Learning DBN parameters (transition probablilities) in GeNIe

[mark]

A few remarks:

It seems to me that the amount of data is insufficient to learn the number of parameters you have in the model. Is it possible to obtain more data?

Did you include the learned network in the zip file? All the probabilities in the CPTs in DBN_mark_test_17411.xdsl are uniform so I don't think so.

In the unrolled network the CPT for node t changes at time 4, which should not be the case. How did you obtain the unrolled network?

[ksm]

A few remarks:

It seems to me that the amount of data is insufficient to learn the number of parameters you have in the model. Is it possible to obtain more data?

Did you include the learned network in the zip file? All the probabilities in the CPTs in DBN_mark_test_17411.xdsl are uniform so I don't think so.

In the unrolled network the CPT for node t changes at time 4, which should not be the case. How did you obtain the unrolled network?

Hi Mark,

In case of the amount of data, I tried adding more biased data generated through GeNIe itself it but I got similar results. However, I would only like the DBN to be learnt through my original data itself.

In my previous post, I included the learnt network (unrolled network). In that case, I unrolled the network (DBN_mark_test_17411.xdsl ) to 5 time slices first and then performed learning on the unrolled network itself using the data file provided. Learning was done by randomizing the parameters with relevance enabled. I guess due to randomization, the CPTs in slice 4 are changing a bit. From my current understanding, after running experiments several times is that, when learning is performed on the unrolled network, the network at each time slices are treated as single BNs separately and the transition probabilities are not used for learning and that's why nodes at time slice 5 are not predicted at all. However, it is the other way around when learning is performed when the network is not unrolled.

I have now included the learnt network (DBN_mark_test_18411_learnt_WITHOUT_UNROLL.xdsl) which was learnt without unrolling. As we can see now, the values at time slice 5 are predicted but I am unsure whether the results are OK.

If I again try to learn the network after unrolling, the network learns and does not predict the values at time slice 5. Do you think the network is not able to learn and predict correctly when the learning is performed on the unrolled network? I guess, I should perform learning on the network which is not unrolled? What are your thoughts on that? Can you please take a look at the file provided in this post?

Thanks and regards,
\ksm

[mark]

The right way is to perform learning on the original network and never on the unrolled network. This way your predictions should be much more accurate. Can you give this a try?

[ksm]

The right way is to perform learning on the original network and never on the unrolled network. This way your predictions should be much more accurate. Can you give this a try?

Hi Mark,

Thanks for the reply. Yes, I tried it like that as mentioned in my previous post. Did you manage to take a look at the DBN (DBN_mark_test_18411_learnt_WITHOUT_UNROLL.xdsl) I attached in my previous post? I think it looks alright now. I would like to understand why there is a difference between network that is learnt with and without unrolling.

Thanks and Regards,
'\ksm

[mark]

It look ok, although many entries in the CPTs do not seem to be updated because of a lack of data.

There is a big difference between learning with and without unrolling. If you unroll, the CPTs at each time step are learned separately. However, these CPTs are usually assumed to be identical and then it makes much more sense to pool the data and learn a CPT only once, which is what happens if you don't unroll.

[ksm]

It look ok, although many entries in the CPTs do not seem to be updated because of a lack of data.

There is a big difference between learning with and without unrolling. If you unroll, the CPTs at each time step are learned separately. However, these CPTs are usually assumed to be identical and then it makes much more sense to pool the data and learn a CPT only once, which is what happens if you don't unroll.

Hi Mark,

Thank you very much for your quick responses. I really appreciate your help!

Regards,
\ksm