How to set maximum iterations and criterion for convergence in parameter learning

The engine.
Post Reply
Bo Hu
Posts: 18
Joined: Tue Aug 02, 2016 4:13 am

How to set maximum iterations and criterion for convergence in parameter learning

Post by Bo Hu »

Hello,

Does DSL provide any functions to set or change the maximum iterations and criterion for convergence in parameter learning using static EM? And what is the default setting for the maximum iteration in static EM?

Thank you,
Bo
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: How to set maximum iterations and criterion for convergence in parameter learning

Post by shooltz[BayesFusion] »

Does DSL provide any functions to set or change the maximum iterations and criterion for convergence in parameter learning using static EM? And what is the default setting for the maximum iteration in static EM?
Unfortunately, SMILearn currently uses a hardcoded convergence crtiterion in its EM implementation. The next release planned for late spring this year will fix this issue.

Just for the record, SMILE and SMILearn are now developed by BayesFusion, not by Decision Support Laboratory. We keep the DSL_ prefix on the library class names for backward compatibility.
Bo Hu
Posts: 18
Joined: Tue Aug 02, 2016 4:13 am

Re: How to set maximum iterations and criterion for convergence in parameter learning

Post by Bo Hu »

So, at this point, the maximum number of iteration cannot be re-set either, right?
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: How to set maximum iterations and criterion for convergence in parameter learning

Post by shooltz[BayesFusion] »

EM in SMILearn does not have a check based on iteration count. The exit from main EM loop is based on the ratio of log likelihood from previous and current iteration. The threshold ratio, as mentioned earlier in this thread, is currently hardcoded.
Bo Hu
Posts: 18
Joined: Tue Aug 02, 2016 4:13 am

Re: How to set maximum iterations and criterion for convergence in parameter learning

Post by Bo Hu »

Thank you very much!
Post Reply