Bayes nets are used to determine new beliefs (in the form of probabilities) as observations are made or facts are gathered. They are composed only of nature nodes. To form a decision net (also known as an “influence diagram”), you add decision nodes and utility nodes (also known as “value” nodes). Decision nodes, which are traditionally drawn as rectangular boxes, correspond to variables or events that you will be able to control. Utility nodes are drawn with a diamond or flattened hexagon shape, and correspond to quantities you want to maximize. Example
The decision net as a whole represents a decision or planning problem that you face, or that some other agent, often called “the decision maker”, faces. Netica can find values for the decision nodes that will result in the largest possible expected value for the utility node (or sum of them if there is more than one). This is known as “solving” the decision net. Once the net is solved, you (or the decision maker) can begin to enact the plan by taking the prescribed action for each decision as it arises.
Links that go into a decision node have a special meaning and are called informational links. They indicate what will be known at the time the decision is to be made. That is, the decision maker will know the values of all the nodes which have links into that decision node, and will not know the values of any other nodes.
Earlier it was stated that when solving a decision net Netica finds values for each of the decision nodes. But if there are some links entering a decision node, it actually finds a decision value for each possible configuration of values of the parents. This is a contingent plan, with a form like: “if I observe A=high and B=no, then I will do X, if I observe A=low and B=yes, then I will do Y, and so on. So, solving a net finds a decision function for each decision node (i.e. a function which gives a decision value for each possible set of parent values).
If there are a number of decision nodes, possibly corresponding to decisions made at different points in time, then solving the net will find a decision function for each of them, and this set of decision functions is known as a policy. It is a full conditional plan, specifying what to do in each possible contingency, based on the information that will be available.
Working through an example
may make these ideas clearer, and then you may want to create
your own.