learning new network

[JTJ]

Hi,
I am new to GeNIe and have the following question. I have a network to start with (say for example the Alarm.xdsl found in the example folder with GeNIE) which I open in GeNIe and use it to generate data (using Network -> Generate Data File. If I now use this data file to learn the network (using Data -> Learn New Network) then irrespective of the algorithm I use, i dont get the similar network. Infact, some of the edges are in the opposite direction to what it was in the original network. I dont follow why this is so. I would be happy if someone could hep me. Thanks
JTJ

[marek [BayesFusion]]

Hi,
I am new to GeNIe and have the following question. I have a network to start with (say for example the Alarm.xdsl found in the example folder with GeNIE) which I open in GeNIe and use it to generate data (using Network -> Generate Data File. If I now use this data file to learn the network (using Data -> Learn New Network) then irrespective of the algorithm I use, i dont get the similar network. Infact, some of the edges are in the opposite direction to what it was in the original network. I dont follow why this is so. I would be happy if someone could hep me.

This is normal in case of structure recovery algorithms. The main reason for this is structural indistinguishability, which comes down to the fact that some structures cannot be (statistically) distinguished from eaech other. For example, a structure X->Y cannot be without additional information distinguished from the structure X<-Y. Both result in the same set of independencies. Another reason is, of course, the possibility of errors in statistical decisions made in the process of searching for the graph. The first reason is most important, however, as it has no theoretical solution (the second could be addressed, for example, by increasing the size of the data file).

This having said, there are structures that are unique and should be retrieved if you have enough data. Please try the network credit.xdsl, included among the example networks -- it should be uniquely identifiable.
I hope this helps.
Cheers,

Marek

[fulu]

What are the exact learning algorithms implemented in Genie? Anyone can give me some pointers on related papers that describe the learning algorithms that are implemented in Genie?

Thanks in advance,

Fulu

[mark]

Please search the forum for explanations and references. If you still have questions, please let me know.