next up previous contents
Next: A Human Interface for Up: Evaluating Qualitative Robot Reasoning Previous: Reasoning with Incomplete Information

Providing Graceful Degradation

In the event that there is insufficient information for a robot reasoning system to continue with its task, it should be able to ``degrade gracefully''. There are several components to graceful degradation (which is actually an objective for any well designed computer system):

There are graceful degradation facilities built into the PDO/EPB representation - in particular the partial distance ordering itself, and the ``vague direction'' handling features. In either directional or magnitude reasoning, information of varying precision can be accomodated, and the system is always aware of the possibility that the information available may not be sufficient. The system is able to distinguish between reasoning failures caused by insufficient information, and failures for other reasons, because all the qualitative geometry functions can return tokens to represent ``unknown'' as a valid result.

These facilities mean that the system can notify an operator of failure resulting from incomplete information. It can continue reasoning as normal, with the unknown token only propagating into areas that depend on the missing value. When this value does become known it cannot automatically be substituted back along the path where it was propagated (which might be an ideal for recovery after graceful degradation), but the parts of the reasoning process which need to be re-run can be directly identified.

Graceful degradation was one area in which the ASSF representation was noticeably inferior to the PDO/EPB representation. Because it depended on chains of relationships to establish relative positions or orientations of any two parts, it was unable to continue reasoning if any piece of information was missing. This handicap arose mainly because the way in which numeric descriptions were converted to qualitative ones was clumsy (especially the use of axis length to create local quantity spaces, and the axis-relative location descriptions).


next up previous contents
Next: A Human Interface for Up: Evaluating Qualitative Robot Reasoning Previous: Reasoning with Incomplete Information
Alan Blackwell
2000-11-17