private PrecalculatedTransitionTarget[] CreatePrecalculatedTransitionTargets(Formula phi, Formula psi)
{
// [phi U<=steps psi]
var psiEvaluator = Ltmdp.CreateFormulaEvaluator(psi);
var precalculatedTransitionTargets = CreateEmptyPrecalculatedTransitionTargetArray();
CalculateSatisfiedTargets(precalculatedTransitionTargets, psiEvaluator);
if (phi != null)
{
// excludedStates = Sat(\phi) \Cup Sat(psi)
var phiEvaluator = Ltmdp.CreateFormulaEvaluator(phi);
Func <int, bool> calculateExcludedStates = target =>
{
if (precalculatedTransitionTargets[target] == PrecalculatedTransitionTarget.Satisfied)
{
return(false); //satisfied states are never excluded
}
if (!phiEvaluator(target))
{
return(true); //exclude state if it does not satisfy phi
}
return(false);
};
CalculateExcludedTargets(precalculatedTransitionTargets, calculateExcludedStates);
}
return(precalculatedTransitionTargets);
}
internal double CalculateMaximumFinalProbability(AutoResizeBigVector <double> cache, PrecalculatedTransitionTarget[] precalculatedTransitionTargets, double[] initialStateProbabilities)
{
var cid = Ltmdp.GetRootContinuationGraphLocationOfInitialState();
cache?.Clear(cid); //use cid as offset, because it is the smallest element
return(CalculateMaximumProbabilityOfCid(cache, precalculatedTransitionTargets, initialStateProbabilities, cid));
}
private void CalculateUnderlyingDigraph()
{
// We use the underlying digraph to interfere the transitionTargets with the final probability
// of 0 or 1.
// I think, the data from the graph is also valid for the states. So, if a state-node
// is in Prob0 or Prob1, then we can also assume that the state has always probability 0 or 1,
// respectively. One further check can could be introduced. When we know that the probability
// of a state is 0 or 1, then we do not have to check the outgoing transitionTargets anymore.
if (_underlyingDigraph != null)
{
return;
}
_output.WriteLine("Creating underlying digraph");
_underlyingDigraph = Ltmdp.CreateUnderlyingDigraph();
_output.WriteLine("Finished creating underlying digraph");
}
internal int CountTargetStatesOfCid(long cid)
{
var targetStatesCount = 0;
Action <LabeledTransitionMarkovDecisionProcess.ContinuationGraphElement> counter = cge =>
{
if (cge.IsChoiceTypeUnsplitOrFinal)
{
targetStatesCount++;
}
};
var traverser = Ltmdp.GetTreeTraverser(cid);
traverser.ApplyActionWithStackBasedAlgorithm(counter);
return(targetStatesCount);
}
internal double SumProbabilitiesOfCid(long cid)
{
Func <LabeledTransitionMarkovDecisionProcess.ContinuationGraphElement, double> leafCounter = (cge) =>
{
var myProbability = cge.Probability;
return(myProbability);
};
Func <LabeledTransitionMarkovDecisionProcess.ContinuationGraphElement, IEnumerable <double>, double> innerCounter = (cge, childProbabilities) =>
{
var myProbability = cge.Probability;
var childProbability = 0.0;
foreach (var p in childProbabilities)
{
childProbability += p;
}
return(myProbability * childProbability);
};
var traverser = Ltmdp.GetTreeTraverser(cid);
var probabilties = traverser.ApplyFuncWithRecursionBasedAlgorithm(innerCounter, leafCounter);
return(probabilties);
}
internal double[] MaximumIterator(AutoResizeBigVector <double> cache, PrecalculatedTransitionTarget[] precalculatedTransitionTargets, int steps)
{
var stopwatch = new Stopwatch();
stopwatch.Start();
var stateCount = Ltmdp.SourceStates.Count;
var xold = new double[stateCount];
var xnew = new double[stateCount];
var loops = 0;
while (loops < steps)
{
// switch xold and xnew
var xtemp = xold;
xold = xnew;
xnew = xtemp;
loops++;
for (var i = 0; i < stateCount; i++)
{
var cid = Ltmdp.GetRootContinuationGraphLocationOfState(i);
cache?.Clear(cid); //use cid as offset, because it is the smallest element
xnew[i] = CalculateMaximumProbabilityOfCid(cache, precalculatedTransitionTargets, xold, cid);
}
if (loops % 10 == 0)
{
stopwatch.Stop();
var currentProbability = CalculateMaximumFinalProbability(cache, precalculatedTransitionTargets, xnew);
_output?.WriteLine(
$"{loops} Bounded Until iterations in {stopwatch.Elapsed}. Current probability={currentProbability.ToString(CultureInfo.InvariantCulture)}");
stopwatch.Start();
}
}
stopwatch.Stop();
return(xnew);
}
private double CalculateMaximumProbabilityOfCid(AutoResizeBigVector <double> cache, PrecalculatedTransitionTarget[] precalculatedTransitionTargets, double[] stateProbabilities, long currentCid)
{
if (cache != null && HasCacheEntry(cache, currentCid))
{
return(cache[(int)currentCid]);
}
var result = double.NaN;
LabeledTransitionMarkovDecisionProcess.ContinuationGraphElement cge = Ltmdp.GetContinuationGraphElement(currentCid);
if (cge.IsChoiceTypeUnsplitOrFinal)
{
var transitionTargetPosition = cge.To;
var transitionTarget = Ltmdp.GetTransitionTarget(transitionTargetPosition);
if (precalculatedTransitionTargets[transitionTargetPosition].HasFlag(PrecalculatedTransitionTarget.Satisfied))
{
result = 1.0;
}
else if (precalculatedTransitionTargets[transitionTargetPosition].HasFlag(PrecalculatedTransitionTarget.Excluded))
{
result = 0.0;
}
else
{
result = stateProbabilities[transitionTarget.TargetState];
}
}
else
{
if (cge.IsChoiceTypeForward)
{
// Note, cgi.Probability is used in the branch "else if (cge.IsChoiceTypeProbabilitstic)"
result = CalculateMaximumProbabilityOfCid(cache, precalculatedTransitionTargets, stateProbabilities, cge.From);
}
if (cge.IsChoiceTypeNondeterministic)
{
var biggest = double.NegativeInfinity;
for (var i = cge.From; i <= cge.To; i++)
{
var resultOfChild = CalculateMaximumProbabilityOfCid(cache, precalculatedTransitionTargets, stateProbabilities, i);
if (resultOfChild > biggest)
{
biggest = resultOfChild;
}
}
result = biggest;
}
else if (cge.IsChoiceTypeProbabilitstic)
{
var sum = 0.0;
for (var i = cge.From; i <= cge.To; i++)
{
var transitionProbability = Ltmdp.GetContinuationGraphElement(i).Probability;
var resultOfChild = CalculateMaximumProbabilityOfCid(cache, precalculatedTransitionTargets, stateProbabilities, i);
sum += transitionProbability * resultOfChild;
}
result = sum;
}
}
if (cache != null)
{
cache[(int)currentCid] = result;
}
return(result);
}
internal int CountTargetStatesOfState(int state)
{
var cidRoot = Ltmdp.GetRootContinuationGraphLocationOfState(state);
return(CountTargetStatesOfCid(cidRoot));
}
internal int CountTargetStatesOfInitialState()
{
var cidRoot = Ltmdp.GetRootContinuationGraphLocationOfInitialState();
return(CountTargetStatesOfCid(cidRoot));
}
internal double SumProbabilitiesOfState(int state)
{
var cidRoot = Ltmdp.GetRootContinuationGraphLocationOfState(state);
return(SumProbabilitiesOfCid(cidRoot));
}
