Emergent properties of referral systems
Proceedings of the second international joint conference on Autonomous agents and multiagent systems - AAMAS '03
Agents must decide with whom to interact, which is nontrivial when no central directories are available. A classical decentralized approach is referral systems, where agents adaptively give referrals to one another. We study the emergent properties of referral systems, especially those dealing with their quality, efficiency, and structure. Our key findings are (1) pathological graph structures can emerge due to some neighbor selection policies and (2) if these are avoided, quality and
... depend on referral policies. Further, authorities emerge automatically and the extent of their relative authoritativeness depends on the policies. with one another to find trustworthy agents with whom to interact. Thus, they form a referral system. Notice that trust applies both to the ultimate service provider and to the agents who contribute referrals to that provider. The agents are autonomous. That is, an agent may or may not respond to another agent by providing a service or a referral. When an agent does respond, there are no guarantees about the quality of the service or the suitability of a referral. Likewise, we do not assume that any agent should necessarily be trusted by others: an agent unilaterally decides how to rate another principal. The above model addresses the important challenge of finding trustworthy agents, which is nontrivial in open systems. First, referrals can apply even in the absence of centralized authorities and even when regulations may not ensure that services are of a suitable quality. Second, because service needs are often context-sensitive, a response from an agent can potentially benefit from the knowledge that it has of the other's needs. Contributions. The idea of referrals in multiagent systems goes back a long time. Referrals have been used in specific applications (see Section 5). However, we treat referrals as the key organizing principle for large-scale multiagent systems. Our objective in this paper is to study the conceptual aspects of referral systems, especially with respect to their emergent properties. Consequently, we first study how referral policies influence the quality and efficiency of referral systems and how these policies influence the way authorities emerge. Next, we study the structural properties of referral systems. Recent work has studied the structure of the Web as it happens to have emerged mostly through links on human-generated, static pages. Whereas existing work takes an after-the-fact look at Web structure, we can study the emerging structure of a referral system as it relates to the local policies of its members. Links over which parties request or give referrals and the referrals they give induce a natural structure on a referral system, leading to two important consequences. First, major application classes can be modeled via different structures. Second, the structure evolves in interesting ways based on the policies followed by the different parties. To this end, we identify pathological graph structures that can emerge. Organization. Section 2 gives additional details on our model of referrals among autonomous agents, possible applications domains, and our experimental setup. Section 3 introduces metrics to measure quality of networks, and evaluates the performance of agent policies. Section 4 characterizes some possible structures of the network in terms of their emergence and desirability for referral systems. Section 5 discusses the relevant literature and motivates directions for further work. © denotes the weight of the sociability in choosing a neighbor. When © is set to 0, the Providers policy is in effect. When © is set to 1, the Sociables policy is in effect. Other values of © measure weighted averages of the sociability and expertise. In our simulation, each agent selects neighbors after every two queries, and each simulation is run for four neighbors changes. The four points on the plot lines correspond to each neighbor change. OBSERVATION 7. Reflexive interest clustering decreases with an increase in quality.