Tech Report CS-92-61

Reasoning about Time and Probability

Keiji Kanazawa

May 1992

Abstract:

Reasoning about Time and Probability

Keiji Kanazawa

Decision making requires reasoning about chance and change in the world: how things change over time, the chances that things change in different possible ways, and how our actions affect these chances. Together with knowledge about preferences, it is possible to choose the best action in a given situation. As an example, consider treatment planning in a hospital emergency room. It requires reasoning about the likely changes in a patient's condition, when important changes are likely to occur, and the probable effects of different treatment and testing actions.

This dissertation is an account of reasoning about chance and change for planning and decision-making. It builds upon earlier results in developing the original treatment of reasoning about time and probability in artificial intelligence. The following distinguish my approach:

--- I explicitly reason about chance with probability, and I use decision theory to choose the best action based on preferences. Probability and decision theory are mature, normative theories of reasoning about chance and choice (they tell you the right thing to do).

--- I develop logics and graph models for expressing knowledge about time, probability, and preference. Logic makes it easy to express knowledge about time and probability, and to apply this knowledge as needed. Graph models of knowledge about probability and preferences make it easy to reason about time and probability efficiently.

The program developed in my thesis, Goo, is a deductive database for knowledge about time and probability. Goo can answer queries such as ``What is the chance that I will receive the test results within half an hour?'' and ``Are there any surgical procedures where the probability of success is more than .5 for this patient?''. Goo automatically constructs and maintains graph models based on logical knowledge in response to queries and assertions.

(complete text in pdf or gzipped postscript)