SystemIntelligenceModel,Bounds.PDFVIP

System Intelligence: Model, Bounds and Algorithms Longbo Huang IIIS@Tsinghua University longbohuang@ ABSTRACT to enhance user experience. For instance, recommendation We present a general framework for understanding system systems [1], smart home [2], arti?cial intelligence engines intelligence, i.e., the level of system smartness perceived by [3], and user behavior prediction [4]. Despite the prevail- users, and propose a novel metric for measuring intelligence ing success in practice, there has not been much theoretical levels of dynamical human-in-the-loop systems, de?ned to understanding about system smartness. In particular, how be the maximum average reward obtained by proactively do we measure the intelligence level of a system? How do serving user demands, subject to a resource constraint. Our we compare two systems and decide which one is smarter? metric captures two important elements of smartness, i.e., What elements in a system contribute most to the level of being able to know what users want and pre-serve them, smartness? Can the intelligence level of a system be pushed and achieving good resource management while doing so. arbitrarily high? We provide an explicit characterization of the system in- Motivated by these fundamental questions, in this paper, telligence, and show that it is jointly determined by user we propose a general framework for modeling system smart- demand volume (opportunity to impress), demand correla- ness and propose a novel metric for measuring system intelli- tion (user predictability), and system resource and action gence. Speci?cally, we consider a discrete time system where costs (?exibility to pre-serve). a server serves a set