Accuracy info for numerical results is computed and stored during model checking. For now,
it is always the worst-case accuracy over all states for which results are computed. The
information is stored in a class prism.Accuracy, instances of which are created in a
consistent fashion using prism.AccuracyFactory. Accuracy info is stored in the
prism.Result class.
Accuracy information is generated for all 4 main engines (sparse/hybrid/MTBDD/explicit),
and attached to the Result object when model checking the filter that is wrapped around a
property (currently, just for the default "state" filter). The accuracy is printed with
the main result at the end of model checking. e.g.:
Result: 0.1666666865348816 (+/- 1.1920928955078125E-7; rel err 7.152556520395694E-7)
For now, accuracy info is generated for the majority of properties for DTMCs and MDPs
(no support yet for multi-objective model checking). For methods like value iteration,
where a guaranteed error bound is not generated, convergence information is used, but
reported as being only an estimate.
Test mode has been updated (code in prism.ResultTesting) to use the reported accuracy of
a result to check correctness against a reference result. If it is missing, or only
estimated, the old approach of using default check of 1e-5 with relative error is used.
For the explicit engine, accuracy info is stored in explicit.ModelCheckerResult, returned
by most numerical computation methods i.e. those in the explicit.ModelChecker and
explicit.IterationMethod (sub)classes. At a higher level, the accuracy info is then stored
in explicit.StateValues, when this created from ModelCheckerResult in
explicit.ProbModelChecker.
For symbolic (sparse/hybrid/MTBDD) engines, the error bound info from C++ code is stored
in a variable last_error_bound that be accessed (or reset) from prism.PrismNative, using
the methods getLastErrorBound() and resetModelCheckingInfo(). At a higher level, this is
then stored in subclasses of prism.StateValues, which are created in the
prism.*ModelChecker classes from the results of computations.
PR #116 (in particular 5c33a555ac) moved
the post-processing of the BSCC steady state probabilities (weighing
by the exit rates of the CTMC) into computeSteadyStateProbsForBSCC().
This introduced a regression, leading to wrong results, as the
post-processing would be applied on the solution vector that was
returned by the method, but not necessarily on the values passed back
out via the result argument vector.
We fix this by moving the post-processing up, before the copy and
adding a note of caution.
Failure e.g. for
prism prism-tests/papers/Par02/Par02-3.3.3.sm prism-tests/papers/Par02/Par02-3.3.3.sm.props -explicit -testall
The computation reuses the computeSteadyStateBackwardsProbs() method
used for the computation of S=?[ phi ]. We tweak the variable names
and comments in that method to make clear that the solution values of
a BSCC - after multiplication with the mult array - are not
necessarily probabilities anymore.
Provide variants of computeSteadyStateProbs() and computeSteadyStateBackwardsProbs() in DTMCModelChecker
that take a BSCCPostProcessor, which can be used to reweight the steady-state probabilities.
(preparation for steady-state computation for CTMCs)
To guarantee convergence, the power method requires the precomputation
P = (Q * deltaT + I)
from: William J. Stewart: Introduction to the Numerical Solution of Markov Chains p. 124.
Sometimes, we want to use the prob0/1... methods to generate schedulers instead of doing
the standard precomputations and don't want to have the normal log output.
Not inherited using inheritSettings()
To obtain the previous behaviour, change
SCCComputer sccs = SCCComputer.createSCCComputer(parent, model);
sccs.computeSCCs();
... = sccs.getSCCs();
to
SCCConsumerStore sccs = new SCCConsumerStore();
SCCComputer sccComp = SCCComputer.createSCCComputer(parent, model, sccs);
sccComp.computeSCCs();
... = sccs.getSCCs();
This additional flexibility in how SCCs can be consumed can be used
in the future for example to handle SCCs on-the-fly, without
having to store all of them at the same time.
git-svn-id: https://www.prismmodelchecker.org/svn/prism/prism/trunk@12110 bbc10eb1-c90d-0410-af57-cb519fbb1720
The DFA needed for co-safe LTL reward computation need to have a
specific structure, i.e., all states that satisfy Sigma^omega have
to be already goal states.
The previous approach did not necessarily ensure this.
git-svn-id: https://www.prismmodelchecker.org/svn/prism/prism/trunk@12065 bbc10eb1-c90d-0410-af57-cb519fbb1720
Joachim Klein