Learning by Experimentation

Abstract

Autonomous systems require the ability to plan effective courses of action under potentially uncertain or unpredictable contingencies. Effective planning requires knowledge of the environment, and if the environment is too complex or changes dynamically, goal-driven learning with reactive feedback becomes a necessity. This paper addresses the issue of learning by experimentation as an integral component of PRODIGY, a flexible planning system augmented with capabilities for execution monitoring and dynamic replanning upon receiving adverse feedback. In particular, experiment formulation seeks to acquire precisely the domain knowledge needed to complete a partial plan, or to correct an errant one. Thus, experimentation is demand-driven and exploits both the internal state of the planner and any external feedback received. A detailed example of integrated experiment formulation in presented as the basis for a systematic approach to extending an incomplete domain theory or correcting a potentially inaccurate one.11This research was sponsored in part by the Defense Advanced Research Projects Agency (DOD), ARPA order No. 4976, monitored by the Air Force Avionics Laboratory under contract F33615-84-K-1520, in part by the Office of Naval Research under contract N00014-84-K-0345, and in part by a gift from the Hughes Corporation. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of DARPA, AFOSR, ONR, or the US government. The authors would like to acknowledge the other members of the PRODIGY project at CMU: Oren Etzioni, Craig Knoblock, Dan Kuokka, Steve Minton, Henrik Nordin and Ellen Riloff.