COLLABORATIVE DEVELOPMENT OF LARGE BAYESIAN

unrestricted directed acyclic graphs. Creating BN models of this type is difficult, time consuming, and error prone. While the graph basis for the data naturally leads to a graphical editor, it is difficult to work with models larh layout tasks become difficult, consuming time that could be spent expanding the model.
INTRODUCTION
Bayesian networks (BN) have been shown to be a very effective form of models for diagnostic assistants, [3]. One of the main problems in development of such systems for complex domains is an efficient design of the BN model. The models are graphs of nodes and arcs, annotated with probabilities, [7]. Models of complex domains can consist of thousands of nodes and directed arcs and tens of thousands of parameters. The graph structures and node definitions are typically assumed to be
COLLABORATIVE DEVELOPMENT OF LARGE BAYESIAN NETWORKS
K. Wojtek Przytula HRL Laboratories, LLC 3011 Malibu Cyn Rd. Malibu, CA 90265 310.317.5892 wojtek@
Gerald B. lsdale HRL Laboratories, LLC 3011 Malibu Cyn Rd. Malibu, CA 90265 310.317.5883 gbisdale@
Tsai-Ching Lu HRL Laboratories, LLC 3011 Malibu Cyn Rd. Malibu, CA 90265 310.317.5187 tlu.fn@
1-4244-0052-X/06/$20.00 ©2006 IEEE
515
One way to overcome the difficulties and cost of conventional BN construction is to generate BN automatically. For many diagnostic domains there already exist troubleshooting flow diagrams, which contain most of the information needed to create initial form BN models. The automatically created initial form networks can be later expanded manually or updated by learning from data, [9]. Another useful representation of information for diagnostic BN models is a fault tree. The trees along with lists of tests and observation can also be automatically converted into diagnostic BN models, [15].
Alternative way to accelerate development of diagnostic BN is by means of graphical templates. Here, the network is assembled using domain specific templates, which are part of a network development software tool. An example of such a tool for medical diagnosis in outpatient clinic is Promedas (PRObabilistic MEdical Diagnostic Advisory System), [4]. The approach proposed in this paper simplifies model development by using a layered structure of BN graph with "noisy" node types. This structure and nodes significantly reduce the complexity of model creation, verification and reasoning on the models. It has been shown that quality of reasoning in such layered networks is not significantly degraded,[8]. Moreover, the structure makes it possible to create large networks by automatically merging multiple smaller networks. These smaller networks can be created in parallel by different experts and then collaboratively integrated into a single systemwide network. For example, one expert develops model of an automobile brake subsystem, while others develop respectively models of electrical storage subsystem, engine control subsystem, etc. The model information is collected much faster and with fewer errors than in the conventional, sequential modeling, when one expert at a time works on the model of the entire domain using graphical editors
In conventional development of BN a knowledge engineer, who is intimately familiar with BN, uses a graphical editor to create model graph and to enter parameters from information provided by domain expert. The knowledge engineer and domain expert also work together on verification of model correctness. This is accomplished by examining diagnoses produced by the model for cases known to the expert. If a disagreement is encountered, the knowledge engineer identifies the part of the graph that is responsible for the error and modifies it. For large and complex system, the conventional network development requires multiple interviews with many different experts.
hundred of arcs. The constructed BN became part of a commercial diagnostic system and was deployed in the field. The network was created by a dozen of domain experts, each focusing on a subsystem of his specialty and working largely independently. The information for the subsystems was merged and converted using the editor. The initial data entry process required approximately 10 person-days. We have previously used conventional method for development of a much smaller diagnostic model (approximately 300 nodes) for a very similar domain. It involved a knowledge engineer and domain experts and required almost the same amount of labor. We are at present performing a more rigorous comparison of our approach with conventional model development.