Keith S. Decker

Associate Professor
Dept. of Computer and Information Sciences
University of Delaware
77 E. Delaware Ave. (the AI/NLP GreenHouse)
Newark, DE 19716-2586
(302) 831-1959 (office)
(302) 831-4091 (fax)


Other Pages

EASSS-03 Complete Lecture Slides

EASSS-03 Changed Lecture Slides only

My UD Courses  Pictures

UMass Distributed AI Page CMU Intelligent Software Agents Page

My Thesis Page  Misc. Links 


My research program is concerned with the areas of Distributed AI and Multi-Agent Systems. There are many problems that do not easily submit to centralized, monolithic computational solutions. One reason that these problems arise is because computer applications support people who are doing tasks in the context of an existing organization. Just as a human organization was designed or evolved so that individuals have different roles, responsibilities, power, control over others, decision-making authority, etc., computer systems that work within existing organizations need to respect these relationships. Consider a nurse attempting to schedule a patient's tests, including an x-ray, and the x-ray unit staff's need to provide services for many patients while using their equipment efficiently. No centralized solution is possible because 1) the humans involved will not give up their authority to a centralized administrator/computer program, 2) the groups involved are actually attempting to maximize different sets of performance/utility criteria. Other examples include collaborative engineering design and agile manufacturing systems.


Much of my research has focussed on the analysis and design of coordination mechanisms for groups of collaborative software agents. Coordination can be defined as the act of managing interdependencies between activities. Many researchers have shown that there is no single best organization or coordination mechanism for all task environments. The design of coordination mechanisms for intelligent agents cannot rely on the principled construction of agents alone, but must also rely on the structure and other characteristics of the agents' task environment. One focus of this activity has been the development of a framework called TAEMS (Task Analysis, Environment Modeling, and Simulation) for representing and reasoning about the salient features of a computational task environment. Another focus has been on how software coordination mechanisms can be designed to respond to particular features of the task environment structure; one result has been the Generalized Partial Global Planning family of coordination algorithms.

A second focus is on Integrating Organizational Style with Environmental Characteristics. Organizational styles can be operationalized as sets of coordination mechanisms that enforce certain behaviors as part of the style. However, not all coordination behavior arises from abstract organizational styles; much arises from standard responses to specific environmental (problem domain) characteristics.  This research aims to extend environment and organization models, represent standard responses to common environmental problems, and develop operational specifications for organizational styles. This allows the interactions between environmental constraints and organizational styles to be reasoned about analytically. Designers are not doomed to create coordinated systems with only a few simple organizational behaviors and large numbers of brittle domain-specific coordination heuristics. Instead, given good environment models and certain other constraints (often dictated by existing human organizations), desiginers can make principled choices of computational organizations and standard environment-influenced coordination behaviors. This project will result in a much richer set of tools for building and analyzing both multi-agent computer systems and integrating such systems with existing human organizations in applications such as distributed information gathering, distributed scheduling, and concurrent engineering.

Distributed Information Gathering

It has been clear for some time that the Internet is a viable medium for supplying the data needed for making various types of decisions. As more useful data become available at different times and in multiple locations, however, it becomes more difficult and time-consuming for a person to collect and evaluate that data. Most current agent-oriented approaches to this problem have focussed on single agents with general knowledge and capabilities to perform a wide range of user-delegated information-finding tasks. Such centralized approaches have several limitations: the need for an enormous amount of knowledge in order to provide coverage for a variety of tasks; the implied centralized processing bottleneck; the inability of most such single agents to deal dynamically with the appearance of new agents and information sources. One solution is to use multi-agent computer systems to access, filter, evaluate, and integrate this information. We have been developing multi-agent systems that can compartmentalize specialized task knowledge, organize themselves to avoid processing bottlenecks, and cope with dynamic changes in the agent and information-source landscape. We have developed the DECAF agent architecture and toolkit for quickly prototyping multi-agent information gathering systems.


Our most comprehensive set of information gathering applications centers around the area of bioinformatics. Today biological information and algorithms for the analysis of biological data are available on the Internet in many different locations with overlapping content, different structure, and varied amounts of curation. Our approach to these problems, called multi-agent information gathering, is to apply multi-agent systems technologies to create software agents for information retrieval, filtering, integration, analysis and display. Currently we have developed a prototype system for the automated annotation of herpesvirus sequences with homologs, motifs, domains, and sub-cellular location predictions. The system automatically produces a searchable database of this information HERE (coming soon, an automated version produced for the coding segments of the chicken genome). Other projects in progress include a new subsystem for assisting biologists in functional annotation using the GO ontologies (try out our visual display system GOFigure!); automated UNIGENE cluster analysis; automated EST processing and annotated consensus sequence database/web site creation; and soon the integration of various forms of gene expression data.

Limited Rationality

Herbert Simon's theory of bounded rationality can be construed as a theory of why humans form organizations---that no one person can process all the necessary information and make all the correct decisions for everyone. Thus control is shared, information filtered, decision-making authority is passed to subordinates, we satisfice rather than optimize. Our work considers problems where time and computational resources are limited for computational agents also. This work on soft real-time approaches has resulted in algorithms for making appropriate commitments to other agents in time-constrained situations, and for using "design-to-time" and "anytime" algorithms to schedule an agent's problem solving activities.

Last Update: 5/24/02