topic Re: Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time in SAS Data Science

Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time

DocMartin — Fri, 06 Apr 2018 14:14:28 GMT

I've got an algorithm with around 300 parameters that I'm trying to optimize using PROC OPTLSO. When I run the optimization using 1,990 observations the routine takes 10 seconds to complete. Yet when I add observation #1991 the routines takes 9 minutes to finish. I tried two things: 1. Using a different data set, and trying it out with 1991 observations. No problem. So it's not that the optimization became saturated. 2. Looked at observation #1991 to see if it differed substantially from the others. The answer was NO. Have any of you experienced this kind of problem? If so, how did you fix it. Thanks! Andrew

Re: Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time

RobPratt — Fri, 06 Apr 2018 20:24:31 GMT

Can you please share your code and data?

Re: Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time

ballardw — Fri, 06 Apr 2018 22:35:21 GMT

@DocMartin wrote:
. 2. Looked at observation #1991 to see if it differed substantially from the others.

Depending on how you look at this record and compare with the remaining what you see may not be what the algorithms see.

If you have a variable with a format assigned with the other option in proc format it may be that a view in a table or proc print does not show an extreme value.

proc format library=work;
value extreme
0-100 = [f5.2]
other = 100;
run;

data example;
   do x= 1,10,99.4, 10e18;
   output;
   end;
   format x extreme.;
run;

proc print data=example noobs;
run;

Or perhaps no single value is extreme but you have added a combination of values that might be considered an outlier that requiring many more iterations to solve.

I would suggest splitting the data into two pieces without about half the records in each and see if the reduced set with the problem record continues to be a problem. I suspect it may. Were there any interesting items in the log for the longer run that don't appear in the shorter?

Without seeing code to know how many variables from your set are being used to create the parameters it is hard to see if a specific minimum sized data set would be needed.

Re: Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time

DocMartin — Sat, 07 Apr 2018 12:07:39 GMT

Thanks! But I'm not using any formatted variables. I've split my data set numerous ways, and each time there seems to be at least one observation that's causing it to hang. I'll check again to make sure the values that are being used are what I expected them to be.

Re: Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time

DocMartin — Sat, 07 Apr 2018 12:09:44 GMT

Because the data are hospitalized patients, I can't share it.

Re: Adding one more observation in PROC OPTLSO leads to 10-fold increased execution time

RobPratt — Sat, 07 Apr 2018 13:55:33 GMT

Can you anonymize the data? Or at least share the code and logs?