An interactive approach to building an influence diagram based on neural networks

Abstract

Building an influence diagram in decision analysis is known to be a most complicated and burdensome process. The use of neural networks to generate influence diagrams in the topological level results in a good performance, but the generated ID is not usually a well—formed influence diagram. It needs futher modification to be applicable to real decision problems, especially when group decision participants are involved. This research suggests an interactive procedure to build a well—fanned influence diagram from the initial influence diagram generated from neural networks, which is thought to be an approximation of experts (explicit or implicit) interpretation of the decision problem. Our procedure is composed of three phases: the first phase is to generate an initial influence diagram from trained neural networks, the second is modification of the influence diagram by each domain expert, and the third is resolution of the differences of the groups influence diagrams interactively. We explain our procedure by applying it to a land development and conservation problem. When the problem is more complicated and group decision participants are involved, this research is expected to be more useful to model a decision problem inexpensively with less time and effort.