Browse Source

Added on-the-fly belief grid generation for rewards

tempestpy_adaptions
Alexander Bork 5 years ago
parent
commit
b9c0af6628
  1. 262
      src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.cpp
  2. 4
      src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.h

262
src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.cpp

@ -293,6 +293,268 @@ namespace storm {
} }
template<typename ValueType, typename RewardModelType>
std::unique_ptr<POMDPCheckResult<ValueType>>
ApproximatePOMDPModelchecker<ValueType, RewardModelType>::computeReachabilityRewardOTF(storm::models::sparse::Pomdp<ValueType, RewardModelType> const &pomdp,
std::set<uint32_t> targetObservations, bool min, uint64_t gridResolution) {
STORM_PRINT("Use On-The-Fly Grid Generation" << std::endl)
RewardModelType pomdpRewardModel = pomdp.getUniqueRewardModel();
uint64_t maxIterations = 100;
bool finished = false;
uint64_t iteration = 0;
std::vector<storm::pomdp::Belief<ValueType>> beliefList;
std::vector<bool> beliefIsTarget;
std::vector<storm::pomdp::Belief<ValueType>> beliefGrid;
std::map<uint64_t, ValueType> result;
std::map<uint64_t, ValueType> result_backup;
//Use caching to avoid multiple computation of the subsimplices and lambdas
std::map<uint64_t, std::vector<std::vector<ValueType>>> subSimplexCache;
std::map<uint64_t, std::vector<ValueType>> lambdaCache;
std::map<uint64_t, uint64_t> chosenActions;
std::deque<uint64_t> beliefsToBeExpanded;
// Belief ID -> Observation -> Probability
std::map<uint64_t, std::vector<std::map<uint32_t, ValueType>>> observationProbabilities;
// current ID -> action -> next ID
std::map<uint64_t, std::vector<std::map<uint32_t, uint64_t>>> nextBelieves;
// current ID -> action -> reward
std::map<uint64_t, std::vector<ValueType>> beliefActionRewards;
uint64_t nextId = 0;
storm::utility::Stopwatch expansionTimer(true);
// Initial belief always has ID 0
storm::pomdp::Belief<ValueType> initialBelief = getInitialBelief(pomdp, nextId);
++nextId;
beliefList.push_back(initialBelief);
beliefIsTarget.push_back(
targetObservations.find(initialBelief.observation) != targetObservations.end());
// for the initial belief, add the triangulated initial states
std::pair<std::vector<std::vector<ValueType>>, std::vector<ValueType>> initTemp = computeSubSimplexAndLambdas(
initialBelief.probabilities, gridResolution);
std::vector<std::vector<ValueType>> initSubSimplex = initTemp.first;
std::vector<ValueType> initLambdas = initTemp.second;
subSimplexCache[0] = initSubSimplex;
lambdaCache[0] = initLambdas;
bool initInserted = false;
for (size_t j = 0; j < initLambdas.size(); ++j) {
if (!cc.isEqual(initLambdas[j], storm::utility::zero<ValueType>())) {
uint64_t searchResult = getBeliefIdInVector(beliefList, initialBelief.observation, initSubSimplex[j]);
if (searchResult == uint64_t(-1) || (searchResult == 0 && !initInserted)) {
if (searchResult == 0) {
// the initial belief is on the grid itself
initInserted = true;
beliefGrid.push_back(initialBelief);
beliefsToBeExpanded.push_back(0);
} else {
// if the triangulated belief was not found in the list, we place it in the grid and add it to the work list
storm::pomdp::Belief<ValueType> gridBelief = {nextId, initialBelief.observation, initSubSimplex[j]};
beliefList.push_back(gridBelief);
beliefGrid.push_back(gridBelief);
beliefIsTarget.push_back(
targetObservations.find(initialBelief.observation) != targetObservations.end());
beliefsToBeExpanded.push_back(nextId);
++nextId;
}
}
}
}
//beliefsToBeExpanded.push_back(initialBelief.id); TODO I'm curious what happens if we do this instead of first triangulating. Should do nothing special if belief is on grid, otherwise it gets interesting
// Expand the beliefs to generate the grid on-the-fly to avoid unreachable grid points
while (!beliefsToBeExpanded.empty()) {
uint64_t currId = beliefsToBeExpanded.front();
beliefsToBeExpanded.pop_front();
bool isTarget = beliefIsTarget[currId];
result.emplace(std::make_pair(currId, storm::utility::zero<ValueType>()));
result_backup.emplace(std::make_pair(currId, storm::utility::zero<ValueType>()));
if (!isTarget) {
uint64_t representativeState = pomdp.getStatesWithObservation(beliefList[currId].observation).front();
uint64_t numChoices = pomdp.getNumberOfChoices(representativeState);
std::vector<std::map<uint32_t, ValueType>> observationProbabilitiesInAction(numChoices);
std::vector<std::map<uint32_t, uint64_t>> nextBelievesInAction(numChoices);
std::vector<ValueType> actionRewardsInState(numChoices);
for (uint64_t action = 0; action < numChoices; ++action) {
std::map<uint32_t, ValueType> actionObservationProbabilities = computeObservationProbabilitiesAfterAction(pomdp, beliefList[currId], action);
std::map<uint32_t, uint64_t> actionObservationBelieves;
for (auto iter = actionObservationProbabilities.begin(); iter != actionObservationProbabilities.end(); ++iter) {
uint32_t observation = iter->first;
// THIS CALL IS SLOW
// TODO speed this up
uint64_t idNextBelief = getBeliefAfterActionAndObservation(pomdp, beliefList, beliefIsTarget, targetObservations, beliefList[currId], action,
observation, nextId);
nextId = beliefList.size();
actionObservationBelieves[observation] = idNextBelief;
//Triangulate here and put the possibly resulting belief in the grid
std::vector<std::vector<ValueType>> subSimplex;
std::vector<ValueType> lambdas;
if (subSimplexCache.count(idNextBelief) > 0) {
// TODO is this necessary here? Think later
subSimplex = subSimplexCache[idNextBelief];
lambdas = lambdaCache[idNextBelief];
} else {
std::pair<std::vector<std::vector<ValueType>>, std::vector<ValueType>> temp = computeSubSimplexAndLambdas(
beliefList[idNextBelief].probabilities, gridResolution);
subSimplex = temp.first;
lambdas = temp.second;
subSimplexCache[idNextBelief] = subSimplex;
lambdaCache[idNextBelief] = lambdas;
}
for (size_t j = 0; j < lambdas.size(); ++j) {
if (!cc.isEqual(lambdas[j], storm::utility::zero<ValueType>())) {
if (getBeliefIdInVector(beliefGrid, observation, subSimplex[j]) == uint64_t(-1)) {
// if the triangulated belief was not found in the list, we place it in the grid and add it to the work list
storm::pomdp::Belief<ValueType> gridBelief = {nextId, observation, subSimplex[j]};
beliefList.push_back(gridBelief);
beliefGrid.push_back(gridBelief);
beliefIsTarget.push_back(
targetObservations.find(observation) != targetObservations.end());
beliefsToBeExpanded.push_back(nextId);
++nextId;
}
}
}
}
observationProbabilitiesInAction[action] = actionObservationProbabilities;
nextBelievesInAction[action] = actionObservationBelieves;
actionRewardsInState[action] = getRewardAfterAction(pomdp, pomdp.getChoiceIndex(storm::storage::StateActionPair(representativeState, action)),
beliefList[currId]);
}
observationProbabilities.emplace(
std::make_pair(currId, observationProbabilitiesInAction));
nextBelieves.emplace(std::make_pair(currId, nextBelievesInAction));
beliefActionRewards.emplace(std::make_pair(currId, actionRewardsInState));
}
}
expansionTimer.stop();
STORM_PRINT("Grid size: " << beliefGrid.size() << std::endl)
STORM_PRINT("#Believes in List: " << beliefList.size() << std::endl)
STORM_PRINT("Belief space expansion took " << expansionTimer << std::endl)
storm::utility::Stopwatch overApproxTimer(true);
// Value Iteration
while (!finished && iteration < maxIterations) {
storm::utility::Stopwatch iterationTimer(true);
STORM_LOG_DEBUG("Iteration " << iteration + 1);
bool improvement = false;
for (size_t i = 0; i < beliefGrid.size(); ++i) {
storm::pomdp::Belief<ValueType> currentBelief = beliefGrid[i];
bool isTarget = beliefIsTarget[currentBelief.id];
if (!isTarget) {
// we can take any state with the observation as they have the same number of choices
uint64_t numChoices = pomdp.getNumberOfChoices(
pomdp.getStatesWithObservation(currentBelief.observation).front());
// Initialize the values for the value iteration
ValueType chosenValue = min ? storm::utility::infinity<ValueType>()
: -storm::utility::infinity<ValueType>();
uint64_t chosenActionIndex = std::numeric_limits<uint64_t>::infinity();
ValueType currentValue;
for (uint64_t action = 0; action < numChoices; ++action) {
currentValue = beliefActionRewards[currentBelief.id][action];
for (auto iter = observationProbabilities[currentBelief.id][action].begin();
iter != observationProbabilities[currentBelief.id][action].end(); ++iter) {
uint32_t observation = iter->first;
storm::pomdp::Belief<ValueType> nextBelief = beliefList[nextBelieves[currentBelief.id][action][observation]];
// compute subsimplex and lambdas according to the Lovejoy paper to approximate the next belief
// cache the values to not always re-calculate
std::vector<std::vector<ValueType>> subSimplex;
std::vector<ValueType> lambdas;
if (subSimplexCache.count(nextBelief.id) > 0) {
subSimplex = subSimplexCache[nextBelief.id];
lambdas = lambdaCache[nextBelief.id];
} else {
//TODO This should not ne reachable
std::pair<std::vector<std::vector<ValueType>>, std::vector<ValueType>> temp = computeSubSimplexAndLambdas(
nextBelief.probabilities, gridResolution);
subSimplex = temp.first;
lambdas = temp.second;
subSimplexCache[nextBelief.id] = subSimplex;
lambdaCache[nextBelief.id] = lambdas;
}
auto sum = storm::utility::zero<ValueType>();
for (size_t j = 0; j < lambdas.size(); ++j) {
if (!cc.isEqual(lambdas[j], storm::utility::zero<ValueType>())) {
sum += lambdas[j] * result_backup.at(
getBeliefIdInVector(beliefGrid, observation, subSimplex[j]));
}
}
currentValue += iter->second * sum;
}
// Update the selected actions
if ((min && cc.isLess(storm::utility::zero<ValueType>(), chosenValue - currentValue)) ||
(!min &&
cc.isLess(storm::utility::zero<ValueType>(), currentValue - chosenValue)) ||
cc.isEqual(storm::utility::zero<ValueType>(), chosenValue - currentValue)) {
chosenValue = currentValue;
chosenActionIndex = action;
}
}
result[currentBelief.id] = chosenValue;
chosenActions[currentBelief.id] = chosenActionIndex;
// Check if the iteration brought an improvement
if (cc.isLess(storm::utility::zero<ValueType>(), result_backup[currentBelief.id] - result[currentBelief.id]) ||
cc.isLess(storm::utility::zero<ValueType>(), result[currentBelief.id] - result_backup[currentBelief.id])) {
improvement = true;
}
}
}
finished = !improvement;
// back up
result_backup = result;
++iteration;
iterationTimer.stop();
STORM_PRINT("Iteration " << iteration << ": " << iterationTimer << std::endl);
}
STORM_PRINT("Overapproximation took " << iteration << " iterations" << std::endl);
auto overApprox = storm::utility::zero<ValueType>();
for (size_t j = 0; j < initLambdas.size(); ++j) {
if (initLambdas[j] != storm::utility::zero<ValueType>()) {
overApprox += initLambdas[j] *
result_backup[getBeliefIdInVector(beliefGrid, initialBelief.observation,
initSubSimplex[j])];
}
}
overApproxTimer.stop();
storm::utility::Stopwatch underApproxTimer(true);
ValueType underApprox = computeUnderapproximationWithMDP(pomdp, beliefList, beliefIsTarget, targetObservations, observationProbabilities, nextBelieves,
result, chosenActions, gridResolution, initialBelief.id, min, true);
underApproxTimer.stop();
STORM_PRINT("Time Overapproximation: " << overApproxTimer
<< std::endl
<< "Time Underapproximation: " << underApproxTimer
<< std::endl);
STORM_PRINT("Over-Approximation Result: " << overApprox << std::endl);
STORM_PRINT("Under-Approximation Result: " << underApprox << std::endl);
return std::make_unique<POMDPCheckResult<ValueType>>(
POMDPCheckResult<ValueType>{overApprox, underApprox});
}
template<typename ValueType, typename RewardModelType> template<typename ValueType, typename RewardModelType>
std::unique_ptr<POMDPCheckResult<ValueType>> std::unique_ptr<POMDPCheckResult<ValueType>>
ApproximatePOMDPModelchecker<ValueType, RewardModelType>::computeReachabilityProbability(storm::models::sparse::Pomdp<ValueType, RewardModelType> const &pomdp, ApproximatePOMDPModelchecker<ValueType, RewardModelType>::computeReachabilityProbability(storm::models::sparse::Pomdp<ValueType, RewardModelType> const &pomdp,

4
src/storm-pomdp/modelchecker/ApproximatePOMDPModelchecker.h

@ -30,6 +30,10 @@ namespace storm {
std::set<uint32_t> targetObservations, bool min, std::set<uint32_t> targetObservations, bool min,
uint64_t gridResolution); uint64_t gridResolution);
std::unique_ptr<POMDPCheckResult<ValueType>>
computeReachabilityRewardOTF(storm::models::sparse::Pomdp<ValueType, RewardModelType> const &pomdp, std::set<uint32_t> targetObservations, bool min,
uint64_t gridResolution);
std::unique_ptr<POMDPCheckResult<ValueType>> std::unique_ptr<POMDPCheckResult<ValueType>>
computeReachabilityProbability(storm::models::sparse::Pomdp<ValueType, RewardModelType> const &pomdp, computeReachabilityProbability(storm::models::sparse::Pomdp<ValueType, RewardModelType> const &pomdp,
std::set<uint32_t> targetObservations, bool min, std::set<uint32_t> targetObservations, bool min,

Loading…
Cancel
Save