lansky@VENICE.AI.SRI.COM (Amy Lansky) (12/04/87)
COMPOSING AND DECOMPOSING UNIVERSAL PLANS
Marcel Schoppers
Advanced Decision Systems (MARCEL@ADS.ARPA)
11:00 AM, MONDAY, December 7
SRI International, Building E, Room EJ228
``Universal plans'' are representations for robot behavior; they are
unique in being both highly reactive and automatically synthesized. As
a consequence of this plan representation, subplans have conditional
effects, and hence there are conditional goal conflicts. When block
promotion (= subplan concatenation) cannot remove an interaction, I
resort not to individual promotion (= subplan interleaving) but to
confinement (falsifying preconditions of the interaction). With
individual promotion out of the way, planning is a fundamentally
different problem: plan structure directly reflects goal structure,
plans can be conveniently composed from subplans, and each goal
conflict needs to be resolved only once during the lifetime of the
problem domain. Conflict analysis is computationally expensive,
however, and interactions may be more easily observed at execution
time than predicted at planning time.
All conflict elimination decisions can be cached as annotated
operators. Hence it is possible to throw away a universal plan, later
reconstructing it from its component operators without doing any
planning. Indeed, an algorithm resembling backchaining mindlessly
reassembles just enough of a universal plan to select an action that
is helpful in the current world state. Since the selected action is
both a situated response and part of a plan, recent rhetoric about
situated action as *opposed* to planning is defeated.
VISITORS: Please arrive 5 minutes early so that you can be escorted up
from the E-building receptionist's desk. Thanks!