Analyzing the Need for Meta-Level Communication
Keith Decker, Victor Lesser · 1993
In naturally distributed, homogeneous, cooperative problem solving environments where welldefined tasks arrive at multiple locations, decisions must be made about the extent of, and overlap between, each agent’s area of responsibility—the agents ’ organization. The organization may be constructed statically by a system designer, or dynamically by the agents during problem solving. No one organization is optimal across environments or even specific problem solving instances [6, 7, 8]. This paper presents an analysis of static and dynamic organizational structures for this class of environments, exemplified by distributed sensor networks. We first show how the performance of any static organization can be statistically described, and then show under what conditions dynamic organizations do better and worse than static ones. Finally, we show how the variance in the agents’ performance leads to uncertainty about whether a dynamic organization will perform better than a static one given only agent aprioriexpectations. In these cases, we show when meta-level communication about the actual state of problem solving will be useful to agents in constructing a dynamic organizational structure that outperforms a static one. Viewed in its entirety, this paper also presents a methodology for answering questions about the design of distributed problem solving systems by analysis and simulation of the characteristics of a complex environment rather than by relying on single-instance examples.