Calibration with NL2SOL not exploring full range of design space

[philippe-manne]

Hi. I am using NLSOL for calibration work. Stubbornly, it keeps exploring a very limited range of the design space instead of reaching out towards the lower and upper bounds despite residues being high. I am wondering if my problem is ill defined or not. I don't see why it behaves like this. Activating scaling, convergence, ... or not doesn't help. Switching the order of parameters or removing some doesn't change the phenomenon. To facilitate the investigation, I boiled down my problem to this very simple setup which suffers from the same symptoms. In other words, what is preventing NLSOL from checking parameter values like (beta=16, gamma=23) after the initial sensitivity assessments??? It then results into false convergence... Any insight will be appreciated. Thank you in advance. -- Philippe

variables
continuous_design = 2
initial_point 20. 18.
lower_bounds 15. 15.
upper_bounds 25. 25.
descriptors 'beta' 'gamma'

yields:
%eval_id interface beta gamma RESIDUE1 RESIDUE2 RESIDUE3 RESIDUE4
1 NO_ID 20 18 -31.097 -111.867 -229.416 -226.87
2 NO_ID 19.99674203 17.99821682 -34.908 -115.406 -229.829 -229.983
3 NO_ID 19.99962816 17.99979595 -36.339 -114.153 -231.036 -228.648
4 NO_ID 19.99996282 17.99997959 -31.62 -113.892 -229.653 -227.731
5 NO_ID 19.99999628 17.99999796 -30.869 -113.024 -229.645 -229.138
6 NO_ID 19.99999963 17.9999998 -32.005 -113.687 -228.849 -228.536

dakota.log
rim_opt_nls.dat.txt
rim_opt_nls.in.txt
rim_opt_nls.out.txt

[jadamstephens]

NL2SOL is a local, gradient-based optimizer. It proceeds "downhill" to the (local) minimum associated with the basin it begins in. Dakota also has global optimizers and a multi-start strategy that would be a better (but more expensive) job of exploring the design space and identifying the global minimum.

Here's a listing of Dakota's optimization methods broken down by their characteristics.

[brunorossin]

Im having the same the problem. Is it possible to make Dakota try a larger step for the next iteration? I believe that would solve it in my case. Thank you!

[Ramzy1990]

NL2SOL is a local, gradient-based optimizer. It proceeds "downhill" to the (local) minimum associated with the basin it begins in. Dakota also has global optimizers and a multi-start strategy that would be a better (but more expensive) job of exploring the design space and identifying the global minimum.

Here's a listing of Dakota's optimization methods broken down by their characteristics.

Apologies for bumping this topic up but a lot of collaborators I work with also had this issue. One indeed has to know which optimizer to use and note the properties of the optimizer. In general, if someone would like to explore their design space, then heuristic methods like evolutionary algorithms (e.g., genetic algorithm) are better for such purposes.

Just one comment though; one probably will not find the global minimum numerically and will be a local minimum at best. Unless it is a very limited design space with many constraints (typically toy problems or verification problems).