Multi-Tier Communication Abstractions for Distributed Multi-Agent Systems Michael Thome [email protected].

Multi-Tier Communication Abstractions for Distributed

Multi-Agent Systems

Michael Thome [email protected]

1 Oct 2003 KIMAS03 2

Motivation

• Problem:– Applications naturally imply multiple levels of

interaction abstractions– Cramming multiple levels into a single

implementation protocol is inefficient– Optimal use of CPU versus Bandwidth is difficult

with a single solution

• Solution:– Parallel mechanisms tailored to different

abstractions– Coordinated, interacting mechanisms allow for

balanced resource use

1 Oct 2003 KIMAS03 3

Observations

An Agent as a model of a “real” entity:

Coarse-grain, asynchronous/loosely-coupled, addressable

An Agent as a model of a business logic element:

Fine-grain, tightly-coupled, anonymous

A model of interacting element sets:

Group semantics, group addressable, peering

1 Oct 2003 KIMAS03 4

The Cougaar Solution

• A Plugin is:– Atomic unit of business logic/opaque processing

• An Agent is:– A set of Plugins– An identity (name, address, db key, creds)– A common infrastructure

• A Community is:– A named, addressable group of agents– Maintained by another agent as a service

1 Oct 2003 KIMAS03 5

Technical approach

• Plugin– Tightly coupled, sequential, influence restricted to an agent– Implemented using shared memory, transactional

blackboard

• Agent– Agents generally are loosely-coupled, operate in parallel,

interacts as peer to other agents– Implemented using message passing

• Community– Partially-specified addressing, collective behavior– Group-level semantics and dialects– Implemented as an orthogonal application

1 Oct 2003 KIMAS03 6

Example: a Large Cougaar application

• Entities:– 40 CPUs– 589 Agents– 12465 Plugins (~50 classes)

• Interactions:– 670K agent-agent messages– 16M plugin-plugin interactions (~24i/m)

1 Oct 2003 KIMAS03 7

Example: an Agent network

1 Oct 2003 KIMAS03 8

Example: a Plugin network

1 Oct 2003 KIMAS03 9

Example: Communities of Agents

1 Oct 2003 KIMAS03 10

Conclusion

A hybrid of shared-memory blackboard interactions for tightly coupled members

and distributed message-passing interactions for loosely coupled

members is an effective and efficient solution for complex DMAS applications

1 Oct 2003 KIMAS03 11

For more information…

• BBN Technologies: http://www.bbn.com• Cougaar: http://www.cougaar.org• CougaarForge: http://cougaar.org• UltraLog: http://www.ultralog.net• Other KIMAS03 papers:

– An Infrastructure for Adaptive Control of Multiagent Systems, Kleinmann, et al

– Multi-resolutional Knowledge Representation for Logistics Systems using Prototypes,Properties and Behaviors, Berliner, et al

1 Oct 2003 KIMAS03 12

Backup slides

1 Oct 2003 KIMAS03 13

What’s a Cougaar?

• Primary Application– Military Logistics Planning and execution– Very deep, wide, and complex domain

• Cougaar– DMAS infrastructure tuned for app– Java, OSS, general purpose

• UltraLog– Making Cougaar Applications survivable– Robust, secure, Scalable… adaptively

1 Oct 2003 KIMAS03 14

Alternate Community view