The current approach to blending various contributions to the objective function (evaluates to the Objective Value) involves model variables smin, smax & sdlim and middleware functions setSclDen.js & sclden.js. It is possible that limitations of this approach have a role in creating "pathological cases" where Pattern Search converges (delmin) to a solution that is significantly inferior to a solution that is known to be available and reachable from a (perhaps only slightly) different starting point. Issue #730 and the v4.3rc2 version of the search at script tutor8 p3 illustrate one such case.
The discussion in #730 has included the ideas of:
introducing a logarithmic value for the scaling denominator (smin, smax) and its limit (sdlim)
scaling all constraint contributions to a 0-1 scale
These and other approaches to improve objective function scaling should be investigated as part of resolving this issue.
The current approach to blending various contributions to the objective function (evaluates to the Objective Value) involves model variables smin, smax & sdlim and middleware functions setSclDen.js & sclden.js. It is possible that limitations of this approach have a role in creating "pathological cases" where Pattern Search converges (delmin) to a solution that is significantly inferior to a solution that is known to be available and reachable from a (perhaps only slightly) different starting point. Issue #730 and the v4.3rc2 version of the search at script tutor8 p3 illustrate one such case.
The discussion in #730 has included the ideas of:
These and other approaches to improve objective function scaling should be investigated as part of resolving this issue.