Adele E. Howe
Computer Science Department
Paul R. Cohen
Experimental Knowledge Systems Laboratory
Colorado State University
Fort Collins, CO 80523
howe@cs.colostate.edu
\
Department of Computer Science
University of Massachusetts
Amherst, MA 01003
cohen@cs.umass.edu
AI systems in complex environments can be hard to understand. We present a simple method for finding dependencies between actions and later failures in execution traces of the Phoenix planner. We also discuss failure recovery analysis, a method for explaining dependencies discovered in the execution traces of Phoenix's failure recovery behavior.
Dependencies are disproportionately high co-occurrences of particular precursors and later events. For the execution traces described in this paper, the precursors are failures and failure recovery actions; the later events are later failures. In complicated environments, it can be difficult to know whether actions produce long-term effects, in particular, whether certain actions cause or contribute to later plan failures. Statistical techniques such as those discussed in this paper can help designers determine how recovery actions affect the long-term function of a plan and whether recovery actions are helping or hindering the progress of plans.