In a cooperative problem-solving environment; such as an office, a hierarchical planner can be incorporated into an intelligent interface to accomplish tasks. During plan execution monitoring, user actions may be inconsistent with system expectations. In this paper, we present an approach towards reasoning about these exceptions in an attempt to accommodate them into an evolving plan. We propose a representation for plans and domain ing about exceptions. objects facilitates reasonI. Interactive planning and tional occurrences TT. nierarchicai pianners incrementaiiy deveiop a pian at different levels of abstraction, imposing linear orderings at each stage of the expansion to eliminate subgoal interactions [Sacerdoti( 1977)) Tate( 1977)) Wilkins( 1984)]. The --_----A.: --I Ll--I--,--:--1L1----A.---_-----I Lexecubwn 01 bne plan s prmu~ive acblons must oe monitored to ensure success. Exceptions and interruptions are common occurrences, and the planner must react to new information made available during the various stages of plan construction and execution. Existing plans may require modification or new plans may have to be generated. --We are concerned with using a pianner as a support tool in a cooperative problem-solving environment such as an office [Broverman and Croft( 1985)) Croft and Lefkowitz( 1984)]. I n such an environment, the planner is not viewed as an omnipotent agent with complete knowledge of the domain and procedures for accomplishing all plan steps. Rather, it aids the user in performing correct and consistent tasks . The operation of the planner depends heavily on interaction with the user in order to allow user controi and to draw On the usels’ domain lrnowiedge. IIlteractive planners necessarily interleave plan generation and execution since user actions determine the course of future events. Previous planners have provided general replanning actions which are invoked in response to problems in iThis work is supported by the Air Force Systems Command, Rome Air Development Center, Griffiss Air Force Base, New York 13441-5700, the Air Force Office of Scientific Research, Bolling Air Force Base, District of Columbia 20332, under contract F30602-85C-0008, and by a contract with Ing. C. Olivetti & C. the plan resulting from the introduction of an arbitrary state predicate or “fact” (Hayes(1975), Sacerdoti(l977)! Wilkins(1985)]. In these systems, the replanning techniques provided do not attempt to reason about failing conditions or possible serendipitous effects of the exception. These methods simply make use of the explicitly linked plan rationale to detect problems and determine what violated goals need to be reachieved. We view this type QfrepianniQ 2s $ "rc=mrt.innatQ tactic invnlvinrr little ..-"=----a J 1-x 1-1 . "'6 ‘1 u "nx, intelligence, and reserve its use for exceptions generated by external agents2. To address the problems associated with interactive planning, we propose extending the traditional replanning approach. When a user action deviates from the planner’s predictions, the system should exploit available knowledge in an attempt to expiain the exceptionai behavior. Such a constructive approach is preferred to replanning, since replanning, in this case, would attempt to achieve goals that the user deliberately chose not to pursue. This paper discusses reasoning about exceptionai occurrences as an approach towards incorporating exceptions into a consistent plan. In the next two sections, we describe an interactive planner and the elements of our representation which are used to support the reasoning process. We then outline the types of exceptions that can occur and algorithms for handling them, within the context of an example taken from the domain of real estate. Input to our interactive planner is provided as an abstract goal specification, and the output is a partially or fully expanded procedural net, with partial temporal ordering (similar to other hierarchical planners [Sacerdoti(l977), Tate( 1977)) Wilkins( 1984)]). A procedural net contains goal nodes? action nodes? and phantom nodes (goal nodes which are trivially true), along with links representing the causal structure of the plan. Since complete expansion of the initial goal may require additional information from the user, only action nodes are considered primitive, and thus executable. 2The planner attempts to satisfy a number of agents. The user(s) are regarded as internal agents, while agents are considered to be e&err& if the system lacks a model for their behavior (e.g., the real world). 190 Planning From: AAAI-87 Proceedings. Copyright ©1987, AAAI (www.aaai.org). All rights reserved.
[1]
Josh D. Tenenberg,et al.
Planning with Abstraction
,
1986,
AAAI.
[2]
David E. Wilkins,et al.
Domain-Independent Planning: Representation and Plan Generation
,
1984,
Artif. Intell..
[3]
Philippe Coiffet,et al.
Representation of a robot
,
1983
.
[4]
W. Bruce Croft,et al.
A Knowledge-Based Approach to Data Management for Intelligent User Interfaces
,
1985,
VLDB.
[5]
G. G. Stokes.
"J."
,
1890,
The New Yale Book of Quotations.
[6]
Patrick J. Hayes,et al.
A Representation For Robot Plans
,
1975,
IJCAI.
[7]
Richard Alterman,et al.
An Adaptive Planner
,
1986,
AAAI.
[8]
Drew McDermott,et al.
Planning and Acting
,
1978,
Cogn. Sci..
[9]
Richard Fikes,et al.
A Commitment-Based Framework for Describing Informal Cooperative Work*
,
1982
.
[10]
W. Bruce Croft,et al.
Task support in an office system
,
1984,
TOIS.