Utility Based Termination of Anytime Algorithms

This paper presents a method for automatically deriving a time and content dependent information value function for probabilistic information This function de scribes analytically what real world value an agent can obtain by using a certain piece of information at a certain time The general form of this function is formulated and a speci c example with two valued outputs is presented in the factory scheduling domain The information value function forms a formal basis for decision theoretic deliberation control because the control decisions can be made in order to maximize a value di rectly derived from the agent s situation in its environment We show how an expert agent can use another agent s communicated information value function to allocate the right amount of time to an anytime algorithm whose results the other agent will use The particular anytime algorithm is generated from an incomplete algorithm by using prior execution statistics This method enables a rational use of incomplete algorithms that are often e ective but su er from not halting on every input As an example we present an anytime algorithm for determining satis ability SAT The algorithm approximates SAT probabilistically by re ning a satis ability probability estimate over time To enhance accuracy both the initial satis ability estimate and the performance pro le of the anytime algorithm are parameterized by problem instance features The result of each execution step of the algorithm is used to dynamically predict the results of future execution steps