[BuiltConstraintSolver] Add SVD based regularization strategy #5722
Conversation
…info in GenericCosntraintSolver data that are dependent on the method type. Need to find a way of setting the solving method by adding a component in the scene.
…ritance. Now the constraintProblem is only a container
… still being used
…_constraint_solver
bakpaul
added
pr: status to review
hugtalbot
changed the title
| { | ||
| if (regularization>std::numeric_limits<SReal>::epsilon()) | ||
| { | ||
| auto * FullW = dynamic_cast<sofa::linearalgebra::LPtrFullMatrix<SReal> * >(&W); |
There was a problem hiding this comment.
I see that this cast is ensured by how the ComplianceWrapper sets the matrix type in BuiltConstraintSolver. Still, you could add a check for nullptr if anything changes in the future.
| int nullSpaceBegin = -1; | ||
| for(size_t i=0; i<problemSize; i++) | ||
| { | ||
| if (fabs(svd.singularValues()(i)) < fabs(svd.singularValues()(0))/100.0) |
There was a problem hiding this comment.
Might be good to add the 1% criterion as a parameter so that the user can limit/expand the filter.
In a more complex problem, wouldn't it be possible that constraint feasibility is not so clearly separated and you end up with over-regularization. Or on the contrary being too strict with the criterion?
There was a problem hiding this comment.
Yes, it is highly probable. But in that case, if no singular value is smaller than the rest, then all the problem should be regularized in my mind.
On the parameter side yes, I also wanted to do that. But I dislike adding new parameters that only apply if another parameter is true. LEt's discuss this in the sofa-dev-meeting
|
[ci-build][with-all-tests] |
bakpaul
added
the
pr: status wip
fredroy
removed
the
pr: status to review
3 participants
Based on #5666
This PR adds a regularization strategy that uses SVD to explore the NullSpace of the matrix to only add regularization to constraints that have antagonists constraints.
meaningful commit: 30f81e4
Introduced example
The added example shows a beam attached at both sides and which center is constrained in two different positions, meaning that two antagonists constraints are imposed to the center. The idea is that, because the fixations are not faulty, then we want them to be fully solved and we don't want to perturb them while still releasing the constraints on the antagonists ones.
With SVD
Without SVD
By submitting this pull request, I acknowledge that
I have read, understand, and agree SOFA Developer Certificate of Origin (DCO).
Reviewers will merge this pull-request only if