Tricks of the trade
NOMAD has default values for all algorithmic parameters. These values represent a compromise between robustness and performance obtained by developers on sets of problems used for benchmarking. But you might want to improve NOMAD performance for your problem by tuning the parameters or use advanced functionalities. The following sections provide tricks that may work for you.
Here are a few suggestions for tuning NOMAD when facing different symptoms. The suggestions can be tested one by one or all together.
Symptom |
Suggestion |
Ref. |
---|---|---|
I want to see more display |
Increase display degree |
|
Quantifiable constraints |
Try PB EB or combinations |
|
Difficult constraint |
Try PB instead of EB |
|
No initial point |
Add a LH search |
|
Variables of different magnitudes |
Change blackbox input scaling |
|
Change \(\Delta_0\) per variable |
||
Tighten bounds |
||
Many variables |
Fix some variables |
|
Use PSD-MADS |
||
Unsatisfactory solution |
Change direction type to |
|
Change initial point |
||
Add a LH search |
||
Add a VNS Mads search |
||
Tighten bounds |
||
Change \(\Delta_0\) |
||
Modify seeds that affect algorithms |
||
Disable quadratic models |
set |
|
Unable SGTELIB models |
set |
|
Disable opportunistic evaluations |
set |
|
Disable anisotropic mesh |
set |
|
Change anisotropy factor |
set |
|
Improvements get negligible |
Change stopping criteria |
Type |
Disable quadratic models |
set |
|
It takes long to improve \(f\) |
Decrease \(\Delta_0\) |
|
Optimization is time consuming |
Perform parallel blackbox evaluations |
Blackbox evaluation of a block of trial points and Parallel evaluations |
Blackbox is not that expensive |
Setup maximum wall-clock time |
remove |
Add a LH search |
||
Add a VNS Mads search |