MultiAgent Dynamics Laboratory

We summarize the research interests of Dr. José M Vidal and the Multiagent Dynamics Laboratory (MDL) as of November 2002.

You can find these slides at http://jmvidal.cse.sc.edu/talks/researchsum/

We study the emergent dynamics of multiagent systems in order to devise ways of engineering them from first principles.

• We are interested in finding ways to build adaptive multiagent systems whose behaviors can be predicted and which can satisfy any given requirements specification.
• We use these tools in order to understand the behavior of these systems at the theoretical level:
  • Machine Learning
  • Complexity theory
  • Game theory
  • Economics
  • Agent-based modeling
  • Operations research
  • Cybernetics
  • Management Science
• We use these technologies in order to implement and test our ideas:
  • Distributed programming (DCOM, CORBA, RMI, SOAP, Jini).
  • Peer-to-peer systems (JXTA).
  • Agent communication languages (FIPA-ACL)
  • Ontology languages (DAML)
  • Agent description languages (RDF, DAML-S)
• The applications we develop fit within Berners-Lee's Semantic Web [1] vision.
• See our Engineering Adaptive Multiagent Systems [2] whitepaper.

• Human cognitive expansion via human-agent and agent-agent cooperation.
• Immune systems for computers: security as an emergent property.
• The world-wide mind (WWM) where the aggregate knowledge of the world's experts can be queried at any time [When brains meet computer brawn [3], Cnet news][NSF/DOC report [4]].
• Self-organizing micro-robotic swarms that do your bidding.

Yes, this is science fiction.

• Research projects are roughly divided into two areas:
• Agent-Oriented Software Engineering
• Multiagent dynamics

• PhD thesis by Paul Buhler.
• A component-based agent architecture that will enable agents to acquire complex behaviors at run-time.
• Synthesize DAML-S [5] service descriptions into Picolla programs which can then be executed.
• Prototype will use agentcities [6] as platform.
• Whitepaper [7]

• MS thesis by Sharad Bansal
• ServiceProfile describes what the service does in terms of its interface. It is similar to a WSDL description.
• The ServiceModel part provides a lot more details into how the service is implemented, which other services it uses, etc.
• Current query languages support search on the ServiceProfile
• Project aims to develop a query language and matching algorithm to support querying on the ServiceModel.

• by Paul Buhler (original implementation), with Shaun P. Wood and J. Vidal.
• The Biter [8] project implements a basic RoboCup [9] client. It is used in CSCE 782.
• In the Robocup competition teams play soccer in a realistic simulation.
• Agents must
  • coordinate,
  • handle noisy and limited sensors,
  • trade-off reactivity versus goal-driven behavior,
  • take timely actions.
• US-only practice competition this Fall. Anyone interested?
• José M. Vidal and Paul Buhler. Using RoboCup to teach multiagent systems and the distributed mindset [10]. In Proceedings of the 33rd ACM Technical Symposium on Computer Science Education, 2002.

• MS thesis by Hrishikesh Goradia.
• How do we expedite the construction of software agents? Can we blend components (software engineering) with BDI and subsumption architectures (agents research)?
• Develop architecture for connecting basic behaviors using Java's Bean Development Toolkit.
• Develop basic behaviors.

• MS thesis by Kapil Rajendra Dukle. Project page. [11]
• Codify, share (publish), and manipulate real-time information about football teams, statistics, and games.
• A Semantic Web application using DAML and FIPA-OS.
• NFL.com website is scrapped for timely information and the results are made available on a FIPA back-end using DAML/RDF as the content language with an football ontology developed by Kapil.
• The future: Make this service available to anyone. Create other agents that can integrate this information with other functionality. Help?

• MS thesis by Taraka Peddireddy.
• Security breaches, DoS attacks, viruses, and worms are common on the Internet.
• Fixing them requires the cooperation of many companies who might not trust each other, and overworked system administrators.
• Agents represent the interests of a company.
• Agents browse the log files, find interesting events and communicate with each other in order to diagnose problem.
• Future architecture will incorporate use of Common Vulnerabilities and Exposures [12]. Want to help?
• Whitepaper [13]
• Taraka D. Peddireddy and José M. Vidal. Multiagent network security system using FIPA-OS [14]. In Proceedings of the IEEE SoutheastCon, 2002.

• A: Set of agents.
• T: Set of tasks.
• c: cost function which maps tasks subsets to a Real.
• Goal is to allocate tasks to agent so as to minimize sum of costs.
• Examples include FedEx delivery trucks, work schedules in business, etc.
• We have found that best global solutions are found when the systems' dynamics lie at the edge of chaos.
• Looking for help.
• José M. Vidal. The effects of cooperation on multiagent search in task-oriented domains [15]. In Proceedings of the Autonomous Agents and Multi-Agent Systems Conference, 2002

• Agents in a system where agents are learning face a moving target function.
• As their behavior changes the others' goal changes, and viz.
• My CLRI theory can predict the expected error of an agent given its changing rate, retention rate, learning rate, and impact. But this in only a first step:
  • What are the systems' dynamics?
  • Does the addition of learning really make a system more robust?
  • When is learning needed? useful?
  • How does a selfish agent determine when learning could be a strategic advantage?
• Specifically, we are studying the behavior of reinforcement learning agents.
• Looking for help.
• José M. Vidal and Edmund H. Durfee. Predicting the expected behavior of agents that learn about agents: the CLRI framework. Autonomous Agents and Multi-Agent Systems, December 2002.

• by Sujay E. Jacob
• Each blue dot is a radar and each yellow dot is a target.
• The radars have to be allocated to targets in order to maximize a utility function.
• Radars have only local information.
• Also, minimize communications.

• People decide what to read/watch/eat based on others' recommendations. The is no absolute "truth", only relative opinions.
• Q: How can an agent tell which other one are like him?
• We assume selfish (utility-maximizing) agents. The parameters of the problem are:
  • Fixed cost of consumption.
  • Reward for consumption: positive or negative.
  • Fixed cost of communications.
  • Processing time and memory are free (unlike humans).
  • Agents do have inherent similarities among themselves.
• Developed protocol to enable agents to operate rationally within this environment while also maximizing global utility.

1. Attend the MAS reading group [16] meetings. Read the papers.
2. Browse my website http://www.multiagent.com [17] to see what others are doing.
3. Master thesis ideas are available at http://jmvidal.cse.sc.edu/thesisprojects.html
4. Take my CSCE 782: Introduction to Multiagent Systems class.