Coordination consists of a set of mechanisms necessary for the effective operation of Intelligent Agent Societies (IASs). In building such societies, it is important to design and implement coordination in accordance with the known requirements and anticipated working conditions of the IAS in question. Currently, there is little theoretical support that could help in this process. We outline the approach and design principles of our work on automatically generating an empirically-based theory of coordination. We also describe the first set of results obtained to prove the feasibility of the approach. We are concerned with Distributed Problem Solving (DPS) systems in which all agents share an identical goal structure and fully collaborate, as opposed to Multi-Agent Systems (MAS) in which agents may also compete with one another. Our investigation is based on an easily modifiable and parametrizable generic IAS, the P-system, a metaphorical and abstract version of our earlier work, the Distributed Control of Nationwide Manufacturing Operations. The P-system shares characteristic properties with most, if not all, IASs. It was used for a sequence of rigorously designed experiments in which control variables operated under well-defined conditions and performance measures were observed. We infer, from the statistical analysis of these data, characteristic and important descriptors of the organization and functioning of IASs in general. The resulting relations should produce insight into the fundamental issues of coordination, provide design tools and guidelines for constructing new IASs, and lend support in trouble-shooting existing ones.
When facing a system consisting of a large number of agents, usual techniques are ineffective and prohibitively expensive. We have defined different forms of social structures each of which is constituted by a group of agents. Such software artifacts can be used for coordination, and controlling and managing unit groups of agents. The artifacts allow conceptual, computational and representational simplicity in the design and implementation of IASs. The basic features of interest, in the context of IASs, are: (1) an individual agent interacts only with only a small and manageable part of the entire universe they can perceive - with fellow agents in its own group of agents and with other group entities; (2) the interactions are not controlled by special internal scripts or protocols embedded within the agents but via external artifacts.
At present our focus is on electronic markets - in particular, consumer commodity markets, such as the electricity market. We have constructed a pattern template on which such markets are based (negotiation-based, auction-based, etc.). The social structure is the exchange, being responsible for different kinds of coordination processes. The latter include matching buyers to sellers, managing ongoing negotiation or auction processes, setting prices and quantities based on buyers' and sellers' bids or offers. We are hopeful that the representational technique of social structures will enable various investigations of computational complexity (for example, the trade-offs between time complexity and space complexity), scaling issues of large agent societies, treating emergent phenomena as a means of effective control, and a unified view of DAI methods and concepts.