EVALUATION OF AGENT ARCHITECTURES

A Special Track at the Sixth International Workshop on
AGENT THEORIES, ARCHITECTURES, AND LANGUAGES (ATAL)

Orlando, Florida, USA July 15-17, 1999

Control architectures for autonomous agents (or, in short: agent architectures) are conceptual or implemented computational models or frameworks that support autonomous hardware or software systems, that need to be designed to:

• plan towards and achieve their goals;
• react to unforeseen events;
• deal with uncertainty and change;
• interacting meaningfully with other agents.

Based on research in Artificial Intelligence (most notably driven by robotics), and largely initiated by Bratman, Pollack, and Israel's seminal paper on the IRMA architecture in 1987, agent architectures have started playing an important role in Artificial Intelligence and Intelligent Agents research, well before the term "Agent" reached its current level of public attention.

To our knowledge, the ATAL workshops have been the first scientific forum that officially recognized the importance of agent architectures, introducing the topic as a part of the name of the ATAL workshop, and as a first class citizen in agent research.

In recent years, a multitude of agent architectures have been proposed in the literature, inside and outside the ATAL series. However, only very little work has been done in evaluating these architectures. As agent technology matures and increasingly influences developments in mainstream computer science, the importance of evaluating and comparing agent architectures with respect to their strengths and weaknesses, and of giving software engineers and system designers guidelines to select suitable architectures for given applications, becomes apparent.

Therefore, the ATAL special track on the EVALUATION OF AGENT ARCHITECTURES seeks insights on the suitability, strengths, and weaknesses of various architectures in different task domains.

SPECIAL TRACK TOPICS

We welcome the submission of all papers on aspects of the evaluation and comparison of agent architectures, but we are particularly interested in "viewpoint" papers, which address any the following questions:

1. Are certain architectures particularly suitable / unsuitable for certain application domains?  When should a hybrid reflective-reactive architecture be deployed? When are reactive or self-organizing architectures suitable?  When are they unsuitable?
2. Which types of architectures work best in uncertain, dynamic or real-time environments?
3. How can the effectiveness of architectures be evaluated? How can two architectures be compared?
4. What are benchmark tasks for various types of environmental niches?  What benchmark tasks require cognitive capabilities?
5. Do you see a common essence underlying most of the agent architectures described so far in the literature? Is it possible / reasonable to identify as much as a minimal set of components that make up an agent? Could evaluation criteria be derived from taking this approach?
6. Is there any need for standardization at the level of control architectures? If so, what are the preconditions that need to hold to make it feasible? Which aspects of  control architectures would benefit from and should be  included into standardization efforts, which ones should not?
7. Are there any empirical or theoretical results to help designers of agent systems to decide which architecture to choose for a certain application domain? Are there any lessons to be learned from the practical design of agents and agent systems?
8. What role play agent architectures and their evaluation in a software engineering context, and what role could they play in the future? How would you explain this role to software designers who are fairly successfully building today's large scale without (consciously) making use of agent technology, and without even thinking about agent architectures?

The focus of contributions to the special track should be on the evaluation of agent architectures rather than on descriptions of individual architectures. Empirical and theoretical work that discusses experiences with existing architectures or that compare different agent architectures are very welcome.

ABOUT ATAL-99

ATAL-99 is the sixth in the international workshop series "Agent Theories, Architectures, and Languages", which to date has alternated annually between North America and Europe. Highly edited post-proceedings of all ATAL workshops have been published as volumes one to five of the "Intelligent Agents" series from Springer-Verlag.

For convenience, ATAL-99 will be held over the three days immediately preceding the AAAI-99 conference, which is also being held in Orlando, Florida. Note, however, that there is no formal connection between AAAI-99 and ATAL-99: they are separate events. The ATAL-99 proceedings will be formally published as volume six of the Intelligent Agents series.

ATAL-99 SUBMISSION DETAILS

Those wishing to participate in the workshop should submit an original research paper of up to 5000 words (approximately 13 pages maximum) to the general chair (Nick Jennings). If you wish your paper to be considered for a special track then mark this clearly on the front page. Electronic submission in PostScript is strongly encouraged, but four single-sided hard copies will also suffice. The first page should include the full name and contact details (including email, full postal address, and telephone number) of at least one author. Formatting instructions are available from the workshop WWW site (see above). The pre-proceedings will be distributed at the workshop; the formal proceedings will be published shortly afterwards.

Those wishing to attend without presenting a paper should send a brief summary of their interests in agents to the organizing committee chair Nick Jennings. Attendance will, of necessity, be limited.

TIMETABLE

Submissions due April 5, 1999 Notifications sent  May 7, 1999
Prefinal versions due  June 11, 1999 Workshop July 15-17, 1999