Details of the OPTEX Procedure |
Search Strategies
General Recommendations
As with all combinatorial optimization problems, finding efficient
experimental designs can be difficult. For this reason, the OPTEX
procedure provides a variety of ways to customize the search.
Although default settings make the procedure simple to use "as is," you
can usually improve the search using knowledge of the specific design
problem. For example, if the default algorithm
(EXCHANGE) runs quickly but it is not clear whether it
finds the best design, you can try a slower but more reliable
search method or use more iterations than the default number of 10.
Set of Candidate Points
The choice of candidate points can profoundly affect both the speed
with which the search converges at a local optimum and the likelihood
that this local optimum is indeed the global optimum. Up to a point,
the more candidate points there are, the better the resulting optimum
design will be but the longer it will take to find. Any prior knowledge
that can be brought to bear on the choice of candidates will almost
certainly improve the search. For example, for first- or second-order
models it is usually adequate to restrict the candidates to just the
center and the edges of the experimental region, or perhaps even less;
refer to Snee (1985), and see the introductory examples
"Handling Many Variables" and "Constructing a Mixture-Process Design"
.
Initial Design
The reliability of the search algorithms in finding the optimal design
can be quite sensitive to the
choice of initial design. The default method of initialization for
each search procedure should achieve good results for a wide variety
of situations (see the
INITDESIGN= option).
However, in
certain situations it is better to override the defaults. For
example, if there are many local optima and you want to find the
exact global optimum, it will probably be best to start each try
with a completely random design (INITDESIGN=RANDOM). On the other
hand, prior knowledge may provide a specific initial
design, which can be placed in a SAS data set and specified with the
INITDESIGN= option.
Copyright © 1999 by SAS Institute Inc., Cary, NC, USA. All rights reserved.