The choice of communication strategy is a basic and important problem in cooperation. Communicating adequate information with suitable persons (relevance) at appropriate times (timeliness) influences the efficiency of cooperative problem-solving. In particular, information which cannot be obtained from other viewpoints or which is time-consuming to get must be effectively used. This is the same as in cooperation between computer agents. The strength and efficiency of problem solving, and the quality of the results, reflect the communication strategy. Communication strategies are closely linked to planning, scheduling, and control strategies. There are several variations of these strategies, such as ``analyzing the data as soon as it is delivered'' and ``postponing the analysis,'' and total coherence and efficiency result from harmony between the various strategies chosen. Although there are many important works on the relationship between communication strategy and performance, we cannot decide a unique best communication strategy, because what is best depends on the features of problems, performance of each agent, kinds of knowledge, internal inference methods, and communication costs.
We are now developing an expert system, LODES, capable of diagnosing internetwork problems, which are increasing due to growth in network size and complexity. As shown in Fig.1(sorry, it is not shown), each network segment has its own LODES that isolates and resolves network problems cooperatively or autonomously, depending on the nature of the problem. Although physical problems are out of their scope, these LODESs can diagnose logical problems such as protocol violations, incorrect IP address assignments, and incorrect routings. Management data for a network segment are kept in the segment's LODES. Packets from hosts or IP routers are observed by the LODES attached to that network segment; other LODES can only observe packets indirectly via a number of intermediary networks and IP routers. Thus, the cause of a problem is detected only by the LODES that monitors the network where the cause is located.
In the case of internetwork problems, communication costs between agents vary according to the network environment. The cooperating agents are dynamically chosen depending on the kind of problem. For example, agent1 and agent2 may be connected with fast lines, but agent1 and agent3 with slow lines. The best communication strategy for cooperation in the former case is often the worst in the latter. Communication lines also include intermediary IP routers and networks that may have a problem, so communication costs can vary. Furthermore, we must avoid making problems worse due to communications between agents. Decision of the communication strategy is, therefore, an important issue.
This paper proposes a number of communication strategies and experimentally analyzes their performances. Moreover, it proposes that, by analyzing these data, a communication strategy can be dynamically determined by network performance observations and negotiations in the first stage of cooperation.