-
assigned issue to
Default potential in super-value nodes
[This issue is related with the issue "Range of utility nodes" that was open in org.openmarkov. I've just eliminated it, and the problem is explained here]
I think we discussed this some time ago, but I'd like to clarify the things before the release.
The default potential in super-value nodes in the GUI is Uniform. That creates problems in the inference, because no function is defined for Uniform potentials in super-value nodes. I propose to have default Sum potential in super-value nodes. I suspect that can be difficult to be managed in the GUI, but sum is the implicit operation among utility nodes in influence diagrams without having super-value nodes. We'll therefore being coherent with that.
If we decide that the default potential in super-value nodes is Uniform we then should consider two things: 1. Uniform is not a valid potential in a super-value node. In my opinion, it is not very user-friendly that the program proposes default values in the potentials that are not valid. 2. The inference should return an exception when the user tries to evaluate a network that has invalid potentials.
We can discuss this on the next meeting.
Comments (5)
-
-
- changed status to resolved
Resolved. Supervalue nodes now have table potential with zero variables when have only utility parents. When parents are a mixture of numerical variables and and utility nodes then potential is a sum. When parents of supervalue nodes are a mixture of utility nodes, numerical, finite states or discretized, then the potential is uniform.
-
Was resolved when adding a new parent but not when changing domain of a node that has a utility child.
-
- changed status to open
-
- changed status to resolved
Resolved
- Log in to comment
We have agreed in today's meeting that initially the potential will be a table (a table of 0 variables is a constant). If the user adds a numerical variable as a parent of this one, the potential will be turned into a sum. If all the numerical parents are removed, the potential will be turned into a table again.
If the parents of a utility node are a mixture of finite state and numeric variables, then the potential will be uniform and the inference algorithms will throw an exception. This behavior will change when we implement LinearCombinationPotential.