Current methods for generating qualitatively different plans are either based on simple randomization of planning decisions and so cannot guarantee meaningful differences among generated plans, or require extensive user involvement to drive the system into different sections of the overall plan space. This paper presents a cost-effective method for automatically generating qualitatively different plans that is rooted in the creation of biases that focus the planner toward solutions with certain attributes. Biases are derived from analysis of a domain metatheory and enforced through compilation into preferences over planning decisions. Users can optionally direct the planner into desired regions of the plan space by designating aspects of the metatheory that should be used for bias generation. Experimental results are provided that validate the effectiveness of the biasing method for reliably generating a range of plans with meaningful semantic differences.
[1]
Greg Linden,et al.
Interactive Assessment of User Preference Models: The Automated Travel Assistant
,
1997
.
[2]
David A. Shlapak,et al.
Perspectives on Theater Air Campaign Planning
,
1995
.
[3]
David E. Wilkins,et al.
Reasoning about Locations in Theory and Practice
,
1998,
Comput. Intell..
[4]
David E. Wilkins,et al.
Practical planning - extending the classical AI planning paradigm
,
1989,
Morgan Kaufmann series in representation and reasoning.
[5]
Karen L. Myers.
Strategic Advice for Hierarchical Planners
,
1996,
KR.
[6]
Dana S. Nau,et al.
Semantics for hierarchical task-network planning
,
1994
.
[7]
John Levine,et al.
Generation of Multiple Qualitatively Different Plan Options
,
1998,
AIPS.