ABDUCTIVE NETWORKS - GENERALIZATION, PATTERN-RECOGNITION, AND PREDICTION OF CHEMICAL BEHAVIOR

Using commercially available software, it is possible to reduce numeri cal data to a mathematical representation called an abductive network (AN). In the current communication, we describe several simple example s which illustrate the interesting, and potentially useful properties of abductive networks. We show that when applied to the correlation of Kovats indices with molecular refractivities and dipole moments of su bstituted phenols, abductive networks more accurately predict Kovats i ndices than do counter-propagation neural networks or linear regressio n equations. When applied to the modeling of quantitative structure-ac tivity relationships (QSAR) for local anesthetics, AN's are marginally superior to regression. AN's offer the advantage that correlations ma y be drawn between variables which are not easily related within a mat hematical context.