search

amidst / toolbox

A Java Toolbox for Scalable Probabilistic Machine Learning
http://www.amidsttoolbox.com
Apache License 2.0
119 stars 35 forks source link

The probability distributions remain uniform #81

Closed gowthamnatarajan closed 6 years ago

gowthamnatarajan commented 6 years ago

The probability distributions for variable without a parent remains a uniform distribution. It does not learn the marginal probability from the data.

andresmasegosa commented 6 years ago

Hi,

could you provide more details, please?

gowthamnatarajan commented 6 years ago

I looked deeper into the model and found that ALL probabilities (marginal and conditional) is uniform. That is its all 0.5 if it has 2 states or all are 0.333 if it has 3 states and so on at Time 0 only. The probabilities at Time T all probabilities are correct. Weird!! Why is that??

Should I create a dynamicDag separately for Time 0 and Time T?

andresmasegosa commented 6 years ago

How is the training data you are using? is it composed of a single sequence with many time steps?

gowthamnatarajan commented 6 years ago

Many sequences and many time steps

On Fri, Nov 3, 2017, 07:46 Andres R. Masegosa notifications@github.com wrote:

How is the training data you are using? is it composed of a single sequence with many time steps?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/amidst/toolbox/issues/81#issuecomment-341724129, or mute the thread https://github.com/notifications/unsubscribe-auth/AE5TwOh_RIyCT6x8ZmOR__FqnOi9j6e9ks5syyc2gaJpZM4QQ7Xp .

andresmasegosa commented 6 years ago

Could you please share the piece of code you are running? Just to know which learning algorithm you are using and how you use it.

gowthamnatarajan commented 6 years ago

I found the issue. The TIME_ID for every sequence was starting at 1 and not 0. In the example data-sets too it was starting at 1. Can some new examples be added which start at 0?