Traditional techniques in artificial intelligence reflect
the fact that early models of behavior assumed actors
were isolated thinkers. These techniques have needed to
be refined or replaced as research focus has shifted towards
groups of collaborating agents who coordinate their efforts
via communication. An important consideration for a collaborative
model is that communication meant to prevent unnecessary
action effort can take longer than the unneeded actions
would have; in the worst case, a near- constant stream
of lengthy conversations can effectively preclude any
action. The work described in this talk extends traditional
techniques in two important ways. First, an activity model
is dgeloped that reduces the amount of information exchanged
and the frequency that communication occurs. Second, agents
are equipped with a memory-based learning framework that
allows the agents, over time, to better coordinate despite
limited communication. A central feature of the learning
technique is that agents independently learn by analyzing
their run-time behavior. Empirical results from an implemented
test-bed verify the efficacy of the technology developed.
An important issue that distinguishes a multiagent system
from a traditional (single agent) one is determining how
the group of agents will coordinate their activities.
Early multiagent planners made sure that agents were properly
coordinated by either having a centralized planner assigning
plans to the agents or by having the agents exchange plan
structures during planning. In either case, the plans
were generated and the execution was assumed. In current
research on autonomous agents, where run-time conditions
are assumed to be uncertain and dynamic rather than pre-determinable
and static, a model of behavior that depends on the creation
of an overarching plan at the outset of a cooperative
activity becomes problematic.
Over the last decade, sound theoretical frameworks have
been developed that specify exactly what information needs
to be communicated during the course of joint activities
in order to maintain coordination at all times. Although
these models "guarantee" coordination, the communication
costs they entail are potentially impractical for dynamic
environments. A critical issue, therefore, is developing
a multiagent system that allows multiple agents to remain
coordinated, while keeping communication costs manageable.
Learning in a dynamic, uncertain, multiagent setting
is a challenging task. Our learning techniques are motivated
by the theory that the development of distributed cooperative
behavior in people is shaped by the accumulated cultural-
historical knowledge of the community. An agent's memory
contains the breadth of knowledge acquired through interacting
with other agents and the world during the course of solving
problems. The cornerstone of the memory-based learning
framework of the agents are case-based reasoning techniques
to convert noisy run-time activity into procedures useful
for future problem- solving activities. The coordinated
procedures created by this conversion process are stored
into memory by each agent independently. Analyzing run-time
performance, rather than planner performance, enables
agents to learn procedures beyond the scope of their first-principles
planner rather than to simply cache previously known ones.
The models of behavior and learning described above have
been implemented into a test-bed system, composed of over
25,000 lines of source code. Empirical studies verify
that communicating only coordination points is advantageous
when communication costs are high and that our learning
techniques are effective at reducing the run-time effort
expended by the community of agents to solve problems.
