Information Avoidance in multi-agent systems

Abstract

This paper introduces a framework of assessing the Value of Information (VoI) for systems managed by multiple agents. The system is composed of some binary components, each managed by one agent. Each agent can decide to inspect and then to repair their component, and they are responsible for the repair cost (of their own component) and for a share of the cost of system failure. The information collected through inspections is available to all agents. The impact of each observation is assessed by predicting the agentsﾒ behavior in the prior and in the posterior setting. We assume agents adopt a Nash Equilibrium (NE) and, when multiple NEs exist, we discuss the adoption of ﾓglobally optimalﾔ and ﾓlocalﾔ (i.e. not optimal) NEs. The VoI of inspecting a component can then be derived as the difference of expected cost in the prior and posterior settings. Discussing various settings for series, parallel and general systems, under global and local NEs, with perfect or imperfect information, we show that the VoI can be negative for some agents, and this can trigger the phenomenon of Information Avoidance (IA), where rational agents prefer not to collect free information. We illustrate under what circumstances IA can occur.