randomize parameter

The front end.
Post Reply
Yan
Posts: 22
Joined: Fri Mar 04, 2022 5:26 am

randomize parameter

Post by Yan »

Dear staff,

Can I ask some questions about "Randomize" option:

1. In the user mannual, it says "Using the Randomize option may be especially useful when learning parameters with latent variables." Could you please explain it in more detail?

2. If we select "radomize" in the parameter learning, is it better to keep this option in validation?

3. In validation, why select "randomize" in the "EM options" produce better accuracy result than "uniformize"?

Thanks a lot.
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: randomize parameter

Post by marek [BayesFusion] »

Hi Yan,

I believe you are asking about randomization in the context of the EM algorithm (please correct me if I am wrong). In validation, this parameter also refers to the EM algorithm part (learning the parameters at each cross-validation step).

It is an empirical finding that randomization gets you more likely to a maximum. You can try it out systematically but I believe that you will just confirm this finding.

Learning and validation are theoretically two different stages of your model development. Still, the finding that randomization is more likely to get you closer to the maximum in EM is valid for the two, so I would use randomization in both.

I hope this helps,

Marek
Yan
Posts: 22
Joined: Fri Mar 04, 2022 5:26 am

Re: randomize parameter

Post by Yan »

marek [BayesFusion] wrote: Tue Jun 21, 2022 7:15 pm Hi Yan,

I believe you are asking about randomization in the context of the EM algorithm (please correct me if I am wrong). In validation, this parameter also refers to the EM algorithm part (learning the parameters at each cross-validation step).

It is an empirical finding that randomization gets you more likely to a maximum. You can try it out systematically but I believe that you will just confirm this finding.

Learning and validation are theoretically two different stages of your model development. Still, the finding that randomization is more likely to get you closer to the maximum in EM is valid for the two, so I would use randomization in both.

I hope this helps,

Marek
Thanks Marek, that makes sense.
Post Reply