An empirical test for patterns of nonmonotonic inference

Abstract

Human inference can be used to test the inference patterns a reliable nonmonotonic consequence should satisfy, because it appears to be nonmonotonic, it is adaptive and it generally achieves efficiency. In this study, an experiment is conducted to investigate whether human inference tends to be consistent with rationality postulates (System P plus Rational Monotony), especially when it no longer satisfies the Monotony property. The experimental protocol uses a possibilistic semantics for plausible rules. Our results appear to be consistent with all the studied properties. Exceptions are the Cut property (with one kind of content out of two) and Left Logical Equivalence which could not be tested. Moreover, when Monotony was not satisfied by participants' inferences, Cut, Cautious Monotony and And properties were corroborated (Rational Monotony was only plausibly supported and the other properties were not tested). Our results emphasize the psychological plausibility of rationality postulates and support the working hypothesis in Artificial Intelligence that System P plus Rational Monotony offer a plausible basic set of properties for nonmonotonic logics.