Apr 10, 2007: Non-Markovian Control in Dynamical Systems
Filed in: Colloquium
Dr. Alfredo Gabaldon, University of New South Wales
It has long been recognized in the field of Artificial Intelligence that one of the fundamentalfeatures of an autonomous intelligent system is the ability to reason about its changing environment. For a number of years, AI researchers have been working on the challenging problem ofdeveloping unifying mathematical and computational foundations for modeling dynamical systems. Reiter's Basic Action Theories is one of a number of prominent formalisms that have been advanced with this goal in mind. In this talk, I will first describe a generalization of Reiter's formalism in order to capture non-Markovian dynamics, where the preconditions and effects of events may depend not only on the current but also on previous states of the system and its environment. I will then describe a procedure for incorporating into a system description, search control knowledge of the form used by two of the most successful planning systems currently available. The procedure is designed based on the generalized framework and the benefitobtained from it is twofold: 1) it clarifies the relationship between search control knowledge as used by the planners andan underlying dynamical system, and 2) it allows a more efficient utilization of the search control knowledge in planning.
Alfredo Gabaldon is a Research Scientist at National ICT Australia and a Visiting Research Fellowat the School of Computer Science and Engineering at the U. of New South Wales, both since 2004. He holds a Ph.D. in Computer Science from the U. of Toronto and B.S. and M.S. degrees from theU. of Texas at El Paso. He is interested in the formalization and modeling of various aspects of AI dynamical systems, such as reasoning about actions, planning, reasoning about knowledge and probability, narrative and temporal reasoning, database transactions and update, and concurrencyin continuous domains, among others.
Abstract
It has long been recognized in the field of Artificial Intelligence that one of the fundamentalfeatures of an autonomous intelligent system is the ability to reason about its changing environment. For a number of years, AI researchers have been working on the challenging problem ofdeveloping unifying mathematical and computational foundations for modeling dynamical systems. Reiter's Basic Action Theories is one of a number of prominent formalisms that have been advanced with this goal in mind. In this talk, I will first describe a generalization of Reiter's formalism in order to capture non-Markovian dynamics, where the preconditions and effects of events may depend not only on the current but also on previous states of the system and its environment. I will then describe a procedure for incorporating into a system description, search control knowledge of the form used by two of the most successful planning systems currently available. The procedure is designed based on the generalized framework and the benefitobtained from it is twofold: 1) it clarifies the relationship between search control knowledge as used by the planners andan underlying dynamical system, and 2) it allows a more efficient utilization of the search control knowledge in planning.
Bio
Alfredo Gabaldon is a Research Scientist at National ICT Australia and a Visiting Research Fellowat the School of Computer Science and Engineering at the U. of New South Wales, both since 2004. He holds a Ph.D. in Computer Science from the U. of Toronto and B.S. and M.S. degrees from theU. of Texas at El Paso. He is interested in the formalization and modeling of various aspects of AI dynamical systems, such as reasoning about actions, planning, reasoning about knowledge and probability, narrative and temporal reasoning, database transactions and update, and concurrencyin continuous domains, among others.