Modelling Dynamic Forgetting in Distributed Information Systems

We describe and model a new aspect in the design of distributed information systems. We build upon a previously described problem on the microlevel, which asks how quickly agents should discount (forget) their experience: If they cherish their memories, they can build their reports on larger data sets; if they discount quickly, they can respond well to change in their environment. Here, we argue that on the macro-level, where agents disseminate information, the coordination of these micro-level strategies of discounting can have significant consequences on the system performance if the environment is uncertain. In our proposed model, a referral network disseminates information about a disruptive environment (a service provider) to a risk-averse client agent, who uses this information to maximise his profit and then gives feedback into the referral system. We model two simple strategies to dynamically find better discounting factors, through central and decentral control. We show that with dynamic discounting rates, the system can become more reactive. We discuss interdependence of the system components in the light of differing discounting scenarios. In this work, we build on a certainty-based trust representation and operators for it in referral systems, developed by Josang [7] and Hang, Wang and Singh [13,2].