Abstract Actions for Stochastic Planning

Jonathan King Tash
Group in Logic and the Methodology of Science
University of California
Berkeley, California 94720
tash@cs.berkeley.edu

Abstract

This paper presents a method for abstracting action representation, reducing the computational burden of stochastic planning. Previous methods of forming abstractions in probabilistic environments have relied on summarizing clusters of states or actions with worst-case or average feature values. In contrast to these, the current proposal treats abstract actions as plans to plan. An underspecified action sequence is used in abstract plans like the expected consequences of its realization. This more accurately reflects our natural use of abstract actions, and improves their utility for planning. An exemplification of this idea is presented for maze route-finding modeled as a Markov decision process.

Appeared in

1995 AAAI Spring Symposium Technical Report on Extending Theories of Action: Formal Theory and Practical Applications, pp. 184-187

Look at the Paper

Go Back to Home Page