Can the parameters of a DBN be learnt from "regular" time series data sets?

[maghnie]

I'd like to learn the parameters of a DBN from the following data set:

"Time","Valve setpoint","Valve actual","Outdoor air temperature"
2023-12-07 06:00:00,38.0,38.0,2.0
2023-12-07 08:00:00,35.5,35.5,3.3
2023-12-07 10:00:00,34.0,34.0,5.6
2023-12-07 12:00:00,33.7,33.7,5.4
2023-12-07 14:00:00,34.9,34.9,4.4
2023-12-07 16:00:00,36.7,36.7,3.7
2023-12-07 18:00:00,40.1,40.1,2.8
2023-12-07 20:00:00,38.5,38.5,3.0
2023-12-07 22:00:00,35.2,35.2,4.8
2023-12-08 00:00:00,34.6,34.6,5.9
2023-12-08 02:00:00,37.3,37.3,4.2
2023-12-08 04:00:00,36.2,36.2,4.1
2023-12-08 06:00:00,36.8,36.8,4.8
2023-12-08 08:00:00,36.7,36.7,5.5
2023-12-08 10:00:00,36.2,36.2,6.4
2023-12-08 12:00:00,36.2,36.2,7.4
2023-12-08 14:00:00,35.1,35.1,7.7
2023-12-08 16:00:00,36.8,36.8,7.4
2023-12-08 18:00:00,40.4,40.4,7.5
2023-12-08 20:00:00,40.4,40.4,7.6
2023-12-08 22:00:00,1.15,1.15,7.7
...

So my two questions are:

• 1. Can DBNs in GeNIe properly model cyber-physical systems, like building automation control? Of course, given that the structure is already defined by the user.
  2. If yes, is there a utility function somewhere that would convert the data into a format acceptable by GeNIe?

I could try coding it myself, but just thought to ask first, since this seems like a common-occurring situation.

[shooltz[BayesFusion]]

1. Can DBNs in GeNIe properly model cyber-physical systems, like building automation control? Of course, given that the structure is already defined by the user.

Yes, the cyber-physical systems can and are actually modelled. One of our examples is based on the network used by one of our customers. It's distributed with GeNIe as one of the examples, and you can use it directly from your browser at our interactive model repository:
https://repo.bayesfusion.com/network/pe ... ybrid.xdsl

2. If yes, is there a utility function somewhere that would convert the data into a format acceptable by GeNIe?

There's no built-in tool to convert the data from the format you've included to the format GeNIe and SMILE use directly. For reference, SMILE requires one column per node and time step, so the file for 3 nodes and 4 time steps would have 12 columns - see the attached image.

[marek [BayesFusion]]

Please note that there are many ways in which you can transform the data set to make it suitable for learning. Essentially, the first decision that you need to make is what order your model is, i.e., what is the highest time degree of a link in your model. You will reflect your decision in the structure of the model. If you decide that it is just a first order Markov model, then you can transform your data set to the format shown by Shooltz in which there are n-1 rows of x_1 and X_2. You treat every transition in your data as generated by the system. If it is a third order system, you will have n-2 triples of x_1, x_2 and x_3, etc. It might be nice to have a utility like the one you decribe but there are many possibilities for transformation of the input data. The utility would have to be very flexible as far as the input format is concerned. We will consider developing such a utility but it may take a while to do it well, as if we undertake anything, we try to do it well.
I hope this helps,

Marek