learning parameter selection

The front end.
Post Reply
wantida
Posts: 6
Joined: Wed Sep 06, 2017 2:13 am

learning parameter selection

Post by wantida »

Hi,

I am not sure that my data fit for which parameter learning function (uniformize, randomize and keep original). The data has 126 samples with 21 discrete nodes and no new data adding, so I decide not to choose a keep original. Then, I try a uniformize and I got all node equal in probabilities. Therefore, I select a randomize (Log(p)=-1218.66). But, I cannot explain how it comes or is it right method to make a decision? In help said that: "Randomize allows for picking random values for parameters, which inserts some randomness in the algorithm's search for the optimal values of parameters." I cannot understand the explanation in help. Please help to provide more details.

Than you.
Wow.
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: learning parameter selection

Post by marek [BayesFusion] »

HI Wow,

Your exclusion of "Keep original" in case you have no meaningful parameters in the network is correct. EM is a heuristic/numerical algorithm that searches for the optimal assignment of parameters. Hence, it is influenced by the starting point. Uniformize makes the starting point very specific -- uniform distribution. The problem with it is that EM sometimes gets stuck in a local maximum that is precisely the uniform distribution. Perhaps this explains why you got uniform distributions. Randomize is safer -- it amounts to starting from some random set of values of parameters. Does this help?

Marek
wantida
Posts: 6
Joined: Wed Sep 06, 2017 2:13 am

Re: learning parameter selection

Post by wantida »

Thanks, it's help!

But I need some explanation more in Randomize technique, for basic understanding. Is Randomize means not every parameter will learn?
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: learning parameter selection

Post by shooltz[BayesFusion] »

Randomize means that the existing parameters in the network will be discarded and replaced by random distributions during the EM init phase. All parameters will be subsequently learned (unless you mark some nodes as fixed).
Post Reply