Increase frequency of Sbplx change without triggering early termination

[A-J-F-Mackintosh]

Hi,

Thanks for all your work on nlopt, it is an excellent library.

I am interested in using the Sbplx algorithm for gradient-free optimisation of functions with high dimensions (e.g. 3 < n < 12) in python. I was hoping that SBPLX would perform better than Nelder-Mead on such problems, but so far I have found that it is less efficient and more likely to get stuck in local optima.

I have observed that the SBPLX algorithm can spend a long time optimising a single sub-space of the problem, even if it is making very little progress. Increasing the ftol or xtol leads to a more dynamic algorithm that will more frequently move between subspaces rather than use many iterations optimising a single one. However, increasing those tolerances also leads to the algorithm terminating before it has found the true optimum value.

I think that a more agile Sbplx behaviour would be helpful for my optimisation problem. So, is there a way to set different stopping criteria for the optimisation of sub-spaces and the problem as a whole? For example, could I set two different maxeval values, one to tell the Sbplx algorithm to switch subspaces and another to stop the optimisation as a whole? Or two ftol values?

Many thanks,

Alex

[stevengj]

One could add an algorithm-specific parameter for this.