Choose Language Hide Translation Bar
Highlighted
Jeff2010
Level I

Questions about "number of starts" and "search time"

Could anyone help me understand "number of starts" and "search time"?

 

1. Will increase "number of starts" or use much longer "search time" provide a better optimal design?

2. What is a good guess for "number of starts"? i.e., let us say a split-plot design with 5 whole plots and 25 subplots. 

 

Thanks!

1 ACCEPTED SOLUTION

Accepted Solutions
Highlighted

Re: Questions about "number of starts" and "search time"

I assume that you already read the documentation about the number of starts and the design search time.

 

The custom design is usually not available from a closed form computation so JMP uses a numerical optimization or search algorithm that is not guaranteed to find the global optimum. The solution depends on the search path, so it might end up at a local minimum instead. The idea behind custom design is to repeat the search to increase the likelihood that it finds the global optimum. So you specify the number of starts regardless of the search time or the search time regardless of the number of starts.

 

You can decrease either parameter if the design takes too long to compute or increase either parameter to improve your chance of the global optimum.

 

Note that there are many cases that have perhaps very different solutions that are equally globally optimal. The solutions are indistinguishable by the optimality criteria. Which of these solutions JMP returns is, of course, a random outcome. You can click Back and then Make Design to get another solution.

 

I recommend that you let JMP determine reasonable parameter values but modify them if you believe the change would help your solution. The default parameter value (guess) is adjusted according to some aspects of your experiment based on empirical performance.

Learn it once, use it forever!

View solution in original post

2 REPLIES 2
Highlighted

Re: Questions about "number of starts" and "search time"

I assume that you already read the documentation about the number of starts and the design search time.

 

The custom design is usually not available from a closed form computation so JMP uses a numerical optimization or search algorithm that is not guaranteed to find the global optimum. The solution depends on the search path, so it might end up at a local minimum instead. The idea behind custom design is to repeat the search to increase the likelihood that it finds the global optimum. So you specify the number of starts regardless of the search time or the search time regardless of the number of starts.

 

You can decrease either parameter if the design takes too long to compute or increase either parameter to improve your chance of the global optimum.

 

Note that there are many cases that have perhaps very different solutions that are equally globally optimal. The solutions are indistinguishable by the optimality criteria. Which of these solutions JMP returns is, of course, a random outcome. You can click Back and then Make Design to get another solution.

 

I recommend that you let JMP determine reasonable parameter values but modify them if you believe the change would help your solution. The default parameter value (guess) is adjusted according to some aspects of your experiment based on empirical performance.

Learn it once, use it forever!

View solution in original post

Highlighted
Jeff2010
Level I

Re: Questions about "number of starts" and "search time"

Thanks for the explanation! It is very helpful! 

Article Labels