Abstract

This paper addresses a computational approach to the automated acquisition of domain knowledge for planning systems via experimentation with the environment. Previous work showed how existing incomplete operators can be re ned by adding missing preconditions and e ects. Here we develop additional methods such as direct analogy to acquire new operators, decomposition of monolithic operators into meaningful sub-operators, and experimentation with partially-speci ed operators.
In order for autonomous systems to interact with their environment in an intelligent way, they must be given the ability to adapt and learn incrementally and deliberately. Our approach is to improve initially hand-coded models for planning by failure-driven experimentation with the environment. Incompleteness is one possible fault in the given domain model. The initial domain may contain incompletely speci ed operators, and may be missing operators for legitimate actions that the planner can use to achieve goals. In previous work [1], we described the use of experimentation within the(ORM) to nd new preconditions and e ects of existing operators. Shen and Simon [3, 4] describe methods to learn new operators: by exploring available actions whose e ects are unknown, or by splitting an existing operator into two di erent ones when an expectation failure occurs. This paper presents some additional methods to acquire new operators. The rst method presented is based on constructing new operators by direct analogy with existing ones through the types of the objects that they are applied to. Then we show how to create micro-operators, which contain only some …

Date

March 16, 1993

Authors

Yolanda Gil

Journal

Proceedings of the AAAI Workshop on Learning Action Models, Washington, DC

Pages

1-4