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

Making sure that we only store the best bounds found so far. Also added some output for the resulting values in each iteration.

tempestpy_adaptions
Tim Quatmann 5 years ago
parent
fa624d2a20
commit
1763f0c582
2 changed files with 39 additions and 8 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. 44
      src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.cpp
  2. 3
      src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.h

44
src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.cpp
View File

@ -47,6 +47,24 @@ namespace storm {
return diff;
}
template<typename PomdpModelType, typename BeliefValueType>
bool ApproximatePOMDPModelchecker<PomdpModelType, BeliefValueType>::Result::updateLowerBound(ValueType const& value) {
if (value > lowerBound) {
lowerBound = value;
return true;
}
return false;
}
template<typename PomdpModelType, typename BeliefValueType>
bool ApproximatePOMDPModelchecker<PomdpModelType, BeliefValueType>::Result::updateUpperBound(ValueType const& value) {
if (value < upperBound) {
upperBound = value;
return true;
}
return false;
}
template<typename PomdpModelType, typename BeliefValueType>
ApproximatePOMDPModelchecker<PomdpModelType, BeliefValueType>::Statistics::Statistics() : overApproximationBuildAborted(false), underApproximationBuildAborted(false), aborted(false) {
// intentionally left empty;
@ -62,13 +80,15 @@ namespace storm {
STORM_LOG_ASSERT(options.unfold || options.discretize, "Invoked belief exploration but no task (unfold or discretize) given.");
// Reset all collected statistics
statistics = Statistics();
statistics.totalTime.start();
// Extract the relevant information from the formula
auto formulaInfo = storm::pomdp::analysis::getFormulaInformation(pomdp, formula);
// Compute some initial bounds on the values for each state of the pomdp
auto initialPomdpValueBounds = TrivialPomdpValueBoundsModelChecker<storm::models::sparse::Pomdp<ValueType>>(pomdp).getValueBounds(formula, formulaInfo);
Result result(initialPomdpValueBounds.lower[pomdp.getInitialStates().getNextSetIndex(0)], initialPomdpValueBounds.upper[pomdp.getInitialStates().getNextSetIndex(0)]);
STORM_PRINT_AND_LOG("Initial value bounds are [" << result.lowerBound << ", " << result.upperBound << "]" << std::endl);
boost::optional<std::string> rewardModelName;
if (formulaInfo.isNonNestedReachabilityProbability() || formulaInfo.isNonNestedExpectedRewardFormula()) {
// FIXME: Instead of giving up, introduce a new observation for target states and make sink states absorbing.
@ -103,6 +123,7 @@ namespace storm {
if (storm::utility::resources::isTerminate()) {
statistics.aborted = true;
}
statistics.totalTime.stop();
return result;
}
@ -117,6 +138,8 @@ namespace storm {
stream << "# Computation aborted early" << std::endl;
}
stream << "# Total check time: " << statistics.totalTime << std::endl;
// Refinement information:
if (statistics.refinementSteps) {
stream << "# Number of refinement steps: " << statistics.refinementSteps.get() << std::endl;
@ -208,9 +231,6 @@ namespace storm {
template<typename PomdpModelType, typename BeliefValueType>
void ApproximatePOMDPModelchecker<PomdpModelType, BeliefValueType>::refineReachability(std::set<uint32_t> const &targetObservations, bool min, boost::optional<std::string> rewardModelName, std::vector<ValueType> const& lowerPomdpValueBounds, std::vector<ValueType> const& upperPomdpValueBounds, Result& result) {
ValueType& overApproxValue = min ? result.lowerBound : result.upperBound;
ValueType& underApproxValue = min ? result.upperBound : result.lowerBound;
// Set up exploration data
std::vector<uint64_t> observationResolutionVector;
std::shared_ptr<BeliefManagerType> overApproxBeliefManager;
@ -231,7 +251,9 @@ namespace storm {
if (!overApproximation->hasComputedValues()) {
return;
}
overApproxValue = overApproximation->getComputedValueAtInitialState();
ValueType const& newValue = overApproximation->getComputedValueAtInitialState();
bool betterBound = min ? result.updateLowerBound(newValue) : result.updateUpperBound(newValue);
STORM_PRINT_AND_LOG("Over-approx result for refinement step #0 is '" << newValue << "' which " << std::string(betterBound ? "improves" : "does not improve") << " the old value. Current runtime is " << statistics.totalTime << " seconds." << std::endl);
}
std::shared_ptr<BeliefManagerType> underApproxBeliefManager;
@ -254,7 +276,9 @@ namespace storm {
if (!underApproximation->hasComputedValues()) {
return;
}
underApproxValue = underApproximation->getComputedValueAtInitialState();
ValueType const& newValue = underApproximation->getComputedValueAtInitialState();
bool betterBound = min ? result.updateUpperBound(newValue) : result.updateLowerBound(newValue);
STORM_PRINT_AND_LOG("Under-approx result for refinement step #0 is '" << newValue << "' which " << std::string(betterBound ? "improves" : "does not improve") << " the old value. Current runtime is " << statistics.totalTime << " seconds." << std::endl);
}
// Start refinement
@ -281,7 +305,9 @@ namespace storm {
overApproxHeuristicPar.optimalChoiceValueEpsilon *= options.optimalChoiceValueThresholdFactor;
buildOverApproximation(targetObservations, min, rewardModelName.is_initialized(), true, overApproxHeuristicPar, observationResolutionVector, overApproxBeliefManager, overApproximation);
if (overApproximation->hasComputedValues()) {
overApproxValue = overApproximation->getComputedValueAtInitialState();
ValueType const& newValue = overApproximation->getComputedValueAtInitialState();
bool betterBound = min ? result.updateLowerBound(newValue) : result.updateUpperBound(newValue);
STORM_PRINT_AND_LOG("Over-approx result for refinement step #" << statistics.refinementSteps.get() << " is '" << newValue << "' which " << std::string(betterBound ? "improves" : "does not improve") << " the old value. Current runtime is " << statistics.totalTime << " seconds." << std::endl);
} else {
break;
}
@ -294,7 +320,9 @@ namespace storm {
overApproxHeuristicPar.optimalChoiceValueEpsilon *= options.optimalChoiceValueThresholdFactor;
buildUnderApproximation(targetObservations, min, rewardModelName.is_initialized(), true, underApproxHeuristicPar, underApproxBeliefManager, underApproximation);
if (underApproximation->hasComputedValues()) {
underApproxValue = underApproximation->getComputedValueAtInitialState();
ValueType const& newValue = underApproximation->getComputedValueAtInitialState();
bool betterBound = min ? result.updateUpperBound(newValue) : result.updateLowerBound(newValue);
STORM_PRINT_AND_LOG("Under-approx result for refinement step #" << statistics.refinementSteps.get() << " is '" << newValue << "' which " << std::string(betterBound ? "improves" : "does not improve") << " the old value. Current runtime is " << statistics.totalTime << " seconds." << std::endl);
} else {
break;
}

3
src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.h
View File

@ -30,6 +30,8 @@ namespace storm {
ValueType lowerBound;
ValueType upperBound;
ValueType diff (bool relative = false) const;
bool updateLowerBound(ValueType const& value);
bool updateUpperBound(ValueType const& value);
};
ApproximatePOMDPModelchecker(PomdpModelType const& pomdp, Options options = Options());
@ -87,6 +89,7 @@ namespace storm {
struct Statistics {
Statistics();
boost::optional<uint64_t> refinementSteps;
storm::utility::Stopwatch totalTime;
boost::optional<uint64_t> overApproximationStates;
bool overApproximationBuildAborted;

Write Preview
Loading…
Cancel
Save