Global Search

Hi,

- Is there any recommendation for setting the NumTrialPoints in Global Search (except the 1000 default value)? How is it related to the number of decision variables? (For instance, in GA, it is recommended in MATLAB that the population size to be 15 times of the number of the decision variables.)

- Is there any specific relationship in setting NumStageOnePoints and NumTrialPoints in Global Search solver?

Thanks, Mehdi

Answer 1

[Paul Kerr-Delworth]

Hi,

Increasing NumTrialPoints can help the GlobalSearch solver search more of the decision variable space. However, unlike GA, I am not aware of a rule of thumb you can use to determine the value you should use.

There is no specific relationship between the NumStageOnePoints and NumTrialPoints options. The idea behind the NumStageOnePoints option is that the first NumStageOnePoints generated by scatter search are less likely to be in the vicinity of global minimum. So, we just take the best one of the NumStageOnePoints and run the local solver from there. In some sense, this option can be considered a trade off between speed and completeness of search. Increase NumStageOnePoints and GlobalSearch should terminate faster because it is likely to call the local solver fewer times. Not calling the local solver as many times means that the search may not cover so much of the decision space.

Sorry for not being very specific, but hope this helps in some way!

Best regards,

Paul