Data-Driven Market Design

市场是通过有约束力的合同促进商品和服务交换的机构；因此，它们是按照规则运行的。有时，这些规则是有机产生的。当买家和卖家很容易找到对方时，当谁进行买卖并不是非常重要的时候，当市场产生简单的合同时(例如，“我以2美元的价格把这罐可乐卖给你”)，这种方法就能很好地发挥作用。否则，市场可能更难确定有效的规则(例如，制定类似“明天我将以60美元的价格将我的空余房间租给你，但你必须在晚上11点后保持安静”的合同)。 当有效市场不是有机地产生时，规则可以被明确地制定出来，这是市场设计和机制设计领域所倡导的想法(分别获得2012年和2007年的诺贝尔奖认可)。目标是证明一个市场在问题施加的约束和对参与者行为的合理假设下实现了理想的结果(例如，匹配从交易中获得最大收益的买家和卖家)。这些假设通常是博弈论的：粗略地说，参与者被充分告知市场是如何运作的，并在市场内“理性”行事，以最大限度地服务于自己的利益。这是一种强大的方法；它既产生了优雅、普遍的理论，也产生了影响深远的应用，从搜索引擎关键字拍卖到肾脏交易。它还对人工智能产生了深远的影响，为解决多智能体系统中的信息融合和任务分配等长期挑战提供了实用的、理论上站稳脚跟的技术。 这种方法有一个严重的缺陷，与20世纪中期该油田奠基时相比，2016年的缺陷更加严重。这个缺陷是，市场设计几乎完全是一种分析(即数学)练习：一旦一个人致力于世界的博弈论模型，就几乎没有空间对现实世界的观察做出反应。相比之下，计算机科学目前正在经历一场数据科学革命：我们现在认为计算机系统不是静态的人工制品，而是记住用户交互并适应它们的不断演变的服务。关于用户与系统(自动驾驶汽车、语音识别系统、搜索引擎)交互的数据越多，它的运行效果就越好，这一点正成为一个不言而喻的事实。 拟议的研究将有助于市场设计成为这一范式转变的一部分，使市场能够联合利用与用户的实际互动和博弈论分析。更具体地说，它将开发数据敏感技术，用于对市场中的人类行为进行建模，构建启发式清算算法，并分析适应机制。其结果将是市场设计可以根据不同的设置进行优化，并在部署后可以适应，就像其他现代计算机系统一样。

Markets are institutions that facilitate the exchange of goods and services via binding contracts; they thus run according to rules. Sometimes these rules arise organically. This can work well when buyers and sellers have little trouble finding each other, when it's not very important who does the buying and selling, and when the market produces simple contracts (e.g., “I'll sell you this can of Coke for $2”). Otherwise, it can be harder for markets to determine effective rules (e.g., to produce contracts like “I'll rent you my spare room tomorrow for $60, but you have to be quiet after 11 PM”). When effective markets do not arise organically, rules can instead be crafted explicitly, an idea championed by the fields of market design and mechanism design (recognized by Nobel prizes in 2012 and 2007 respectively). The goal is to prove that a market achieves desirable outcomes (e.g., matching up buyers and sellers who gain the most by trading) under the constraints imposed by a problem and under reasonable assumptions about the behaviour of participants. These assumptions are typically game theoretic: roughly, that participants are fully informed about how a market works and act “rationally” within it to best serve their interests. This is a powerful approach; it has yielded both elegant, general theory and deeply impactful applications as varied as search-engine keyword auctions and kidney exchanges. It has also had a profound impact on artificial intelligence, providing practical, theoretically grounded techniques for addressing longstanding challenges like information fusion and task allocation in multiagent systems. This approach has a critical flaw, which is more egregious in 2016 than it was when the field's foundations were being laid in the mid-1900s. This flaw is that market design is almost entirely an analytic (i.e., mathematical) exercise: once one has committed to a game theoretic model of the world, there is little room left for responsiveness to real-world observations. In contrast, computer science is currently undergoing a data science revolution: we now think of computer systems not as static artefacts, but as evolving services that remember user interactions and adapt to them. It is becoming a truism that the more data one has about user interactions with a system (a self-driving car; a speech recognition system; a search engine) the better it should work. The proposed research will help market design to become part of this paradigm shift, enabling markets to draw jointly on actual interactions with users and on game theoretic analysis. More specifically, it will develop data-sensitive techniques for modeling human behavior in markets, building heuristic clearing algorithms, and analyzing adaptive mechanisms. The result will be market designs that can be optimized to different settings and that can adapt after being deployed, just like other modern computer systems.