Dynamic Bayesian Network (DBN) for time series

[tengwang]

Hi, thanks for your attention.

I want to use GeNIe to build a recursive time series including uncertainty. The initial value a_0=Normal(1,1). Then, a_{i+1}=a_i+Normal(0,0.01).

When I try DBN in GeNIe, I find I can not make the function node be recursive. The details are given in the picture below,



Thanks a lot for your help!

looking forward to your reply.

[shooltz[BayesFusion]]

Sorry - the equations are not supported in the DBNs.

[tengwang]

Do you mean GeNIe can not do it or no DBN can handle it? I think it is not a difficult issue. There should be some solutions. Thanks a lot

[marek [BayesFusion]]

Hi Teng Wang,

It is the first -- we have not married equations and DBNs but you are right that it is possible in theory. One way that I deal with it now is creating static networks with equations and deriving CPTs for the equation nodes through discretization. Once the nodes are discretized, I copy the CPTs to the DBN. Does this help?

Marek

[tengwang]

Thanks for your reply. At least I know this issue can be solved via DBN. Manually copy this time-consuming. I will explore other tools. Thanks again for the useful information.

[marek [BayesFusion]]

Hi Teng Wang,

It is actually not that bad -- I have created DBN models in GeNIe that were based on systems of difference equations, such as predator-prey models. Unlikely you will find another tool -- GeNIe has the most flexible implementation of hybrid Bayesian networks. Finally, we will extend DBNs with support of equations, so please stay tuned. Currently we are tied up with other important modules.
Cheers,

Marek

[tengwang]

Hi, marek,

Could you share the predator-prey models as you mentioned?

Sure, Iooking forward to this extension in GeNIe. Many thanks.

Teng Wang

[marek [BayesFusion]]

Here are two models that you may want to explore. The "Foxes Rabbits Equations.xdsl" model is a hybrid model that contains the difference equations describing the dynamic model. I used this model to derive the CPTs for the Dynamic Bayesian Network in "Foxes Rabbits DBN 20.xdsl". I discretized the continuous variables F and R into 20 intervals.

[tengwang]

Thanks a lot for the example. It does help me. I will try this strategy for the time being. Hope GeNIe can develop function-based BDN in the future. Thanks again~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~