Reading this presentation http://ice.uchicago.edu/2009_presentations/Skrainka_OptHessians.pdf I am missing one of the most valuable tools in NLP modeling: setting appropriate bounds. In my opinion (and experience) in NLP modeling there are three areas of attention when the model fails to yield an optimal solution:
- Provide better initial solution
- Do better scaling
- Provide better bounds
Fixing these will help most ill-behaved models. These points are somewhat interrelated: a good bound may prevent the solver go into an area with scaling issues and a good bound also can change the initial point. Bounds are a very important tool in NLP modeling and should never be just ignored.