Closed sjunges closed 1 month ago
I believe I understood the issue:
For Rmin kind of properties, we need to clear those choices from the maybeStates
that lead to a state with infinite expected reward. This happens somewhere in the SparseMdpPrctlHelper
. However, the initial policy is currently computed before this step and thus the selected choice indices in the initial policy might become invalid.
For the DRN at the bottom and the following query, we obtain an error which I believe is due to the creation of the initial policy. The error occurs with and without the exact flag (without, one must set ./bin/storm -drn exactbug.drn --prop "R{\"costs\"}min=? [F \"goal\"]" --minmax:method topological --topological:minmax pi).
EDIT: Regular PI also fails on this model
Some helpful assertions that help are presented in https://github.com/moves-rwth/storm/pull/554
LLDB (without the added assertions says):
The drn file content is here;
For exact solving, the current workaround is to use soplex.