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

Nondeterministic Infinite horizon: Split value getters into StateValueGetter and ActionValueGetters. Made VI code more general, so that they may also be used for Markov Automata.

main
Tim Quatmann 5 years ago
parent
31dd1d8f49
commit
fc66e01ed5
2 changed files with 313 additions and 141 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. 438
      src/storm/modelchecker/helper/infinitehorizon/SparseNondeterministicInfiniteHorizonHelper.cpp
  2. 16
      src/storm/modelchecker/helper/infinitehorizon/SparseNondeterministicInfiniteHorizonHelper.h

438
src/storm/modelchecker/helper/infinitehorizon/SparseNondeterministicInfiniteHorizonHelper.cpp
View File

@ -31,36 +31,56 @@ namespace storm {
template <typename ValueType>
std::vector<ValueType> SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLongRunAverageProbabilities(Environment const& env, storm::storage::BitVector const& psiStates) {
return computeLongRunAverageValues(env, [&psiStates] (uint64_t stateIndex, uint64_t) { return psiStates.get(stateIndex) ? storm::utility::one<ValueType>() : storm::utility::zero<ValueType>();});
return computeLongRunAverageValues(env,
[&psiStates] (uint64_t stateIndex) { return psiStates.get(stateIndex) ? storm::utility::one<ValueType>() : storm::utility::zero<ValueType>(); },
[] (uint64_t) { return storm::utility::zero<ValueType>(); }
);
}
template <typename ValueType>
std::vector<ValueType> SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLongRunAverageRewards(Environment const& env, storm::models::sparse::StandardRewardModel<ValueType> const& rewardModel) {
if (_markovianStates) {
return computeLongRunAverageValues(env, [&] (uint64_t stateIndex, uint64_t globalChoiceIndex) {
if (rewardModel.hasStateRewards() && _markovianStates->get(stateIndex)) {
return rewardModel.getTotalStateActionReward(stateIndex, globalChoiceIndex, _transitionMatrix, (ValueType) (storm::utility::one<ValueType>() / (*_exitRates)[stateIndex]));
} else {
return rewardModel.getTotalStateActionReward(stateIndex, globalChoiceIndex, _transitionMatrix, storm::utility::zero<ValueType>());
}
});
std::function<ValueType(uint64_t stateIndex)> stateRewardsGetter;
if (rewardModel.hasStateRewards()) {
stateRewardsGetter = [&rewardModel] (uint64_t stateIndex) { return rewardModel.getStateReward(stateIndex); };
} else {
stateRewardsGetter = [] (uint64_t) { return storm::utility::zero<ValueType>(); };
}
std::function<ValueType(uint64_t globalChoiceIndex)> actionRewardsGetter;
if (rewardModel.hasStateActionRewards() || rewardModel.hasTransitionRewards()) {
if (rewardModel.hasTransitionRewards()) {
actionRewardsGetter = [&] (uint64_t globalChoiceIndex) { return rewardModel.getStateActionAndTransitionReward(globalChoiceIndex, this->_transitionMatrix); };
} else {
actionRewardsGetter = [&] (uint64_t globalChoiceIndex) { return rewardModel.getStateActionReward(globalChoiceIndex); };
}
} else {
return computeLongRunAverageValues(env, [&] (uint64_t stateIndex, uint64_t globalChoiceIndex) {
return rewardModel.getTotalStateActionReward(stateIndex, globalChoiceIndex, _transitionMatrix);
});
stateRewardsGetter = [] (uint64_t) { return storm::utility::zero<ValueType>(); };
}
return computeLongRunAverageValues(env, stateRewardsGetter, actionRewardsGetter);
}
template <typename ValueType>
std::vector<ValueType> SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLongRunAverageValues(Environment const& env, std::vector<ValueType> const& combinedStateActionRewards) {
return computeLongRunAverageValues(env, [&combinedStateActionRewards] (uint64_t, uint64_t globalChoiceIndex) {
return combinedStateActionRewards[globalChoiceIndex];
});
std::vector<ValueType> SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLongRunAverageValues(Environment const& env, std::vector<ValueType> const* stateValues, std::vector<ValueType> const* actionValues) {
std::function<ValueType(uint64_t stateIndex)> stateValuesGetter;
if (stateValues) {
stateValuesGetter = [&stateValues] (uint64_t stateIndex) { return (*stateValues)[stateIndex]; };
} else {
stateValuesGetter = [] (uint64_t) { return storm::utility::zero<ValueType>(); };
}
std::function<ValueType(uint64_t actionIndex)> actionValuesGetter;
if (actionValues) {
actionValuesGetter = [&actionValues] (uint64_t globalChoiceIndex) { return (*actionValues)[globalChoiceIndex]; };
} else {
actionValuesGetter = [] (uint64_t) { return storm::utility::zero<ValueType>(); };
}
return computeLongRunAverageValues(env, stateValuesGetter, actionValuesGetter);
}
template <typename ValueType>
std::vector<ValueType> SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLongRunAverageValues(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter) {
std::vector<ValueType> SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLongRunAverageValues(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateRewardsGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionRewardsGetter) {
// Prepare an environment for the underlying solvers
auto underlyingSolverEnvironment = env;
@ -86,7 +106,7 @@ namespace storm {
std::vector<ValueType> mecLraValues;
mecLraValues.reserve(mecDecomposition.size());
for (auto const& mec : mecDecomposition) {
mecLraValues.push_back(computeLraForMec(underlyingSolverEnvironment, combinedStateActionRewardsGetter, mec));
mecLraValues.push_back(computeLraForMec(underlyingSolverEnvironment, stateRewardsGetter, actionRewardsGetter, mec));
}
// Solve the resulting SSP where end components are collapsed into single auxiliary states
@ -129,36 +149,45 @@ namespace storm {
}
template <typename ValueType>
ValueType SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLraForMec(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter, storm::storage::MaximalEndComponent const& mec) {
bool SparseNondeterministicInfiniteHorizonHelper<ValueType>::isContinuousTime() const {
STORM_LOG_ASSERT((_markovianStates == nullptr) == (_exitRates == nullptr), "Inconsistent information given: Have Markovian states but no exit rates (or vice versa)." );
return _markovianStates != nullptr;
}
template <typename ValueType>
ValueType SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLraForMec(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateRewardsGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionRewardsGetter, storm::storage::MaximalEndComponent const& mec) {
// FIXME: MA
// If the mec only consists of a single state, we compute the LRA value directly
if (++mec.begin() == mec.end()) {
if (mec.size() == 1) {
uint64_t state = mec.begin()->first;
auto choiceIt = mec.begin()->second.begin();
ValueType result = combinedStateActionRewardsGetter(state, *choiceIt);
uint64_t bestChoice = *choiceIt;
for (++choiceIt; choiceIt != mec.begin()->second.end(); ++choiceIt) {
ValueType choiceValue = combinedStateActionRewardsGetter(state, *choiceIt);
if (this->minimize()) {
if (result > choiceValue) {
result = std::move(choiceValue);
if (isContinuousTime()) {
// Singleton MECs have to consist of a Markovian state because of the non-Zenoness assumption. Then, there is just one possible choice.
STORM_LOG_THROW(_markovianStates->get(state), storm::exceptions::InvalidOperationException, "Markov Automaton has Zeno behavior. Computation of Long Run Average values not supported.");
STORM_LOG_ASSERT(mec.begin()->second.size() == 1, "Markovian state has Nondeterministic behavior.");
if (isProduceSchedulerSet()) {
_producedOptimalChoices.get()[state] = 0;
}
return stateRewardsGetter(state) + (*_exitRates)[state] * actionRewardsGetter(*choiceIt);
} else {
// Find the choice with the highest/lowest reward
ValueType bestValue = actionRewardsGetter(*choiceIt);
uint64_t bestChoice = *choiceIt;
for (++choiceIt; choiceIt != mec.begin()->second.end(); ++choiceIt) {
ValueType currentValue = actionRewardsGetter(*choiceIt);
if ((this->minimize() && currentValue < bestValue) || (this->maximize() && currentValue > bestValue)) {
bestValue = std::move(currentValue);
bestChoice = *choiceIt;
}
} else {
if (result < choiceValue) {
result = std::move(choiceValue);
bestChoice = *choiceIt;
}
}
if (isProduceSchedulerSet()) {
_producedOptimalChoices.get()[state] = bestChoice - _transitionMatrix.getRowGroupIndices()[state];
}
return bestValue + stateRewardsGetter(state);
}
if (isProduceSchedulerSet()) {
_producedOptimalChoices.get()[state] = bestChoice - _transitionMatrix.getRowGroupIndices()[state];
}
return result;
}
// Solve MEC with the method specified in the settings
// Solve nontrivial MEC with the method specified in the settings
storm::solver::LraMethod method = env.solver().lra().getNondetLraMethod();
if ((storm::NumberTraits<ValueType>::IsExact || env.solver().isForceExact()) && env.solver().lra().isNondetLraMethodSetFromDefault() && method != storm::solver::LraMethod::LinearProgramming) {
STORM_LOG_INFO("Selecting 'LP' as the solution technique for long-run properties to guarantee exact results. If you want to override this, please explicitly specify a different LRA method.");
@ -169,144 +198,281 @@ namespace storm {
}
STORM_LOG_ERROR_COND(!isProduceSchedulerSet() || method == storm::solver::LraMethod::ValueIteration, "Scheduler generation not supported for the chosen LRA method. Try value-iteration.");
if (method == storm::solver::LraMethod::LinearProgramming) {
return computeLraForMecLp(env, combinedStateActionRewardsGetter, mec);
return computeLraForMecLp(env, stateRewardsGetter, actionRewardsGetter, mec);
} else if (method == storm::solver::LraMethod::ValueIteration) {
return computeLraForMecVi(env, combinedStateActionRewardsGetter, mec);
return computeLraForMecVi(env, stateRewardsGetter, actionRewardsGetter, mec);
} else {
STORM_LOG_THROW(false, storm::exceptions::InvalidSettingsException, "Unsupported technique.");
}
}
/*!
* Abstract helper class that performs a single iteration of the value iteration method
*/
template <typename ValueType>
ValueType SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLraForMecVi(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter, storm::storage::MaximalEndComponent const& mec) {
// Initialize data about the mec
storm::storage::BitVector mecStates(_transitionMatrix.getRowGroupCount(), false);
storm::storage::BitVector mecChoices(_transitionMatrix.getRowCount(), false);
for (auto const& stateChoicesPair : mec) {
mecStates.set(stateChoicesPair.first);
for (auto const& choice : stateChoicesPair.second) {
mecChoices.set(choice);
class LraViHelper {
public:
LraViHelper(storm::storage::MaximalEndComponent const& mec, storm::storage::SparseMatrix<ValueType> const& transitionMatrix) : _mec(mec), _transitionMatrix(transitionMatrix) {
// Intentionally left empty
}
virtual ~LraViHelper() = default;
/*!
* performs a single iteration step.
* If a choices vector is given, the optimal choices will be inserted at the appropriate states.
* Note that these choices will be inserted w.r.t. the original model states/choices, i.e. the size of the vector should match the state-count of the input model
* @return the current estimate of the LRA value
*/
virtual void iterate(Environment const& env, storm::solver::OptimizationDirection const& dir, std::vector<uint64_t>* choices = nullptr) = 0;
struct ConvergenceCheckResult {
bool isPrecisionAchieved;
ValueType currentValue;
};
/*!
* Checks whether the curently computed value achieves the desired precision
*/
virtual ConvergenceCheckResult checkConvergence(bool relative, ValueType precision) = 0;
/*!
* Must be called between two calls of iterate.
*/
virtual void prepareNextIteration() = 0;
protected:
/*!
*
* @param xPrevious the 'old' values
* @param xCurrent the 'new' values
* @param threshold the threshold
* @param relative whether the relative difference should be considered
* @return The first component is true if the (relative) difference between the maximal and the minimal entry-wise change of the two value vectors is below or equal to the provided threshold.
* In this case, the second component is the average of the maximal and the minimal change.
* If the threshold is exceeded, the computation is aborted early and the second component is only an approximation of the averages.
*/
std::pair<bool, ValueType> checkMinMaxDiffBelowThreshold(std::vector<ValueType> const& xPrevious, std::vector<ValueType> const& xCurrent, ValueType const& threshold, bool relative) const {
STORM_LOG_ASSERT(xPrevious.size() == xCurrent.size(), "Unexpected Dimension Mismatch");
STORM_LOG_ASSERT(threshold > storm::utility::zero<ValueType>(), "Did not expect a non-positive threshold.");
auto x1It = xPrevious.begin();
auto x1Ite = xPrevious.end();
auto x2It = xCurrent.begin();
ValueType maxDiff = (*x2It - *x1It);
ValueType minDiff = maxDiff;
bool result = true;
// The difference between maxDiff and minDiff is zero at this point. Thus, it doesn't make sense to check the threshold now.
for (++x1It, ++x2It; x1It != x1Ite; ++x1It, ++x2It) {
ValueType diff = (*x2It - *x1It);
// Potentially update maxDiff or minDiff
bool skipCheck = false;
if (maxDiff < diff) {
maxDiff = diff;
} else if (minDiff > diff) {
minDiff = diff;
} else {
skipCheck = true;
}
// Check convergence
if (!skipCheck && (maxDiff - minDiff) > (relative ? (threshold * minDiff) : threshold)) {
result = false;
break;
}
}
ValueType avgDiff = (maxDiff + minDiff) / (storm::utility::convertNumber<ValueType>(2.0));
return {result, avgDiff};
}
boost::container::flat_map<uint64_t, uint64_t> toSubModelStateMapping;
uint64_t currState = 0;
toSubModelStateMapping.reserve(mecStates.getNumberOfSetBits());
for (auto const& mecState : mecStates) {
toSubModelStateMapping.insert(std::pair<uint64_t, uint64_t>(mecState, currState));
++currState;
}
// Get a transition matrix that only considers the states and choices within the MEC
storm::storage::SparseMatrixBuilder<ValueType> mecTransitionBuilder(mecChoices.getNumberOfSetBits(), mecStates.getNumberOfSetBits(), 0, true, true, mecStates.getNumberOfSetBits());
std::vector<ValueType> choiceValues;
choiceValues.reserve(mecChoices.getNumberOfSetBits());
uint64_t currRow = 0;
ValueType selfLoopProb = storm::utility::convertNumber<ValueType>(env.solver().lra().getAperiodicFactor());
ValueType scalingFactor = storm::utility::one<ValueType>() - selfLoopProb;
for (auto const& mecState : mecStates) {
mecTransitionBuilder.newRowGroup(currRow);
uint64_t groupStart = _transitionMatrix.getRowGroupIndices()[mecState];
uint64_t groupEnd = _transitionMatrix.getRowGroupIndices()[mecState + 1];
for (uint64_t choice = mecChoices.getNextSetIndex(groupStart); choice < groupEnd; choice = mecChoices.getNextSetIndex(choice + 1)) {
bool insertedDiagElement = false;
for (auto const& entry : _transitionMatrix.getRow(choice)) {
uint64_t column = toSubModelStateMapping[entry.getColumn()];
if (!insertedDiagElement && entry.getColumn() > mecState) {
mecTransitionBuilder.addNextValue(currRow, toSubModelStateMapping[mecState], selfLoopProb);
insertedDiagElement = true;
storm::storage::MaximalEndComponent const& _mec;
storm::storage::SparseMatrix<ValueType> const& _transitionMatrix;
};
/*!
* Abstract helper class that performs a single iteration of the value iteration method
* @see Ashok et al.: Value Iteration for Long-Run Average Reward in Markov Decision Processes (CAV'17), https://doi.org/10.1007/978-3-319-63387-9_10
*/
template <typename ValueType>
class MdpLraViHelper : public LraViHelper<ValueType> {
public:
MdpLraViHelper(storm::storage::MaximalEndComponent const& mec, storm::storage::SparseMatrix<ValueType> const& transitionMatrix, std::function<ValueType(uint64_t stateIndex)> const& stateRewardsGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionRewardsGetter, ValueType const& aperiodicFactor) : LraViHelper<ValueType>(mec, transitionMatrix), _x1(mec.size(), storm::utility::zero<ValueType>()), _x2(_x1), _x1IsCurrent(true) {
// We add a selfloop to each state (which is necessary for convergence)
// Very roughly, this selfloop avoids that the values can flip around like this: [1, 0] -> [0, 1] -> [1, 0] -> ...
ValueType selfLoopProb = aperiodicFactor;
// Introducing the selfloop also requires the rewards to be scaled by the following factor.
_scalingFactor = storm::utility::one<ValueType>() - selfLoopProb;
uint64_t numMecStates = this->_mec.size();
boost::container::flat_map<uint64_t, uint64_t> toSubModelStateMapping;
toSubModelStateMapping.reserve(this->_mec.size());
uint64_t currState = 0;
uint64_t numMecChoices = 0;
for (auto const& stateChoices : this->_mec) {
toSubModelStateMapping.insert(std::pair<uint64_t, uint64_t>(stateChoices.first, currState));
++currState;
numMecChoices += stateChoices.second.size();
}
assert(currState == numMecStates);
// Get a transition matrix that only considers the states and choices within the MEC
storm::storage::SparseMatrixBuilder<ValueType> mecTransitionBuilder(numMecChoices, numMecStates, 0, true, true, numMecStates);
_choiceValues.reserve(numMecChoices);
uint64_t currRow = 0;
for (auto const& stateChoices : this->_mec) {
auto const& mecState = stateChoices.first;
auto const& mecChoices = stateChoices.second;
mecTransitionBuilder.newRowGroup(currRow);
for (auto const& choice : mecChoices) {
bool insertedDiagElement = false;
for (auto const& entry : this->_transitionMatrix.getRow(choice)) {
uint64_t column = toSubModelStateMapping[entry.getColumn()];
if (!insertedDiagElement && entry.getColumn() > mecState) {
mecTransitionBuilder.addNextValue(currRow, toSubModelStateMapping[mecState], selfLoopProb);
insertedDiagElement = true;
}
if (!insertedDiagElement && entry.getColumn() == mecState) {
mecTransitionBuilder.addNextValue(currRow, column, selfLoopProb + _scalingFactor * entry.getValue());
insertedDiagElement = true;
} else {
mecTransitionBuilder.addNextValue(currRow, column, _scalingFactor * entry.getValue());
}
}
if (!insertedDiagElement && entry.getColumn() == mecState) {
mecTransitionBuilder.addNextValue(currRow, column, selfLoopProb + scalingFactor * entry.getValue());
insertedDiagElement = true;
} else {
mecTransitionBuilder.addNextValue(currRow, column, scalingFactor * entry.getValue());
if (!insertedDiagElement) {
mecTransitionBuilder.addNextValue(currRow, toSubModelStateMapping[mecState], selfLoopProb);
}
// Compute the rewards obtained for this choice
_choiceValues.push_back(_scalingFactor * (stateRewardsGetter(mecState) + actionRewardsGetter(choice)));
++currRow;
}
if (!insertedDiagElement) {
mecTransitionBuilder.addNextValue(currRow, toSubModelStateMapping[mecState], selfLoopProb);
}
_mecTransitions = mecTransitionBuilder.build();
STORM_LOG_ASSERT(_mecTransitions.isProbabilistic(), "The MEC-Matrix is not probabilistic.");
STORM_LOG_ASSERT(_mecTransitions.getRowGroupCount() == _x1.size(), "Unexpected size mismatch for created matrix.");
STORM_LOG_ASSERT(_x1.size() == _x2.size(), "Unexpected size mismatch for created matrix.");
}
virtual void iterate(Environment const& env, storm::solver::OptimizationDirection const& dir, std::vector<uint64_t>* choices = nullptr) override {
// Initialize a multipler if it does not exist, yet
if (!_multiplier) {
_multiplier = storm::solver::MultiplierFactory<ValueType>().create(env, _mecTransitions);
}
if (choices == nullptr) {
// Perform a simple matrix-vector multiplication
_multiplier->multiplyAndReduce(env, dir, xCurrent(), &_choiceValues, xPrevious());
} else {
// Perform a simple matrix-vector multiplication but also keep track of the choices within the _mecTransitions
std::vector<uint64_t> mecChoices(_mecTransitions.getRowGroupCount());
_multiplier->multiplyAndReduce(env, dir, xCurrent(), &_choiceValues, xPrevious(), &mecChoices);
// Transform the local choices (within this mec) to global indices
uint64_t mecState = 0;
for (auto const& stateChoices : this->_mec) {
uint64_t mecChoice = mecChoices[mecState];
STORM_LOG_ASSERT(mecChoice < stateChoices.second.size(), "The selected choice does not seem to exist.");
uint64_t globalChoiceIndex = *(stateChoices.second.begin() + mecChoice);
(*choices)[stateChoices.first] = globalChoiceIndex - this->_transitionMatrix.getRowGroupIndices()[stateChoices.first];
++mecState;
}
// Compute the rewards obtained for this choice
choiceValues.push_back(scalingFactor * combinedStateActionRewardsGetter(mecState, choice));
++currRow;
}
// Swap current and previous x vectors
_x1IsCurrent = !_x1IsCurrent;
}
auto mecTransitions = mecTransitionBuilder.build();
STORM_LOG_ASSERT(mecTransitions.isProbabilistic(), "The MEC-Matrix is not probabilistic.");
// start the iterations
ValueType precision = storm::utility::convertNumber<ValueType>(env.solver().lra().getPrecision()) / scalingFactor;
bool relative = env.solver().lra().getRelativeTerminationCriterion();
std::vector<ValueType> x(mecTransitions.getRowGroupCount(), storm::utility::zero<ValueType>());
std::vector<ValueType> xPrime = x;
auto dir = this->getOptimizationDirection();
virtual typename LraViHelper<ValueType>::ConvergenceCheckResult checkConvergence(bool relative, ValueType precision) override {
typename LraViHelper<ValueType>::ConvergenceCheckResult res;
std::tie(res.isPrecisionAchieved, res.currentValue) = this->checkMinMaxDiffBelowThreshold(xPrevious(), xCurrent(), precision, relative);
res.currentValue /= _scalingFactor; // "Undo" the scaling of the rewards
return res;
}
auto multiplier = storm::solver::MultiplierFactory<ValueType>().create(env, mecTransitions);
ValueType maxDiff, minDiff;
virtual void prepareNextIteration() override {
// To avoid large (and numerically unstable) x-values, we substract a reference value.
ValueType referenceValue = xCurrent().front();
storm::utility::vector::applyPointwise<ValueType, ValueType>(xCurrent(), xCurrent(), [&referenceValue] (ValueType const& x_i) -> ValueType { return x_i - referenceValue; });
}
uint64_t iter = 0;
private:
std::vector<ValueType>& xCurrent() {
return _x1IsCurrent ? _x1 : _x2;
}
std::vector<ValueType>& xPrevious() {
return _x1IsCurrent ? _x2 : _x1;
}
storm::storage::SparseMatrix<ValueType> _mecTransitions;
std::vector<ValueType> _x1, _x2, _choiceValues;
bool _x1IsCurrent;
std::unique_ptr<storm::solver::Multiplier<ValueType>> _multiplier;
ValueType _scalingFactor;
};
template <typename ValueType>
ValueType SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLraForMecVi(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateRewardsGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionRewardsGetter, storm::storage::MaximalEndComponent const& mec) {
// Collect some parameters of the computation.
ValueType aperiodicFactor = storm::utility::convertNumber<ValueType>(env.solver().lra().getAperiodicFactor());
ValueType precision = storm::utility::convertNumber<ValueType>(env.solver().lra().getPrecision()) / aperiodicFactor;
bool relative = env.solver().lra().getRelativeTerminationCriterion();
boost::optional<uint64_t> maxIter;
if (env.solver().lra().isMaximalIterationCountSet()) {
maxIter = env.solver().lra().getMaximalIterationCount();
}
auto dir = this->getOptimizationDirection();
// Create an object for the iterations
std::shared_ptr<LraViHelper<ValueType>> iterationHelper;
if (isContinuousTime()) {
// TODO
} else {
iterationHelper = std::make_shared<MdpLraViHelper<ValueType>>(mec, _transitionMatrix, stateRewardsGetter, actionRewardsGetter, aperiodicFactor);
}
// start the iterations
ValueType result = storm::utility::zero<ValueType>();
uint64_t iter = 0;
while (!maxIter.is_initialized() || iter < maxIter.get()) {
++iter;
// Compute the obtained values for the next step
multiplier->multiplyAndReduce(env, dir, x, &choiceValues, x);
// update xPrime and check for convergence
// to avoid large (and numerically unstable) x-values, we substract a reference value.
auto xIt = x.begin();
auto xPrimeIt = xPrime.begin();
ValueType refVal = *xIt;
maxDiff = *xIt - *xPrimeIt;
minDiff = maxDiff;
*xIt -= refVal;
*xPrimeIt = *xIt;
for (++xIt, ++xPrimeIt; xIt != x.end(); ++xIt, ++xPrimeIt) {
ValueType diff = *xIt - *xPrimeIt;
maxDiff = std::max(maxDiff, diff);
minDiff = std::min(minDiff, diff);
*xIt -= refVal;
*xPrimeIt = *xIt;
}
if ((maxDiff - minDiff) <= (relative ? (precision * minDiff) : precision)) {
iterationHelper->iterate(env, dir);
// Check if we are done
auto convergenceCheckResult = iterationHelper->checkConvergence(relative, precision);
result = convergenceCheckResult.currentValue;
if (convergenceCheckResult.isPrecisionAchieved) {
break;
}
if (storm::utility::resources::isTerminate()) {
break;
}
iterationHelper->prepareNextIteration();
}
if (maxIter.is_initialized() && iter == maxIter.get()) {
STORM_LOG_WARN("LRA computation did not converge within " << iter << " iterations.");
} else if (storm::utility::resources::isTerminate()) {
STORM_LOG_WARN("LRA computation aborted after " << iter << " iterations.");
} else {
STORM_LOG_TRACE("LRA computation converged after " << iter << " iterations.");
}
if (isProduceSchedulerSet()) {
std::vector<uint_fast64_t> localMecChoices(mecTransitions.getRowGroupCount(), 0);
multiplier->multiplyAndReduce(env, dir, x, &choiceValues, x, &localMecChoices);
auto localMecChoiceIt = localMecChoices.begin();
for (auto const& mecState : mecStates) {
// Get the choice index of the selected mec choice with respect to the global transition matrix.
uint_fast64_t globalChoice = mecChoices.getNextSetIndex(_transitionMatrix.getRowGroupIndices()[mecState]);
for (uint_fast64_t i = 0; i < *localMecChoiceIt; ++i) {
globalChoice = mecChoices.getNextSetIndex(globalChoice + 1);
}
STORM_LOG_ASSERT(globalChoice < _transitionMatrix.getRowGroupIndices()[mecState + 1], "Invalid global choice for mec state.");
_producedOptimalChoices.get()[mecState] = globalChoice - _transitionMatrix.getRowGroupIndices()[mecState];
++localMecChoiceIt;
}
// We will be doing one more iteration step and track scheduler choices this time.
iterationHelper->prepareNextIteration();
iterationHelper->iterate(env, dir, &_producedOptimalChoices.get());
}
return (maxDiff + minDiff) / (storm::utility::convertNumber<ValueType>(2.0) * scalingFactor);
return result;
}
template <typename ValueType>
ValueType SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLraForMecLp(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter, storm::storage::MaximalEndComponent const& mec) {
ValueType SparseNondeterministicInfiniteHorizonHelper<ValueType>::computeLraForMecLp(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateRewardsGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionRewardsGetter, storm::storage::MaximalEndComponent const& mec) {
std::shared_ptr<storm::solver::LpSolver<ValueType>> solver = storm::utility::solver::getLpSolver<ValueType>("LRA for MEC");
solver->setOptimizationDirection(invert(this->getOptimizationDirection()));
@ -330,7 +496,7 @@ namespace storm {
for (auto element : _transitionMatrix.getRow(choice)) {
constraint = constraint + stateToVariableMap.at(element.getColumn()) * solver->getConstant(element.getValue());
}
constraint = solver->getConstant(combinedStateActionRewardsGetter(state, choice)) + constraint;
constraint = solver->getConstant(stateRewardsGetter(state) + actionRewardsGetter(choice)) + constraint;
if (this->minimize()) {
constraint = stateToVariableMap.at(state) <= constraint;

16
src/storm/modelchecker/helper/infinitehorizon/SparseNondeterministicInfiniteHorizonHelper.h
View File

@ -44,13 +44,13 @@ namespace storm {
* Computes the long run average value given the provided action-based rewards
* @return a value for each state
*/
std::vector<ValueType> computeLongRunAverageValues(Environment const& env, std::vector<ValueType> const& combinedStateActionRewards);
std::vector<ValueType> computeLongRunAverageValues(Environment const& env, std::vector<ValueType> const* stateValues = nullptr, std::vector<ValueType> const* actionValues = nullptr);
/*!
* Computes the long run average value given the provided state-action-based rewards
* @return a value for each state
*/
std::vector<ValueType> computeLongRunAverageValues(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter);
std::vector<ValueType> computeLongRunAverageValues(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateValuesGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionValuesGetter);
/*!
* Sets whether an optimal scheduler shall be constructed during the computation
@ -81,21 +81,27 @@ namespace storm {
storm::storage::Scheduler<ValueType> extractScheduler() const;
protected:
/*!
* @return true iff this is a computation on a continuous time model (i.e. MA)
*/
bool isContinuousTime() const;
/*!
* @pre if scheduler production is enabled, the _producedOptimalChoices vector should be initialized and sufficiently large
* @return the (unique) optimal LRA value for the given mec.
* @post _producedOptimalChoices contains choices for the states of the given MEC which yield the returned LRA value.
*/
ValueType computeLraForMec(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter, storm::storage::MaximalEndComponent const& mec);
ValueType computeLraForMec(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateValuesGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionValuesGetter, storm::storage::MaximalEndComponent const& mec);
/*!
* As computeLraForMec but uses value iteration as a solution method (independent of what is set in env)
*/
ValueType computeLraForMecVi(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter, storm::storage::MaximalEndComponent const& mec);
ValueType computeLraForMecVi(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateValuesGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionValuesGetter, storm::storage::MaximalEndComponent const& mec);
/*!
* As computeLraForMec but uses linear programming as a solution method (independent of what is set in env)
*/
ValueType computeLraForMecLp(Environment const& env, std::function<ValueType(uint64_t stateIndex, uint64_t globalChoiceIndex)> const& combinedStateActionRewardsGetter, storm::storage::MaximalEndComponent const& mec);
ValueType computeLraForMecLp(Environment const& env, std::function<ValueType(uint64_t stateIndex)> const& stateValuesGetter, std::function<ValueType(uint64_t globalChoiceIndex)> const& actionValuesGetter, storm::storage::MaximalEndComponent const& mec);
/*!
* @return Lra values for each state

Write Preview
Loading…
Cancel
Save