Congestion Management by Interval Signaling

We consider a multi-agent system where, at every time instant, many users choose to use one of multiple resources, whose performance depends on the number of concurrent users. In addition, a central agent has up-to-date knowledge of the congestion across all resources. What signals should the central agent broadcast to the individual users in order to reduce the total congestion? We model users with varying levels of aversion to risk, and a heterogeneous population of users over these levels. We consider signaling schemes that communicate for each resource an interval of possible congestion values, instead of scalar values. We show how to optimise over these intervals with respect to the social cost, under constraints that these intervals be consistent with past observations. Moreover, under mild assumptions, the resulting distribution of users across resources converges.