Agent Systems and Applications

As the number of deployed multi-agent applications increases, further and better experience with the technology is gained, enabling a strong evaluation of the ®eld from a more practical perspective. In particular, questions relating to how the theory of multi-agent systems impacts on practice, and how the practical development itself compares with other technologies, can be answered in the light of a heightened level of maturity. Given the tensions between theoreticians and practitioners in computing in general, let alone their spats in AI or multi-agent systems in particular, the discussion on agent systems and applications was both vigorous and enthusiastic. 1 Characteristics of multi-agent applications What are the characteristics of applications which make them suitable for a multi-agent system solution? Are there applications which are unsuitable for multi-agent systems? The obvious response to the question of characteristics of suitable applications is that they are worth considering when the application requires both distribution and intelligence. Thus, a multi-agent approach would be sensible for problems that are inherently (physically or geographically) distributed where independent processes can be clearly distinguished. Such problems include, for example, distributed sensor networks, decision support systems, air trac control, or other networked or distributed control systems. A distributed approach is not in itself enough, however, and there should also be requirements for intelligence or adaptivity in the sub-processes that involve explicit reasoning about behaviour, for example. If the problem can be solved by means of a look-up table at each node of a network, a multi-agent system would be excessive. Physical distribution may not be the only reason for a distributed approach. Minsky's Society of Mind paradigm (Minsky, 1986) suggests the use of a multi-agent system where there is a wide range of reasonably selfcontained pieces of functionality that require the use of AI, especially if they run asynchronously, or are distributed or independent in the sense of timing. In this way, a multi-agent approach might be applied to a single robot manipulator taking each joint as an agent, or to a single static system with sensors and e€ectors (an immobot) such as a smart building or a spacecraft. A range of further application areas that qualify for multi-agent solutions can also be enumerated. These include those requiring the interconnection and inter-operation of multiple autonomous, self-interested existing legacy systems, expert systems, and decision systems, or those requiring solutions that draw from distributed autonomous and sel®sh information sources, such as the Personal Travel Assistance demo from BT; those where the solutions draw from di€erent This report is the result of a panel discussion at the Second UK Workshop on Foundations of Multi-Agent Systems (FoMAS'97). All members of the panel are authors, listed alphabetically. distributed experts, such as health care provisioning, in which some central agent cannot possibly perform the task without help from other experts; problems that naturally cross organizational boundaries for which an understanding of the interactions among societies and organizations is needed; and problems where no single agent has a total view, but several agents have local views. The notion of ownership of information and strategies in the application is important here, and in particular when it is distributed over di€erent organizational entities so that no single entity can (or does) have access to all the information. Final particular examples include traders in a marketplace (Chavez et al.,1997), and di€erent entities in a business working on other's tasks (Jennings et al., 1996). In these situations, the problems to be tackled do not have one overall goal, but rather consist of balancing the (possibly con ̄icting) goals of di€erent entities. As with any other piece of technology, there are plenty of applications that do not require a multiagent approach. Multi-agent systems are not required merely to produce modularity (though they reduce complexity), extra speed (though this may be an e€ect of their inherent parallelism), reliability (though they provide redundancy), ̄exibility or re-usability at Newell's knowledge level (Newell,1982). In the same way, they are not required simply because a problem is too large for a centralized single agent due to resource limitations, nor because of the sheer risk of a centralized system, nor merely for reasons of eciency, hetergenuous reasoning, etc. For example, a payroll system might bene®t in a software engineering sense from an objectoriented approach (providing modularisation and re-use), but such standard data-processing problems do not really need the communications overhead or functionality of a multi-agent approach. Such applications usually require neither distribution nor intelligence. In the same way, a small free-standing expert system used in a single location requires the intelligence but not the distribution, and the human interface is insuciently complex to be worth thinking of it as an agent. Finally, it is worth noting that a multi-agent system approach may be useful, though not necessary, when tackling problems that are easiest visualized in a way that appears to have the above characteristics, such as combat simulation (Rao and Selvestrel, 1992). 2 Applications development What does a multi-agent approach to applications development buy you over more standard approaches such as object-oriented, expert systems, or distributed computing approaches. In general, Object-Oriented (OO) systems, expert systems and distributed computing techniques do not o€er solutions to the kind of problems for which multi-agent systems are used, for a range of reasons (Wooldridge, 1997). OO techniques are good in general, but are rather low-level for intelligent applications. They can be used, for instance, to implement knowledge representations, but they do not themselves provide a knowledge representation. OO development methodologies can, however, be seen as a low-level underpinning for a multi-agent methodology. The same might be said of distributed computing methodologies and indeed, many multi-agent systems (e.g. ADEPT) are built on top of distributed platforms such as CORBA. However, if it can be argued that if OO approaches are still relatively new, these are even newer and less generally accepted. Again, however, the level is wrongÐfor example, communications protocols do not operate at the high level of Speech Acts as one might wish for a multi-agent system. Yet distributed computing approaches could deal with some of the lower levels as well as providing some of the basic techniques (such as protocol de®nition and validation). An OSI model with an elaborated applications level might be ideal or, alternatively, current approaches to multi-agent applications development might form such a layer. Expert systems are even more problematic in terms of development methodology. There is a reasonable consensus on the life-cycle, but while KADS has made some impression, many avoid it. Developments such as KRL and other movements for standardization are very much in their infancy, while ontologies are helpful but not widely used. The most successful parts of knowledge r . ay l e t t e t a l . 304