsChecked(true);
cachedData.minMaxSolver->setOptimizationDirection(storm::solver::OptimizationDirection::Maximize);
// Clear the scheduler choices so that an update of the linEqSolver is enforced
cachedData.schedulerChoices.clear();
cachedData.schedulerChoices.reserve(epochModel.epochMatrix.getRowGroupCount());
// Update data for linear equation solving
cachedData.bLinEq.resize(epochModel.epochMatrix.getRowGroupCount());
cachedData.xLinEq.resize(this->objectives.size());
for (auto& x_o : cachedData.xLinEq) {
x_o.assign(epochModel.epochMatrix.getRowGroupCount(), storm::utility::zero<ValueType>());
}
}
}
template <class SparseMdpModelType>
std::vector<typename SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::ValueType> SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::getUnderApproximationOfInitialStateResults() const {
STORM_LOG_THROW(underApproxResult, storm::exceptions::InvalidOperationException, "Tried to retrieve results but check(..) has not been called before.");
return underApproxResult.get();
}
template <class SparseMdpModelType>
std::vector<typename SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::ValueType> SparseMdpRewardBoundedPcaaWeightVectorChecker<SparseMdpModelType>::getOverApproximationOfInitialStateResults() const {
STORM_LOG_THROW(overApproxResult, storm::exceptions::InvalidOperationException, "Tried to retrieve results but check(..) has not been called before.");
return overApproxResult.get();
}
template class SparseMdpRewardBoundedPcaaWeightVectorChecker<storm::models::sparse::Mdp<double>>;
#ifdef STORM_HAVE_CARL
template class SparseMdpRewardBoundedPcaaWeightVectorChecker<storm::models::sparse::Mdp<storm::RationalNumber>>;
#endif
}
}
}