- Lots of debugging
- Changed the way the filter keeps information about the scheduler to use for probability/reward queries.
| This was done by keeping a special action at the first position of the action list.
| Which was not exactly consistent with the idea behind the filter actions.
| Now the filter keeps this information as an enum value in a member variable.
- All but one tests are green. So we almost reestablished full functionality.
|- The last test that still fails is SparseMdpPrctlModelCheckerTest.Dice where the second to last model check returns the wrong result.
Next up: Debug. Then introduce the full range of filter actions.
Former-commit-id: fd311966cc
- This change splits the path formulas into probabilistic path formulas like Next or Until and reward path formulas like InstantaneousReward or SteadyStateReward.
|- That way it is assured at compile time that no reward path formula can ever be subformula of any probabilistic bound operator and vise versa.
Next up: Adopt changes in the Csl formula tree to the Csl parser.
Former-commit-id: d74c88bbf8
- Every PRCTL formula that worked before works now and behaves in the same way. One exception:
|- Formulas of the type Pmin<0.5[Phi] and Rmin<0.5[Psi] result in a parsing error, as the comparison operator already implies the scheduler to be used.
| Also, the modelchecker now actually uses the comparison operator in order to choose the correct scheduler for MDPs.
Former-commit-id: d942d18e7e
- Fixed similar undefined behavior for the MarkovAutomaton Csl modelchecker.
Next up: Make necessary changes to the formula parsers.
Former-commit-id: e8765fe58b
- General function of the filters: The filter as an abstraction layer between the control flow and the formula/modelchecker.
|- It has a formula as child but is not a formula itself.
|- It invokes the modelchecking process on the child formula and manipulates the result.
|- For the purpose of result manipulation it keeps a list of filter actions.
|- Each action manipulates the result in a certain way. For example: It returns only the results for states 25 to 140.
|- Furthermore the printing of the result to standard out and the log is no longer done by the modelchecker but by the filter.
|- That way the tasks of each class becomes more clear: Modelchecker to compute the results, filter to prepare the computed results for write out.
- Battled with a major design problem: How to integrate the optimizing operator (aka. min or max probs for non-det. models) into the filter scheme.
|- It is now integrated as a separate filter action, which does not touch the results but hold the flag determining whether to maximize or to minimize.
|- This action must be the innermost filter action (i.e. the first list entry) to have any effect.
|- This is combined with a special fuction of the modelchecker that manipulates the mutable minimizationStack calculates the modelchecking result and resets the stack to its original state.
|- This way the information whether to min, to max or not to try is managed by the filter and propagated as needed.
|- Remark: Fixed a major risk of undefined behavior in the SparseMdpPrctlModelChecker.
|- If the formula to be checked did not have a NoBoundFormula as root then the minimumOperatorStack would be empty and minimumOperatorStack.top() would result in undefined behavior.
|- Added tests whether the stack is empty before trying to read out the possibly non existant top element.
Next up: Implement similar filters for LTL and CSL and try to get it compiled.
Former-commit-id: 577998e027
- Began removing NoBoundFormulas, since they might not be needed anymore. This task will be taken over by filters if they are to be implemented.
Next up: CSL
Former-commit-id: 6164f73737
- Fixed one problem marked FIXME in the transpose function. The need for a "sentinel" element was created by an off by one in the prior loop.
- Changed all occurences of SparseMatrix<bool> to SparseMatrix<T>. Now the only two types for which SparseMatrix is instantiated are double and int.
- Compiles again.
|-> Compile time seems to be roughly the same for clean; make all. For incremental builds I haven't tested yet.
Former-commit-id: 6d829e0903
Reverted unnecessary changes to the AbstractModel checker.
Next on the list: Adapting the subsystem generation routine to the new method of providing the subsystem to the model checker.
Former-commit-id: 6c90c064a2
Ran into problem with the AbstractModelChecker being declared const for the model check.
I use it for the subsystem generation and tell it what the current subsystem is. so I have two options:
1. Carry the subsystem as argument through all checking functions of the complete checking tree
2. Store the subsystem in the checker and use it in checkAp to induce the correct result back through the tree.
In the original implementation I used option 2.
But that does only work if it is not constant.
Former-commit-id: 8a833cc05e
* included small consensus example
* made backward-transition generation more beautiful and versatile
* included Dijkstra search for most probable paths
* included first rough scheduler-guessing (there's room for improvement though)
Former-commit-id: db795fa1bf
dehnert
sjunges
masawei
PBerger
David_Korzeniewski