Verifying Parameters of Bayesian Belief Networks by Exploring the Impact Intensity
Baofeng Guo · 2006
Constructing a Bayesian belief network (BBN) consists of two main tasks, which we called, structure design and parameter design respectively. The structure design decides the network's topology and the parameter design, the conditional probability for each node. Correspondingly, the verification of a Bayesian network has two parts, namely structure verification and parameter verification. Basically, the structure verification is relatively easier because it is not difficult to elicit such knowledge by experts. But the parameter verification is quite difficult especially when network is becoming large. In this paper, we discuss the problem of parameter verification by investigating BBN from two aspects, i.e., impact direction and impact intensity. We propose the concept of impact intensity coefficient to characterize the implicit relationship between the note probability tables (NPTs) and nodes' impact intensity. Through two simplified examples, we found that it is potential to extract the impact intensity coefficients (IICs) by analyzing NPTs' numeric characteristics. Based on these coefficients, we devise a set of heuristic rules that offer a simple way to verify BBN's parameters.