Choice, autonomy, multiple goals or tasks, situated, learning, flexible, communicating, cooperating, resource-bounded - the intelligent agent software paradigm is becoming more widespread and better understood. The continued growth and development of the agent paradigm is due in large part to researchers interacting, pushing each other to consider new ideas and different views. The ATAL series of workshops (Agent Theories Architectures and Languages) has historically served as one such forum and has contributed significantly to the development of agent computing. This trend continues; ATAL-99 follows prior ATALs in bringing together researchers with a broad range of core agent interests and emphasizing a synergy between theory, methodology, formal methods and software infrastructure. This year's ATAL also focused on the issue of agent oriented software engineering and evaluation of agent architectures. The rationale behind the focus is that to facilitate widespread use of the agent paradigm, we must provide a bridge from intellectual exploration to industrial development and application.
The research presented at the workshop was fairly diverse. Seventy-five papers were submitted to ATAL from nineteen different countries. Twenty-two were accepted for presentation (acceptance rate of 29%). Sixty people attended the workshop, which was held July 15-17, 1999, in Orlando, Florida, USA. Aside from the intellectual activities, one of the more noteworthy features of the workshop was the two-story plastic bear (and large friends) located across the street (outside of the FAO Schwartz toy store).
As my perceptions necessarily bias my perception of the workshop and my reporting thereof, I must digress and note that my research goal is to create reusable agent control problem solving components (e.g., scheduling, modeling, coordination) that can be bundled with domain problem solvers to create intelligent social agents (and reduce the cost of agent development). I am primarily interested in technologies that address the uncertainty and complexity of real environments while being tractable, approximate, or otherwise addressing resource boundedness so that they can be used for actual on-line agent control. Though much of the work presented at ATAL does not address the tractability side of my own research agenda, I found several areas of work interesting and useful.
Best paper awards went to ``Operational Semantics of a Role-based Agent Architecture'' by Jacques Ferber and Olivier Gutknect and to ``Extending ConGolog to Allow Partial Ordering'' by Chitta Baral and Tran Cao Son. The former paper formalizes the semantics of AALAADIN models of multi-agent systems enabling verification of system properties, e.g., agents acting in multiple groups simultaneously. The work is potentially widely useful as it is based on an organization/role-view model of multi-agent systems and removed from the details of particular implementations. The latter paper extends the Golog/ConGolog high-level execution language to include htn-like non-determinism. I found the extension notable as, in my view, online choice is central to agents adapting to different problem solving contexts and is one of the key aspects of our research in TÆMS, Design-to-Criteria scheduling, and GPGP coordination. The best student paper award went to ``Incorporating Uncertainty in Agent Commitments'' by Ping Xuan and Victor Lesser. The paper describes the addition of uncertainty to commitments in GPGP, its impact on the agent reasoning process, and its use in negotiation. Empirical results suggest that in certain situations, incorporating uncertainty into the agent's coordination process leads to better coordination and thus higher overall system utility.
The workshop included two invited talks. John Pollock's talk, Rational Cognition in OSCAR, discussed the use of defeasible epistemic reasoning in OSCAR. I found the talk particularly interesting as Pollock's work is motivated by a belief that in complex agents, operating in complex (uncertain) environments, the search for threats in plans via traditional methods is computationally infeasible. Tractability is achieved in OSCAR using defeasible reasoning in conjunction with evaluation and revision. Sarit Kraus' talk, Agents for Information Broadcasting, examined the problem of broadcasting customized content. The problem is framed as a single-source multiple-consumer problem and the objective is to characterize consumers' needs to improve the targeting of content. Kraus presented an innovative hierarchical multi-agent solution in which agents cluster consumers into communities with similar interests, construct a representative profile for the communities, and customize content accordingly.
The workshop also included two panel discussions: agent-oriented software engineering and evaluation of agent architecture. Rather than attempting to summarize a wide range of commentary, I'll instead point out a few of the more notable intellectual themes. One such theme is the view of agents as components from which a larger multi-agent system can be constructed. Related to this is the issue of organizing agents and organizing distributed computation. Possibly both culminate into organizational-level programming rather than programming at the individual agent level - several papers at the workshop dealt with social views or local-control based on a local-view of the larger organizational context. Architecturally, adaptability, responsiveness, and flexibility were identified as key concepts for dealing with dynamic domains - possibly achieved through soft-real-time control problem solving, choice, and learning. Robustness and fault-tolerance were also discussed as possible desirable characteristics of agent architectures and dimensions along which they may be evaluated. With respect to FIPA (Foundation for Intelligent Physical Agents), it was observed that while important, it has little impact on the internal structure of each local agent and thus offers little help when designing the local agent. However, several panelists also felt that the field is currently too diverse to attempt to impose standards on the structure of individual agents.
This is the first year that I attended ATAL and overall I felt that it was a good use of my time and beneficial to the development of my research. I particularly appreciated the blend of basic agent research and work in agent oriented software engineering. The papers will appear in: N.R. Jennings and Y. Lesperance (eds) (2000), Intelligent Agents VI, Springer Verlag. Information on ATAL-2000 will be available via http://www.atal.org; Yves Lesperance may also be contacted directly at lesperan@cs.yorku.ca. Contact information for authors and invited speakers mentioned in this article follows: Jacques Ferber and Olivier Gutknect {ferber,gutkneco}@lirmm.fr, Chitta Baral and Tran Cao Son {chitta,tson}@cs.utep.edu, Ping Xuan and Victor Lesser {pxuan,lesser}@cs.umass.edu, John Pollock pollock@u.arizona.edu, and Sarit Kraus sarit@macs.biu.ac.il.