Minimizing Information-Centric Convergence Cost in Multi-Agent Agreement Problems

The basic paradigm of learning has shifted significantly, from single agents that learn in single, static environments, to collective learning: multiple, interacting agents with diverse goals learning from each other across different local environments. Instances of collective learning abound in sensor networks, peer-to-peer systems, distributed recommender systems, and in social systems in general. A critical, but unexplored, activity in collective learning is information gathering: the exchange of information about other agents’ individual preferences, that will guide collective decision making processes. A set of tradeoffs exists between the amount of information agents gather, the effort of this information gathering, and agents individual and collective performance. Reducing the amount of information gathered may reduce information costs, but reduced information can produce interpretation errors that create suboptimal behavior in the agents and the collective. In this paper we define and study the impact of these tradeoffs using the well known “Generalized Simple Majority” decision process in the model problem of norm emergence, a type of multi-agent agreement process in which agents converge to a common strategy. We present a new metric,“Information-Centric Convergence Cost”(ICCC), that combines information cost with the cost of time, and a new decision process, “Generalized Simple Sampled Majority,” and we study these in several agent network topologies. Surprisingly, we find that as the level of information gathering is reduced the amount of error increases non-linearly, giving a non-linear impact on performance. Thus the careful manipulation of information-processing effort can minimize ICCC while still achieving quick norm emergence.