Improved Methods for building large-scale Bayesian Networks Statement of Interest for the Third Bayesian Modelling Applications Workshop at UAI 2005 Martin Neil and Norman Fenton Agena Ltd and Risk Assessment and Decision Analysis Research (RADAR) group, Department of Computer Science, Queen Mary, University of London 1. Background Agena Ltd and the RADAR group have been applying Bayesian Networks (BNs) to risk assessment problems in a variety of problem domains for the last five years. Prominent areas in which we have applied BNs include: • Predicting the risk of mid-air collisions between aircraft in UK airspace [17]; • Predicting software defects in complex consumer electronic devices [2,6,8,9]; • Evaluating the reliability and availability characteristics of military systems [15]; • Modelling the warranty return rates of electronic components [3]; • Modelling operational risk in financial institutions and predicting resulting losses. [14] All of these examples involved building large-scale BN models for a real end-user. To support this work we have developed a general BN decision support tool called AgenaRisk [3]. In addition to this we have also developed a number of heuristics that have been successfully deployed and validated in these application areas. We would hesitate to claim these heuristics constitute a methodology, but are confident that they have provided, and will continue to provide, a productive and reproducible means for modelling uncertainty in risky domains involving limited or ambiguous data and the consequential requirement to accommodate expert judgement. Our heuristics address two different, but universally applicable steps, encountered when using BNs: • Deciding the conditional relations embodied in a Directed Acyclic Graph (DAG) — we address this using the idea of idioms. • Completing the node probability tables (NPTs) — for particular classes of NPTs we use ranked nodes. 2. Idioms When working with experts in a variety of domains a series of fundamental questions arose concerning the structure of the DAG: • Which edge direction to choose? What was the cause and which the effect? • Whether some of the statements they wished to make were actually uncertain and if not whether they could be represented in a BN?