Strategies for reducing set of considered models
Disclaimer: this is a discussion issue, possibly serving as a source for subsequent technical issues, not as an issue to implement.
Status
After the first OAT part of the getViableProjections
one gets a matrix of OATviable
with 0
,1
entries for each of n
parameter points (in rows) and each of p
parameters (in columns). The matrix encodes whether the j
-th OAT parameter projection was viable or not for the i
-th parameter point.
Currently the set of reduced models which is considered in the second part contains all unique projections obtained per row. Up to p*n + n
models are evaluated in both parts, resulting in possibly n
maximally reduced models, one per each parameter point sample.
Moreover, in the OATviable
and over the recursive steps (more important for the running time), viable projection P1
possibly smaller than other already found viable projection P2
can be considered. This seems to make sense in a recursive steps, due to the possibility of finding another viable projection P3
larger than P1
, but non-comparable with P2
(cf. hand-sketch: ProjectionsPoset-SmallerToNoncomparable.pdf). Observe that:
- Projection
P3
could be found only with parameter point samples other than witnesses ofP2
andP1
; - samples with projected
P2\P1
parameters will not be witnesses for any suchP3
.
Considerations
- For a faster, more greedy and inaccurate search of maximal reductions allow to directly skip non-maximal projections at each recursive steps (use
pleq
instead ofpeq
inProjectionSet.has()
), and ingetViableProjections
(on top ofunique
remove also rows withpleq
, or re-implement to fully useProjectionSet
). - Each sample knows which projections it witnessed (viable projections for that sample), and possibly take advantage of the observations above in
getViableProjections
andtopologicalFilteringRecursiveStep{Parallel,Sequential}
. - Reduction to a single unique projection obtained per column (after elimination of all zeros rows), i.e. only parameters non-essential in all considered samples are considered for maximally reduced model. This corresponds to making much smaller jumps in the model space, if such reductions are found at all. Arguably, such strategy would make sense with local analysis based on small sample and re-sampling per reduction, in contrast to a global large initial sample.