Multiagent Systems (MAS) is a recent outgrowth of the field of Distributed Artificial Intelligence (DAI). DAI, in turn, is the intersection of the fields of Distributed Computing and AI as depicted in Figure 1.
Figure 1: A taxonomy of fields containing MAS. DAI is the intersection
of Distributed Computing and AI. DAI consists of MAS and ``not MAS'',
most of the latter of which are DPS systems (but not all, as indicated
by the dotted curve).
The field of Distributed Computing has existed ever since it became possible to use more than one processor to work on a computing problem. Originally, these processors shared the data associated with a given problem. The majority of issues that arose pertained to the parallelization and synchronization of the different processors.
Soon after people started thinking about distributed computing, AI researchers became involved, trying to add automated intelligence to the field. AI efforts examined the prospect of sharing control as well as data among processors. To use constraint satisfaction as an example, different processors were allowed to alter the working solution rather than blindly solving their subproblems and reporting back to a central controlling system. The resulting research became known as DAI. It was distinct from traditional Distributed Computing in that it focussed on problem solving, communication, and coordination as opposed to the lower level parallelization and synchronization issues [6].
DAI systems do not necessarily involve agents. One possible definition of DAI uses the less general term, ``entities:''
Distributed Artificial Intelligence (DAI) is concerned with the study and design of systems consisting of several interacting entities which are logically and often spatially distributed and in some sense can be called autonomous and intelligent. Gerhard Weiß [3]Because of the emphasis on information management in the past, DAI has been decomposed into Parallel AI and Distributed Expert Systems, the latter of which was broken into Distributed Knowledge Sources and Distributed Problem Solving (DPS) [7]. However, these days people tend to break DAI systems into one of two classes: DPS and MAS [3] (see Figure 1).
DPS deals with information management while MAS deals with behavior management. Their intersection arises in systems that use agents to do information management. Thus the boundary between DPS and MAS is not a clear one. The information management of DPS consists of both task decomposition (task sharing), in which a complex task is divided into manageable subtasks and sent to different processors, and solution synthesis (result sharing), in which the results of the different subtasks are combined.
In DPS systems, generally designed by a single person or a small group of people, there are many strong assumptions made about the compatibility of different problem-solving entities. In MAS there are no such guarantees. Different agents could have entirely different designs and yet need to interact.
A final important characteristic of MAS is that the individual behaviors of agents are typically far less complex than the interactions among agents. It is these behavior interaction issues that form the basis for the field of MAS. Table 1 summarizes the definitions of MAS and its related fields.
Table 1: Definitions of MAS and its related fields.