forum.alglib.net

Hi,

I've been using ALGLIB's Levenberg-Marquardt implementation (VJ mode, no acceleration, reusing state) for bounded non-linear optimisation, but it seems that the performance scales really badly. Example timings:

problem size: 360 unknowns and 427 squared sums
objective function total time (for all evaluations of the objective function and jacobian) : 0.485 secs
minlmoptimize total time: 8.836 secs

problem size: 144 unknowns and 169 squared sums
objective function total time (for all evaluations of the objective function and jacobian) : 0.1 secs
minlmoptimize total time: 0.745 secs

problem size: 40 unknowns and 55 squared sums
objective function total time (for all evaluations of the objective function and jacobian) : 0.048
minlmoptimize total time: 0.1 secs


So in the last case, total time is 2x total objective function time.
For the case before, it's 7x.
For my slowest case, total evaluation time is 18x the total objective function time.
I've tried it with same squared sums but less unknowns, and the performance is significantly better.

I'm using all tolerances equal to zero, as I was trying to get the best results possible as parts of my calculations are done in floating point precision.
Is there any quick hint to increase performance? Is there any chance that I can get much better performance with large problems? If I can only use tolerances, which ones would be suggested for such large problems?

Thanks,
Babis

Levenberg-Marquardt scales badly when number of unknowns increases. Iteration cost is proportional to O(M*N^2), where N is for unknowns, and M is for squared sums. I have several ideas about algorithm improvement, but it is impossible to prevent growth of iteration cost.