a-teamsusers.ece.cmu.edu/~talukdar/ateams.pdf · 1998. 11. 19. · institute for complex engineered...
TRANSCRIPT
-
Institute for ComplexEngineered Systems
Institute for ComplexEngineered Systems
Carnegie Mellon UniversityCollege of Engineering
1201 Hamburg HallPittsburgh, PA 15213-3890
ph: 412.268.3372fax: 412.268.5229
http://www.ices.cmu.edu
CarnegieMellon
WHAT IS AN ASYNCHRONOUS TEAM?Asynchronous teams (A-Teams) are open, high-per-formance organizations for solving difficult prob-lems, particularly, problems from the areas of plan-ning, design, scheduling and real-time control. A-Teams get their openness (ease of assembly) from theuse of autonomous agents that can be added orremoved without any supervisory system getting inthe way. A-Teams get their high-performance (quickconvergence to good solutions) by exploiting multi-plicity (large numbers of agents, computers, andother resources). Indeed, what an A-Team does well,perhaps better than any other type of organization, isexploit multiplicity.
WHERE DID THE IDEAS COME FROM?An A-Team is a combination of features taken froma variety of natural and synthetic systems, includinginsect societies, cellular communities, brainstormingprotocols, blackboards, genetic algorithms, simulat-ed annealing, and tabu search.
HOW DOES AN A-TEAM WORK?Structurally, an A-Team is a strongly cyclic networkof autonomous agents and shared memories. Eachagent contains some problem-solving skills; eachmemory contains a population of trial- solutions toa part of the problem-to-be-solved. The trial-solu-tions circulate continually. Agents cooperate byselecting and modifying these solutions as they cir-culate. The modifications can be constructive (animprovement to a part of the selected solution), ordestructive (the erasure of the selected solution). Allthe agents can work asynchronously (each at its ownspeed) and in parallel. Some of the agents are simple,others can be complex, some of the agents can becomputer-based, others can be human.
An A-Team’s structure allows for multiplicity to thepoint of excess, along five dimensions: 1) multiplic-ity of skills, so no relevant skill, no matter how spe-cialized or redundant it may seem, need be excluded;2) multiplicity of agent-types, so a wide variety ofagents, including software and human agents, can beaccommodated; 3) multiplicity of representations, sothe different representational needs of all the agentscan be met; 4) multiplicity of computers, so all theagents can work in parallel, whether they are co-located or distributed; and 5) multiplicity of trial-solutions, so a suitably wide range of promisingalternatives can be simultaneously considered.
Both experiment and theory confirm two key prop-erties. First, construction and destruction are com-plementary processes in an A-Team: adept destruc-tion can compensate for inept construction, andvice-versa. Second, there are profound benefits-of-scale, meaning that the performance of an A-Teamcan invariably be improved by expanding it alongone or more of its five "multiplicity-dimensions."
APPLICATIONSA-Teams can be applied to both off-line problems,like planning, design and scheduling, and real-timecontrol problems, like the operation of electricpower or packet switching networks. At CMU, A-Teams have been developed for traveling salesmanproblems, high-rise building design, reconfigurablerobot design, fault diagnosis, operation planning,and job shop scheduling. Work is now underway todevelop a an A-Team whose agents will be distrib-uted over an electric power network and will contin-ually learn how to do their control jobs better. AtIBM, A-Teams have been used for steel and papermill scheduling.
FOR FURTHER INFORMATIONContact Sarosh Talukdar at 412 268 8778 or [email protected]. Links to papers on A-Teams can befound at: www.ece.cmu.edu/~talukdar/
Sarosh N. Talukdar
A-TEAMSHow Distributed, Autonomous Agents Can Cooperate to
Solve Both Off-Line and Real-Time-Control Problems
For More Information:
Sarosh Talukdarph: 412.268.8778
email: [email protected]://www.ece.cmu.edu/
~talukdar