Domain Monotonicity and the Performance of Local Solutions Strategies for CDPS-based Distributed Sensor Interpretation and Distributed Diagnosis

摘要

The growth in computer networks has created the potential to harness a great deal of computing power, but new models of distributed computing are often required. Cooperative distributed problem solving (CDPS) is the subfield of multi-agent systems (MAS) that is concerned with how large-scale problems can be solved using a network of intelligent agents working together. Building CDPS systems for real-world applications is still very difficult, however, in large part because the effects that domain and strategy characteristics have on the performance of CDPS systems are not well understood. This paper reports on the first results from a new simulation-based analysis system that has been created to study the performance of CDPS-based distributed sensor interpretation (DSI) and distributed diagnosis (DD). To demonstrate the kind of results that can be obtained, we have investigated how the monotonicity of a domain affects the performance of a potentially very efficient class of strategies for CDPS-based DSI/DD. Local solutions strategies attempt to limit communications among the agents by focusing on using the agents' local solutions to produce global solutions. While these strategies have been described as being important for effective CDPS-based DSI/DD, they need not perform well if a domain is nonmonotonic. We had previously suggested that the reason they have performed well in several research systems was that many DSI/DD domains are what we termed nearly monotonic. In this paper, we will provide quantitative results that relate the performance of local solutions strategies to the monotonicity of a domain. The experiments confirm that domain monotonicity can be important to consider, but they also show that it is possible for these strategies to be effective even when domains are relatively nonmonotonic. What is required is that the agents receive a significant fraction of the data that is relevant to their subproblems. This has important implications for the design of DSI/DD systems using local solutions strategies. In addition, while the work indicates that many DSI/DD domains are likely to be “nearly monotonic” according to our original definitions, it also shows that these measures are not as predictive of performance as other measures we define. This means that near monotonicity alone does not explain why local solutions strategies have performed well in previous systems. Instead, a likely explanation is that these systems typically involved only a small number of agents.