Magazine article AI Magazine

Acquiring Planning Knowledge Via Demonstration

Magazine article AI Magazine

Acquiring Planning Knowledge Via Demonstration

Article excerpt

In the mid to late 1980s there was a flurry of papers using explanationbased techniques to learn how to perform complex actions by observing (or interpreting descriptions of) human performance. These techniques were shown to work reasonably well with one or a small number of examples. However, as statistical approaches gained in power and popularity, and some kinds of data became more plentiful, machine learning as a field moved away from this kind of learning, which requires strong domain models. Recently, efforts have begun to look again at explanation-based learning and other approaches in contexts where few examples can be gathered. This workshop focused specifically on techniques for learning planning and procedural knowledge from single demonstrations, a task made more difficult by the fact that some contingency situations and what to do about them are never demonstrated.

The guest speaker was James Allen, who gave an extended talk on the procedural learning on the web (PLOW) system that was also described in his team's outstanding award-winning paper during the main conference. PLOW learns by combining observation of user-demonstrated web procedures with an interpretation of the user's natural language narration of that procedure. This enables learning from one example by filling in critical details about such things as the criteria for decisions about alternate courses of action, when to end loops, and the semantic relationships among parameters to service calls. A similar set of themes was echoed in the talk of James Blythe describing ISI's Tailor system. Here the natural language instruction was replaced by a more structured user interface in order to learn information-gathering procedures. Walsh and Littman also described the modeling, planning, and execution algorithm they call MOPLEX, a system that learns a conceptual model of information query procedures.

Mark Burstein presented an overview of plan order induction by reasoning from one trial (POIROT), a system being developed as part of DARPA's Integrated Learning program that uses a shared blackboard of hypotheses generated by different learning systems observing the same demonstration as the basis for its procedure learning. By combining the hypotheses produced by more bottomup, inductive techniques such as that presented by Tim Oates and Fusun Yaman of the University of Maryland, Baltimore County on the web initiative in teaching (WIT) system, and more top-down explanation-based techniques found in LIGHT, a system from Stanford's ISLE, POIROT is able to produce reasonably complete procedures from one example. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.