The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

Research Paper #478

Title:An Architecture Facilitating Repair and Replanning in Interactive Explanations
Date: 1990
Presented:Proceedings of the 5th UK Explanation Workshop (Ed. N Filer), pp. 1.1 - 1.8, Manchester, 1990
Abstract:This paper describes a planning architecture which allows interruptions, checking moves which ask whether an explanation is understood, and repair and replanning strategies. We define repair as recovering from some misconception or missing information which has prevented an explanation from being understood, whether the repair is initiated by the explainer or the explainee. Repair by the explainer is implemented using Moore's [4] method of inspecting the prerequisites of faulty plans. Repair by the explainee involves allowing restricted interruptions of the explanation. Replanning involves abandoning the current explanation and restarting with a different approach. It is appropriate where the cause of an explanations's failure can not be found or when repair is thought to be more costly than beginning again. Since humans are able to use repair and replanning to fix interactive explanations, explanation systems should take these strategies into account. Peachey and McCalla [5], Moore [4], and Cawsey [1] all have systems which can replan some part of an explanation. However, their systems always choose to replan as little of the explanation as possible, and will only replan when every other means of providing an explanation has failed. Human explanations appear to allow replanning more generally, depending on the estimated difficulties of possible courses of action. The system described can replan any unfulfilled discourse goal at any planning step and decides which recovery strategies to employ by evaluating the different possibilities. The replanning can also be restricted by specifying which kinds of goals and planning steps may be replanned. The planner is being implemented in a domain where one computer process explains to another in an artificial language how to navigate around a simple map.

[Search These Pages] [DAI Home Page] [Comment]