Speci and evaluation of preferences for planning under uncertainty

Abstract This paper describes a framework for specifying behavior in terms of preference sentences of the form "prefer a to --a if3", to be interpreted as "a is preferred to -~a when all else is fixed in any 3 world". We demonstrate how such preference sentences, together with normality defaults, may be used as constraints on admissible preference-ranking of worlds and how they allow a reasoning agent to evaluate queries of the form "would you prefer al over or2 given ¢" where al and ~r2 could be either action sequences or observational propositions. We also prove that by extending the syntax to allow for importance-rating of preference sentences, we obtain a language that is powerful enough to represent all possible preferences among worlds