Hi Srinivasan, Q1: Do you know roughly by when the bug will be fixed? ... A1: The nondeterminism is due to multi-threading of the algorithm. I realized your problem is pretty small--with only 8 decision variables. I replaced the statement in your original SAS program: proc optmodel; with proc optmodel nthreads=1; to run the program in a single threaded mode. And I ran it half a dozen times, all getting the exactly same solutions. The solution time on each run is small, < 2.5 seconds on my PC. If this solution time is acceptable, I'd suggest to run your program in a single threaded mode. Q2: One option we worked with is the msbndrange option. ... A2: MSBNDRANGE affects the decision variable bound range that is used in sampling for starting points. The wider it is, the more scattered the sampling points will be across the domain space where the sampling is conducted. Another factor to consider is that the total sample points is related to MSAMXSTARTS. Given a value of MSMAXSTARTS, the total number of sample points are relatively fixed. When MSBNDRANGE is small, e.g., 50 or 100, the Multistart algorithm can "zoom in" and "see" more local mins. When MSBNDRANGE is large, there is "zoom out" effect--the algorithm does not see finer details of local mins. So if you increase MSBNDRANGE, you should also increase MSMAXSTARTS to have adequate sampling coverage. But note that the number of potential local optima can explode with the increase of MSBNDRANGE. 5000 is a quite big value for MSBNDRANGE in most cases. Thanks. Tao Huang
... View more