Todd Michael Mansell
This paper describes how evidential reasoning (E-R) has been used to manage uncertainty and abstraction in U-Plan, a planning system capable of utilising uncertain and incomplete information. U-Plan uses a unique possible worlds formalism to describes likely representations of the environment at multiple abstraction levels. The possible worlds are generated from information which may have been collected from disparate sources, and expressed in diverse frames of reference. This information is accompanied by a quantitative measure of belief, that is used to weight the evidence supporting each possible world state. In support of hierarchical planning, each possible world contains a description at a number of levels of abstraction. This allows strategic decisions to be made using a coarse description of the world, and tactical decisions using a detailed description. Compatibility relations from E-R provides the mechanism to represent and manipulate information at varying levels of abstraction. The architecture described in this paper also provides for planning in dynamic environments as U-Plan can rapidly assess when changes to the world, during plan generation, may invalidate existing plans.