Combining Nets

The purpose of this feature is to combine the knowledge of two or more Bayes nets into a single net.  The nets may have been learned from data, or created by experts.

You start with one net (the original net), and then add on another (the additional net), to form the resultant net.  This process may be repeated to combine any number of nets.

If nodes in each of the nets have the same name (regardless of states), they will be interpreted as standing for the same variable, and will result in a single node in the resultant net.  Among nodes having the same name, the resultant node will have all the states of the original node and the states of the additional node. CPTs will be combined accordingly.

How to:  Open your original net, or create a new one.  Choose Modify Combine Nets, and from the standard file-open dialog box, pick the Bayes net file to be added on.  If you wish to combine several nets, then after adding the first one, successively add the others in the same way.

Netica will next ask you to enter a degree for the additional net.  It is an amount to weight the knowledge in that net relative to the first net.  You will almost always accept the default of 1.0, since the experience amounts already weight the nets relative to each other, and the degree is just an extra "experience multiplier" to provide added flexibility (see below).

Experience:  If the nets were learned from data, then they will have experience tables as well as CPT tables, which will weight the nets according to how much data was used to learn each one.  If the nets were created by experts rather than learned from data, they may have CPT tables with no experience tables.  In that case you will be prompted for a single experience number to be applied to the whole net.  The effect of the number you enter is just to weight the nets produced by each expert, but the idea is to enter a number that is an estimate of the equivalent number of cases seen by that expert.  If you wish to have more control over the experience amounts (for instance, different numbers for each node, or even a table of different numbers for each node), then enter the experience tables before doing the combining operation.

Structure:  The structure of the original net and the additional net do not have to be the same, but they must be compatible.  If in doubt, try combining them and observing the results obtained.  You may want to adjust the structure of one or both nets before-hand to make them compatible.  For that purpose you can use link reversals, or even better, you can construct the CPTables with Build From Other Net.

Degree:  The way that Netica uses the degree as an experience multiplier is:  If the degree is less than one it multiplies all experience numbers of the additional net by the degree, and if it is more than one, it multiplies all the experience numbers of the original net by 1/degree.  So experience numbers are never increased, and the magnitude of the degree indicates how much more reliable the additional net is compared to the original net.

Alternative:  Another way to combine nets is to copy  nodes from the additional net, and then paste them into the original net.  You might have to hook up some disconnected links.  That method will not combine CPT tables.