MAS Infrastructure: Definitions, Needs and Prospects

This paper attempts to articulate the general role of infrastructure for multi-agent systems (MAS), and why infrastructure is a particularly critical issue if we are to increase the visibility and impact of multi-agent systems as a universal technology and solution. Second, it presents my current thinking on the socio-technical content of the needed infrastructure in four different corners of the multi-agent systems world: science, education, application, and use. 1 Why MAS Infrastructure is an Issue MAS have the potential to meet two critical near-term needs accompanying the widespread adoption of high-speed, mission-critical content-rich, distributed information systems. First, they can become a fundamental enabling technology, especially in situations where mutual interdependencies, dynamic environments, uncertainty, and sophisticated control play a role. Second, they can provide robust representational theories and very direct modeling technologies to help us understand large, multiparticipant, multi-perspective aggregates such as social systems, ecologies, and large information processing systems. Many people inside and outside the MAS community can now legitimately envision a future in which we clearly understand how information and activity of all kinds can be managed by (automated) teams and groups (not individuals), and in which we naturally and ubiquitously manage it that way: a vision of "MAS everywhere." Progress toward systematic scientific principles and robust coordination/interaction technologies for MAS has been underway for the past thirty years. Though more fully articulated knowledge is needed, we are on the way toward developing the knowledge that will eventually give MAS a comprehensible. predictable operational character. MAS researchers have developed some fairly sophisticated theories and technologies for multi-agent interaction; coordination; coalition formation; conflict resolution; dynamic organization and reorganization; network and organization design; faulttolerance, survivability, and robustness; multi-agent learning; and real-time multiagent behavior. In a theoretical sense, there is much interesting work and many good results, many of which point the way to more intriguing questions. It is fair to say that in several of these areas---coordination, teamwork, coalition-formation, dynamic reorganization, for example---the approaches developed in the multidisciplinary MAS community are among the most detailed, sophisticated and general that are available. But from a practical point of view, our understanding is really just beginning. A number of deep scientific issues (such as managing dynamic heterogeneity [11] and understanding system-wide pathologies [10]) are very under-explored and have implications that will only arise clearly when we begin to have widespread, interacting, fielded MAS. Currently there is a very small number of fielded MAS systems, and in general there are very few---if any---systems in routine operational use that actively exploit the most sophisticated MAS theoretical developments such as robust interagent coordination techniques, multi-issue negotiations, dynamic organizational efficiencies, or multi-agent learning. Moreover, even if the next generation of MAS technical milestones are met and new capabilities are created by researchers, widespread use of MAS won't occur until a critical mass of fielded systems, services, and components exists and network effects take hold to blossom the user population and public interest. The public incentives for widespread attention to and use of analogous technologies such as Web browsers and cell phones appeared only with the development of a) a stable, reliable, accessible infrastructures, and b) a critical mass of "content" (e.g., broadly interesting websites) that compelled potential users. Similarly, until we have a stable, operational, widely accessible, and low-apparent-cost MAS infrastructure populated with a critical mass of MAS services and systems that provide value to prospective users, MAS is likely to languish as a technology with potential, not kinetic, energy. Another critical impact of the failure to have a variety of fielded MAS is that we lack practice and experience in building and operating MAS in situ. In virtually every case of implemented experimental or commercial MAS, the theoretical and technological frameworks used rely on standard, homogeneous agents and limited, inflexible standards of interactivity. Each project or application is generally self-contained and its components can accommodate only a very limited, predictable range of interaction. In the research community, each group's projects are quite often similarly isolated. Though there are some widely distributed MAS tools (see e.g., [1]) it is rare that one group's tools and technologies work with those of others in an integrated way, and cross-group testing generally doesn't happen. (However, see [15],[18].) An important exception to this is recent experiments in constructing and using joint simulation environments and shared physical environment, such as the RoboCup simulations. Still, under these conditions there is generally still careful centralized control over interaction possibilities, determined for example by simulator APIs or controlled physical environments. Also worth mentioning are the newly emerging infrastructure tools, such as the Nortel FIPA-OS implementation, many attempts at KQML tools, and many XML frameworks e.g., for e-commerce [20]. It's not yet clear to what extent these will actually serve to integrate agent behaviors (see e.g., [11],[19].) The bottom line is that despite the compelling vision of ubiquitous multiagent technology, we simply have hardly any real experience building truly heterogeneous, realistically coordinated multi-agent systems that work together, and thus almost no basis for systematic reflection and analysis of that experience. Finally, the current prospects for advanced pedagogy in MAS are very weak, especially in terms of demonstration of MAS and experimentation in MAS behavior and implementation. How will the MAS communities create pedagogical environments and tools that will help develop, transfer, and extend the MAS knowledge and skills to impact widening groups of people? Simply put, there are few if any sharable tools with serious pedagogical aims. 2 MAS Infrastructure Elements and Attributes An infrastructure is a technical and social substrate that stabilizes and rapidly enables instrumental (domain-centric, intentional) activity in a given domain. Said another way, (technical) infrastructure solves typical, costly, commonly-accepted community (technical) problems in systematic and appropriate ways. In this way, infrastructure allows much greater community attention to unique, domain-specific activities. As Star and Ruhleder have pointed out [16] infrastructures have the general character of being: embedded "inside" other structures; transparent (not needing reinvention or re-assembly each time); of wide reach or scope; learned as part of community membership; linked to conventions and norms of community practice; embodying standards, shaped by pre-existing installed bases of practice and technology; and invisible in use yet highly visible upon breakdown. Infrastructure is also an effective leveling device: it unifies local practices with global ones, both providing coordination and creating shared knowledge. For the purposes of this paper, I've divided spheres of MAS activity into four categories, each of which has different infrastructure needs---that is to say, the communities in each sphere have different views of their own "typical, costly, commonly-accepted community technical problems" and different notions of what are the most "systematic and appropriate" solutions to them. These four categories are MAS science, MAS education, MAS application, and MAS use. The most critical infrastructure needs are not the same across these focus areas, and not all of these area are developing with equal force or speed. (E.g., MAS science is way ahead of MAS use, and MAS application is somewhat ahead of MAS education). Table 1 presents a schematic view of MAS infrastructure elements and characteristics, and their relationship to each of the four MAS spheres. I'll treat each of these in more detail below. 3 Needed Elements of MAS Infrastructure Rows of Table 1 show the main elements of MAS infrastructure, categorized into System Elements, Services, Capabilities, and Attributes. Table 1 shows the relationships of these infrastructure elements to needs in the four different categories of MAS activity. MAS Infrastructure Needs