Thursday 27 Jan 1994, 1:30pm, WeH 4601 Reasoning about Actions at Multiple Levels of Granularity Lonnie Chrisman ABSTRACT Action models can be used for many purposes, including their use by planning methods for figuring out what to do next. But obtaining good models of physical actions is very difficult due to the infinite complexity of the real-world. Anyone confronted with the task of designing models for a set of physical actions has two basic choices that must be made. First, one must decide what basic behavioral units to encapsulate as actions (the entities being modeled), and second, one must decide what features to include in the description of the world state. The first is the choice of level of Granularity, and the second is the choice of the level of Abstraction. It is important to realize that BOTH (even the choice of what constitutes the basic actions) are choices made by the designer. There is a fundamental tradeoff between grandularity and abstraction. If one models very fine grain actions, small details of the world have greater effects on the outcomes making it more important to include these details in our world state description. Additionally, because no model can exactly mirror the real-world, inaccuracies build up quicker with fine grain models since a greater number must be chained together to predict the outcome of a given behavioral sequence. On the other hand, if one models only coarse grained actions, then a more concise state description may suffice, but there is a greater lack of flexibility since the agent will not be able to reason about any behavioral units finer than the granularity of its models. Because the choice of any single level of abstraction and granularity comes with disadvantages, it is desirable to be able to model behavioral units at multiple models of granularity, and to be able to mix these models freely during reasoning. In this way, models at different levels of granularity can capture distinctly different knowledge about the outcomes of an action, and that knowledge can be combined when needed. Unfortunately, few standard frameworks can handle multiple models at multiple levels of granularity. For example, standard probabilistic action models of the form of Pr(outcomeState|action,prestate) encounter logical contradictions (termed incoherence) when multiple models are mixed. In this talk I will discuss the problems with modeling physical actions and motivate the need for using multiple models at multiple levels of granularity. The problems with standard probabilistic models will be examined, and I will present the beginnings of an alternative approach for modeling action outcome uncertainty based not only the laws of probability but also on an explicit account of abstraction. I will utilize experiments performed within this framework to explicitly demonstrate the tradeoff between abstraction and granularity.