KOO@SUSHI.STANFORD.EDU (Charlie Koo) (07/04/87)
A Model for Distributed Performance --
Synchronizing Plans among Intelligent Agents
via Communication
Charles C. Koo
July 8, Wednesday
9:00am - 10:00am
Room 352
Margaret Jacks Hall
In a society where a group of agents cooperate to achieve certain goals,
agents perform their tasks based on certain plans. Some tasks may interact
with tasks done by other agents. One way to coordinate those tasks is to let
a master planner generate a plan to begin with, and distribute tasks to
individual agents accordingly. However, there are two difficulties
with this approach, given that agents are resource-limited. First, the
master planner needs to know all the expertise that each agent has. The
amount of knowledge sharply increases with the number of specialties.
Second, the centralized planning process takes longer turn-around time than
if each agent plans for itself. This causes a lot of computing resources
not being utilized. Thus, distributed planning is desirable.
In this presentation, I will describe a model for synchronizing and
monitoring plans autonomously made by intelligent agents via communication.
The model suggests an planning algorithm that allows agents to plan in
parallel and then synchronize their plans via a commitment-based
communication vehicle. Represenation as well as reasoning issues in the
distributed environment will be addressed.
Communication plays an integral role in planning for synchronization
purposes. The communication vehicle includes a minimal set of protocols
that enables the synchronization, a set of communication operators and a
set of commitment tracking operators. The tracking operators provide means
to monitor the progress of plan execution, to prevent delays, and to modify
plans with less effort when delays happen. A deadlock detection scheme will
also be described.