Influence Diagram For Decision-making Nodes Aggregation Method

Decision-making is inevitable to individuals and to a country or the whole society. Decision Analysis becomes one of the hottest topics, and making decision scientifically gains more and more attention. As an effective tool of decision analysis, influence diagrams are used in decision support, predictive analysis, and knowledge discovery,etc. However, the complexity of making decision in an influence diagram is exponential, relative to the number of nodes involved in the influence diagram. Due to its accurate representation and decision-making mechanism, the traditional influence diagram can not satisfy the situation that the practical application has too many variables. At the same time, in many applications, users do not need to know the trivial relationships within a field, but only need to know the relationships among fields.Pertinent to the problems mentioned above, we proposed a method to aggregate nodes in an influence diagram based on the aggregating nodes of Bayesian network. The method makes the nodes in a same domain aggregate into a new node, and as a result, it simplifies the structure of the influence diagram. Preliminary experiments show the feasibility of our proposed method for aggregation of nodes in influence diagrams.Generally, the main contributions of this thesis can be summarized as follows:As influence diagrams contain nodes of different types, and the relationships between nodes are not pure probabilistic relationships, we need to partition the nodes according to the type. This partition does not destroy the internal structure of the influence diagram, and get two parts, the probabilistic part which can be equal to a Bayesian network and the utility part. Then we aggregate the two parts respectively.We aggregate the nodes of probabilistic part by using the method of aggregation nodes in Bayesian network. For the utility part, we discuss the conditional utility independence relationship between value nodes, and propose method to aggregate value nodes. We also propose method to refine the utility matrixes.Based on the graphic features of influences diagrams, we combine two new graphs into a new influence diagram according to the parent-child relationships among the value nodes and random nodes.