Show 
Probabilistic Inference by Node Absorption

It is possible to do probabilistic inference using node absorption, by entering all the findings, and then absorbing all the nodes except for a single query node.  The resulting probability CPT for that node will be a single belief vector (because the node won’t have any parents), which is the same as the beliefs that would be obtained by compiling the Bayes net and doing belief updating.  Of course, the net is destroyed in the process, but you can recover it by choosing Edit Undo.  Normally it is better to do inference by compiling the net and doing belief updating, but sometimes additional insights are gained by using node absorption for inference.

It should be mentioned that node absorption will also work with decision nets.  To solve a decision net, select all its nodes, and click the image\StarTOOL.gif  tool button.  The nature nodes will all be absorbed out.  When a decision node is absorbed, it is not removed from the net; instead it is completely disconnected and its decision table set to the optimal decision function.  Utility nodes are also left, so you can see the expected utility.  The algorithm used is described in Shachter86, Shachter88 and Shachter89.

When using node absorption to solve decision problems, the decision nodes must have no-forgetting links.  There must be only one utility node, with no descendants.  And if not all the nodes of the net are being absorbed, they must consist of a descendant subnet.  So, if nodes are absorbed one-by-one, a suitable order must be used.  These restrictions are explained in more detail in the Shachter references mentioned above.  If any of these restrictions are not met, Netica will not produce an erroneous result, but will just absorb as many of the nodes as it can, and then display a message explaining why it was impossible to proceed.
 

Home > Probabilistic Inference by Node Absorption