Invited Speakers -- ICMAS '95


First International Conference on Multiagent Systems - ICMAS '95

June 12-15, 1995
San Francisco, California


Blissful Ignorance: Knowing Just Enough to Coordinate Well

Edmund H. Durfee
EECS Department
University of Michigan
1101 Beal Ave.
Ann Arbor, MI 48109
durfee@caen.engin.umich.edu

Abstract:

Coordination requires making decisions based on expectations of what others will do, which usually implies having models of others. Better (more complete, timely, and precise) models of others can improve the accuracy of expectations, but acquiring and using such models carries costs in perception, communication, computation, and/or memory. Moreover, by committing to actions based on its models of others, and by being committed to actions implicit in any model of itself that it has projected to others, an agent might leave itself without flexibility to respond to unexpected opportunities. A fundamental component of the coordination process is thus to develop models that convey enough knowledge to benefit agents, while at the same time retaining enough ignorance among agents to avoid excessive coordination costs and to respond flexibly to new circumstances. In this talk, I describe various kinds of knowledge that could be captured in an agent model and how the knowledge can impact the quality and costs of coordination. I consider how different restrictions on and simplifications of agent models have been introduced that have led to practical coordination mechanisms for applications including robotic coordination, decentralized resource scheduling, and distributed information gathering. While, in most of these mechanisms, the decision about how agents should model each other is designed in, I also describe early steps in the development of automated techniques that agents can use to find agent models that are most appropriate for their coordination needs.


Computational Organization Research

Les Gasser
IMPACT Lab
School of Engineering
USC
Los Angeles, CA 90089
gasser@usc.edu

Abstract:

Large, dynamic ensembles of machines, intelligent agents, and people interacting clearly exhibit organizational-level phenomena. These "computational organizations"---organizations made up completely or partly of computational participants---raise many exciting R&D opportunities: computational models and representations of organizational knowledge, explicit organizational ontologies, simulating organizational activity or structuring, computational approaches to building organization theories, coordination algorithms, computational approaches to organization design, computational tools for organization analysis, and so on. From a complementary point of view, a growing worldwide community is researching phenomena of human organizations using computational methods. Pressing research issues in human organization theory and organization development are highly amenable to computational modeling, theorizing, and experimentation.

These complementary research directions converge in what might be termed "Computational Organization Research". Computational modeling, analysis and simulation can help us to better understand generic underlying principles of organization in both human and computational settings, and to solve particular, pressing configuration problems---issues that are both highly impactful and scientifically challenging. The impacts on improved effectiveness and flexibility in organizations can be great, and the theory, modeling technology, and infrastructure are ready. This talk will motivate and survey current computational organization research, and pose some core questions for further development.


Parallel, Distributed and Multi-Agent Production Systems: Research Foundations for Distributed Artificial Intelligence

Toru Ishida
Department of Information Science
Kyoto University
ishida@lab7.kuis.kyoto-u.ac.jp

Abstract:

Production systems have been widely used as expert system building tools and recognize-and-act models in cognitive science. This talk is intended to introduce parallel/distributed/multi-agent production systems, and to reveal their possibilities as research foundations for distributed artificial intelligence: a parallel production system as an agent reactive architecture, a distributed production system as an adaptive agent organization, and a multi-agent production system for organizational learning. Production systems already have been equipped with clear syntax and efficient pattern matching algorithms. Their foundations can be further strengthened with recent advanced theories such as linear logic, real-time search and reinforcement learning.


Developing Industrial Multi-Agent Systems

N. R. Jennings
Distributed AI Unit
Dept of Electronic Engineering
Queen Mary & Westfield College
University of London
N.R.Jennings@qmw.ac.uk

Abstract:

The development and deployment of multi-agent systems in real world settings raises a number of important research issues and problems which must be overcome if Distributed AI is to become a widespread solution technology. Work undertaken in the context of the ARCHON project has provided a number of important insights into these issues - especially with respect to the domain of industrial control applications. This talk will draw together many of these experiences and will give details of an Electricity Transportation Management System which is one of the world's first operational Distributed AI systems.


MULTIAGENT PLANNING AS A SOCIAL PROCESS: VOTING, PRIVACY, AND MANIPULATION

Jeffrey S. Rosenschein
Institute of Computer Science
Hebrew University
Jerusalem, Israel
jeff@cs.huji.ac.il

Abstract:

The essence of effective multiagent planning is to coordinate and merge the separate search activity of individual agents. When these agents are self-motivated, however, the coordination and merging process becomes more subtle---different agents are actually interested in transforming the world in different ways, and have an interest in influencing the group activity in certain directions. Under these circumstances, multiagent planning has to incorporate elements of consensus formation. Designed appropriately, the consensus process may be able to protect the group against manipulative individuals, and, as much as possible, maintain the privacy of agents' preferences. This talk explores multiagent planning as a heuristic process of distributed search and merging, and in particular considers techniques (such as voting) that enable a group of self-motivated planning agents to reach consensus effectively.

This is joint work with Eithan Ephrati, Computer Science Department, the University of Pittsburgh.


What We Talk About When We Talk About Agents

Yoav Shoham
Computer Science Dept.
Stanford University
Stanford, CA 94305
shoham@flamingo.stanford.edu

Abstract:

The terms "software agent" and other, related ones, are used very aggressively these days, in academia as well as industry. Some of the work in the area is exciting, but the area suffers from a lack of cohesion and variable quality. In the talk I will do the following. First, I will drive home the diversity point by briefly discussing a half dozen examples of agent-based work. The conclusion will be that there is no definition of software agents that is at once reasonably clear and covers most of the work in the field. Second, in lieu of such a definition, I will outline one approach to understanding the area, based on dimensions of agency. Finally, I will propose that this approach is useful but inherently flawed, and propose an alternative which I find more useful.


Coordination of Multiple Intelligent Agents in the Infosphere

Katia Sycara
School of Computer Science
Carnegie Mellon University
Pittsburgh, PA. 15213
e-mail: katia@cs.cmu.edu

Abstract:

The current explosion of data and information on the Internet-based Infosphere has made the problem of locating information sources, accessing them and combining information from them a very critical task. The notion of Intelligent Software Agents has been proposed to address this challenge, Intelligent Software Agents are programs that must act on behalf of their users in order to perform laborious information gathering tasks, such as locating and accessing information from various on-line information sources, resolve inconsistencies in the retrieved information, filter away irrelevant or unwanted information, integrate information from heterogeneous information sources and adapt over time to their users' information needs and the shape of the Infosphere. These tasks are performed automatically or with little help from the users. Most current research on Intelligent Agents has focused on the development of a single information gathering agent. It is clear, however, that the Infosphere is the natural playground of Multiple Intelligent Coordinating Agents. This arises from the vastness of the available information, the heterogeneity of the information resources, the diversity of information gathering and problem solving tasks that the gathered information supports, and the presence of multiple users with related information needs.

This talk will address issues involved in designing multiple Intelligent Agents that coordinate and learn from each other in order to support information gathering and integration. In addition, the talk will discuss opportunities in the Infosphere for multi-agent research and development. We will draw examples from our work on the PLEIADES project which is developing distributed collections of Intelligent Agents that cooperate, negotiate and learn from each other in performing goal-directed information retrieval and integration to support a variety of problem solving tasks.


Back to ICMAS page.


If you have questions, problems or suggestions send mail to www@icmas.cs.umass.edu
Last Update: 3 April 95