Inducing safer safety trees

Inductive learning techniques offer the potential for learning to classify whether a given situation is safe or unsafe based on past incident data. The most common learning algorithms are developments of ID3, CART, or AQ and include systems such as ASSISTANT-86 (I. Konenko, I. Bratko, 1986), LAIS (M.F.S. Smith, J.H. Donald, 1992) and C4.5 (J.R. Quinlan, 1992). We examine the problems of using such algorithms for learning whether a situation is safe or unsafe. We show that the primary problem with the use of these algorithms in the area of safety remains the manner in which numerical attributes are discretized