English
When the internet became an economic platform — when search engines sold advertising by auction, when marketplaces matched buyers and sellers algorithmically, when platforms designed the rules by which millions of strategic agents interacted — computer science and economics collided, and the collision produced a genuinely new field. Algorithmic game theory and its applied counterpart, mechanism design, sit at the boundary where the strategic reasoning of economics meets the computational concerns of computer science. The questions they address belong fully to neither parent field: economics had the theory of strategic interaction but not the computational concerns that arise when the interactions are mediated by algorithms at scale; computer science had the algorithmic tools but not the framework for reasoning about self-interested agents. The fusion is where the action is.
The relationship has a particular character that distinguishes it from the others in this chapter: it is driven substantially by practice, by the enormous economic systems that computer science now builds and operates. The advertising auctions that fund the internet, the matching markets that assign medical residents to hospitals and students to schools, the pricing systems of ride-sharing and marketplace platforms, the emerging markets in computational resources and attention — these are economic mechanisms implemented as algorithms, designed by people who must understand both the economics (what incentives the mechanism creates) and the computation (whether the mechanism can be computed efficiently). The practical stakes are large, which has driven the theory, and the theory has in turn reshaped how both fields think.
This section examines the relationship through algorithmic game theory (the computational study of strategic interaction), mechanism design (the design of rules to achieve desired outcomes among strategic agents), and the deeper way that computational thinking has constrained and reshaped economic theory. It draws on the algorithmic foundations (§2.6) and complexity theory (§3.3), and engages economics — game theory, microeconomics, market design — that is outside this guide’s scope but whose fusion with computation is the subject.
Background: Algorithms (§2.6) and complexity (§3.3) provide the computational constraints; game theory, microeconomics, and market design provide the economic side of the relationship.
The Relationship: Where Strategy Meets Computation
Algorithmic Game Theory: The Computation of Strategic Interaction
Game theory, developed by von Neumann and Morgenstern (1944) and Nash (1950) and others, is the mathematical study of strategic interaction — how rational agents should act when their outcomes depend on others’ choices. It became central to economics, providing the framework for analyzing competition, bargaining, auctions, and any situation where agents’ interests interact. Algorithmic game theory asks the computational questions about this framework that economics did not: can the equilibria that game theory describes actually be computed? how efficiently? what happens to the theory when computation is taken into account?
The questions turn out to matter for the theory itself. The central solution concept of game theory is the Nash equilibrium — a set of strategies where no agent can improve by unilaterally changing their strategy. Nash proved that every finite game has at least one (possibly in mixed strategies). But Nash’s proof is non-constructive: it shows an equilibrium exists without showing how to find it. The computer science question — how hard is it to compute a Nash equilibrium? — was answered, after decades, by a result that computing a Nash equilibrium is complete for a complexity class called PPAD (Daskalakis, Goldberg, and Papadimitriou, 2009), which is believed to be intractable. This has a striking implication for economics: if computing an equilibrium is intractable, then the equilibrium concept loses some of its force as a prediction of what rational agents will do, because agents cannot be assumed to compute what no efficient algorithm can. The computational complexity of a solution concept bears on its validity as economic theory — a computer science result reshaping an economic concept.
The price of anarchy, introduced by Koutsoupias and Papadimitriou (1999), is another characteristically algorithmic contribution: it measures how much worse the outcome of self-interested behavior (an equilibrium) is than the outcome a benevolent central planner would achieve (the optimum). For a given system — a network where users route traffic selfishly, a market where agents act in self-interest — the price of anarchy quantifies the cost of decentralized selfish behavior relative to centralized optimization. This is a computer science question (it is about the quality of equilibria as measured against an optimization benchmark) applied to economic systems, and it has produced a body of results characterizing when selfish behavior is nearly optimal (low price of anarchy) and when it is disastrous (high price of anarchy). It is a tool for understanding when systems can be left to self-organize and when they need intervention — a question that matters for designing any system where strategic agents interact.
Mechanism Design: Engineering the Rules of the Game
Mechanism design is sometimes called “reverse game theory,” and the inversion captures its character: instead of analyzing how agents behave under given rules, it designs the rules so that the agents’ self-interested behavior produces a desired outcome. It is the engineering discipline of the boundary — given a goal (allocate resources efficiently, elicit truthful information, maximize revenue), design the mechanism (the rules of interaction) so that rational, self-interested agents, acting in their own interest, produce that goal. The fusion with computer science is natural because a mechanism is, increasingly, an algorithm, and designing it requires both the economic analysis (what incentives does it create?) and the computational analysis (can it be computed efficiently?).
The foundational achievement of mechanism design is the design of incentive-compatible mechanisms — mechanisms in which truthful behavior is in each agent’s self-interest, so that the mechanism can rely on agents revealing their true preferences rather than gaming the system. The Vickrey auction (second-price sealed-bid auction, Vickrey 1961) is the classic example: in an auction where the highest bidder wins but pays the second-highest bid, each bidder’s optimal strategy is to bid their true valuation, because bidding otherwise cannot help and can hurt. The mechanism is designed so that truth-telling is the dominant strategy, which makes the auction’s outcome predictable and efficient. The Vickrey-Clarke-Groves (VCG) mechanism generalizes this to a broad class of settings, providing a general method for designing incentive-compatible mechanisms that achieve efficient outcomes. These results are economics (they are about incentives and equilibria), but their application at scale is computer science (the mechanisms must be computed, often in real time, over large numbers of agents and outcomes).
The internet made mechanism design enormously consequential, because the largest economic mechanisms in history are now algorithms running on internet platforms. Search advertising auctions — the mechanism by which search engines sell ad placement — allocate hundreds of billions of dollars and are run billions of times per day; their design (the generalized second-price auction, and its analysis and refinement) is mechanism design at scale, and getting it right has enormous economic stakes. Matching markets — the mechanisms that assign medical residents to hospitals (the National Resident Matching Program, using the Gale-Shapley deferred acceptance algorithm), students to public schools, organs to recipients — are mechanism design applied to allocation problems where money is not used, and their design (drawing on the matching theory that earned Roth and Shapley the 2012 economics Nobel) determines whether the allocation is stable, efficient, and fair. These are economic mechanisms whose design requires the computational thinking of computer science, implemented as the algorithms of computer science, with stakes that are among the largest in the economy.
Computation as a Constraint on Economic Rationality
The deepest contribution of computer science to economics is conceptual: the recognition that computation is a constraint on rationality, which challenges the idealization of perfect rationality at the heart of classical economic theory. Classical economics models agents as perfectly rational — able to compute the optimal action given their preferences and information. Computer science asks whether the optimal action can be computed, and the answer is often no: the computation is intractable, and a perfectly rational agent who must solve an intractable problem is a fiction.
This challenge has force precisely because computer science can make it rigorous. The bounded rationality that economists like Herbert Simon had argued for on general grounds — that real agents have limited computational resources and cannot achieve the perfect optimization that classical theory assumes — can be made precise with complexity theory: if finding the optimal action is NP-hard, then no agent (human or machine) with reasonable computational resources can reliably find it, and a theory that assumes they do is assuming the impossible. The computational complexity of the problems that economic agents are assumed to solve becomes a constraint on what economic theories can plausibly claim. This is computer science reshaping the foundations of economics — not by adding tools but by challenging the central idealization, with the rigor that complexity theory provides.
The reshaping runs deeper into how economic systems should be understood and designed. If agents are computationally bounded, then mechanisms should be designed to be easy for agents to interact with optimally (a mechanism where the optimal strategy is computationally hard to find is a bad mechanism, however good its theoretical properties), and the simplicity of the optimal strategy becomes a design criterion. If equilibria are hard to compute, then the dynamics by which systems reach (or fail to reach) equilibrium become important, not just the equilibria themselves. The computational lens changes what questions economics asks: not just what the equilibrium is, but whether it can be reached, how long it takes, and whether the agents can compute their parts of it. This is a genuine reshaping of economic thinking by computational concerns, and it is among the clearest cases of one field changing another’s foundations.
The dialogue runs back the other way as well: economics has given computer science the framework for reasoning about systems of self-interested agents, which has become essential as computer systems increasingly involve many parties with divergent interests. Distributed systems where nodes may be selfish, networks where users compete for resources, blockchains where participants are economically motivated, multi-agent AI systems where agents have different objectives — these are computer systems whose correct design requires the economic analysis of incentives. The computer scientist designing such a system must think like an economist about what incentives the design creates, and the economic framework for strategic interaction has become part of the computer scientist’s toolkit. The fields have become genuinely interdependent at this boundary.
What This Perspective Changes
Seeing the relationship between computer science and economics changes how a practitioner understands systems where strategic agents interact.
The first change is the recognition that systems involving multiple self-interested parties require economic thinking. The practitioner who has internalized this sees that a distributed system with selfish nodes, a platform with strategic users, a network with competing traffic, or a multi-agent AI system is not just a technical system but an economic one, whose behavior depends on the incentives the design creates. They ask not only “does this work technically?” but “what incentives does this create, and how will strategic agents respond?” — a question that determines whether the system behaves as intended when its participants pursue their own interests.
The second change is the mechanism-design perspective on system design. The practitioner who understands mechanism design approaches the design of systems involving strategic agents as the design of incentives: rather than assuming agents will behave as desired, they design the rules so that self-interested behavior produces the desired outcome. This is a powerful reframing — from controlling behavior to aligning incentives — and it applies wherever a system’s participants have their own interests, from auction design to the design of incentives in distributed systems.
The third change is calibration about rationality and computation. The practitioner who understands that computation constrains rationality holds the assumption of rational agents with appropriate skepticism: real agents are computationally bounded, optimal strategies may be intractable to compute, and a system or theory that assumes perfect rationality assumes something impossible. This calibration applies both to understanding economic systems (where the perfect-rationality assumption may fail) and to designing mechanisms (which should be easy for bounded agents to interact with well).
The fourth change is appreciation of the stakes and the dialogue. The practitioner who sees that the largest economic mechanisms in history are now algorithms — that search auctions and matching markets and platform pricing are mechanism design at enormous scale — appreciates that this boundary is not academic but consequential, governing the allocation of enormous resources. And they see the genuine two-way dialogue: computation reshaping economic theory’s foundations, and economics providing the framework that computer science needs for systems of strategic agents. The fields have become interdependent, and the practitioner who works at their boundary draws on both.
Resources
Foundational and Bridging Texts
The standard text for the fused field is Algorithmic Game Theory (Nisan, Roughgarden, Tardos, and Vazirani, eds., Cambridge, 2007, free online), the edited volume that defined the field, with chapters by its founders on the computation of equilibria, the price of anarchy, mechanism design, and the applications. It is the comprehensive reference for the boundary, written by the people who created it, and freely available. It assumes mathematical maturity and the algorithmic background of §2.6 and §3.3.
Tim Roughgarden’s Twenty Lectures on Algorithmic Game Theory (Cambridge, 2016) is the accessible entry, based on his Stanford course, covering mechanism design, auctions, the price of anarchy, and the computation of equilibria with the clarity of a course designed to teach the material rather than to reference it. His lecture videos (free on YouTube) accompany it, and together they are the best route into the field for someone with the algorithmic background.
For mechanism design and market design specifically, Roth’s Who Gets What — and Why (2015) is the accessible treatment of matching markets and market design by one of the field’s architects (and a Nobel laureate), explaining how markets are designed to be stable, efficient, and fair, with the real applications (medical residents, schools, organs) that make the stakes concrete. For the economics foundations, a game theory text (Osborne’s An Introduction to Game Theory, or the more advanced Fudenberg and Tirole) provides the economic theory that the algorithmic work builds on.
| Resource | Role | Type |
|---|---|---|
| Nisan, Roughgarden, Tardos & Vazirani, Algorithmic Game Theory (free) | The field-defining reference | Reference |
| Roughgarden, Twenty Lectures on Algorithmic Game Theory | Accessible entry | Entry |
| Roth, Who Gets What — and Why | Market design, accessibly | Entry |
| Osborne, An Introduction to Game Theory | Game theory foundations | Depth |
Key Results and Applications
For the computational complexity of equilibria, the Daskalakis-Goldberg-Papadimitriou work on the PPAD-completeness of Nash equilibrium is the landmark result; Papadimitriou’s surveys make it accessible. For the price of anarchy, Roughgarden’s Selfish Routing and the Price of Anarchy (2005) is the thorough treatment by the concept’s co-developer.
For the auction theory that underlies internet advertising, Krishna’s Auction Theory (2nd ed., 2009) is the standard economics text, and the literature on the generalized second-price auction (Edelman, Ostrovsky, and Schwarz, 2007; Varian, 2007) analyzes the specific mechanism that funds the internet. For matching, Roth and Sotomayor’s Two-Sided Matching is the foundational treatment, and the Gale-Shapley deferred acceptance algorithm (1962, the origin) is readable in the original.
Roughgarden’s Stanford courses (CS364A Algorithmic Game Theory, with free lecture notes and videos) are the best structured path through the field for the computationally-oriented learner.
| Resource | Role | Type |
|---|---|---|
| Roughgarden CS364A course (free) | Structured path through the field | Entry |
| Mechanism Design for Social Good (free) | Public-interest mechanism design research community | Reference |
| ACM EC proceedings / SIGecom resources (often free) | Economics and computation research community | Reference |
| Daskalakis et al., PPAD-completeness of Nash | Complexity of equilibria; landmark | Depth |
| Roughgarden, Selfish Routing and the Price of Anarchy | Price of anarchy, thorough | Depth |
| Krishna, Auction Theory (2nd ed.) | Auction theory foundations | Depth |
| Gale & Shapley, deferred acceptance (1962) | Matching; foundational paper | Depth |
Traps
| Trap | Why it misleads | Better response |
|---|---|---|
| Designing systems for strategic agents without economic thinking | A system whose participants have their own interests — a platform, a distributed system with selfish nodes, a marketplace, a multi-agent system — will be gamed if it is designed only for technical correctness without regard to the incentives it creates. Participants will exploit whatever the rules reward, which may not be what the designer intended. Technical correctness does not ensure good behavior when the participants are strategic. | Apply mechanism-design thinking: analyze what incentives the system creates and how strategic agents will respond, and design the rules so that self-interested behavior produces the desired outcome. Ask what behavior the mechanism rewards, and whether a strategic agent could exploit it. The system must be designed for the agents who will actually use it — self-interested ones — not for cooperative ones. |
| Assuming agents are perfectly rational | Both the design and the analysis of economic systems often assume agents are perfectly rational — able to compute their optimal strategy. Computer science shows this is often impossible: the optimal strategy may be intractable to compute, and an agent who cannot compute it will not play it. A mechanism whose good properties depend on agents finding a computationally hard optimal strategy will not achieve those properties in practice. | Account for the computational bounds on agents. Design mechanisms so that the optimal strategy is easy to find (simplicity of the optimal strategy is a design criterion), and analyze systems with the recognition that agents are computationally bounded and may not reach the equilibria that perfect rationality would. The computational complexity of the agents’ problem is part of the analysis, not a detail to be idealized away. |
| Treating equilibrium as automatically achieved | Game-theoretic analysis identifies equilibria, and it is tempting to assume the system will reach the equilibrium. But equilibria can be intractable to compute (Nash equilibrium is PPAD-complete), and even computable equilibria may not be reached by the dynamics of the actual system. The equilibrium is a fixed point, not a guarantee that the system arrives there. | Consider the dynamics, not just the equilibria. Ask whether the equilibrium can be computed, whether the system’s dynamics actually converge to it, and how long convergence takes. For systems that may not reach equilibrium, the out-of-equilibrium behavior is what actually occurs and must be understood. The existence of an equilibrium does not ensure the system realizes it. |
中文
当互联网成为一种经济平台——搜索引擎通过拍卖出售广告位，市场平台用算法匹配买家与卖家，平台设计规则让数百万策略性主体彼此互动——计算机科学与经济学发生了碰撞，而这场碰撞产生了一个真正的新领域。算法博弈论及其应用对应领域机制设计，位于经济学的策略推理与计算机科学的计算关切之间的边界。它们处理的问题并不完全属于任何一个母领域：经济学有关于策略互动的理论，却没有充分处理当互动由大规模算法中介时出现的计算问题；计算机科学有算法工具，却没有一套推理自利主体行为的框架。二者融合的地方，正是最关键的地方。
这种关系具有本章其他跨学科关系所没有的一种特殊性：它在很大程度上由实践推动，由计算机科学如今构建并运行的巨大经济系统推动。支撑互联网的广告拍卖、把住院医师分配给医院和把学生分配给学校的匹配市场、网约车和市场平台的定价系统、正在出现的计算资源与注意力市场——这些都是作为算法实现的经济机制。设计它们的人必须同时理解经济学，也就是机制会创造什么激励；以及计算，也就是机制能否被高效计算。实践后果巨大，因此推动了理论；而理论反过来也重塑了两个领域的思考方式。
本节将通过算法博弈论、机制设计，以及计算思维如何约束并重塑经济理论这条更深层线索，来考察计算机科学与经济学的关系。算法博弈论研究策略互动中的计算问题；机制设计研究如何设计规则，使策略性主体之间产生我们希望的结果。本节依托算法基础（§2.6）和复杂性理论（§3.3），同时涉及经济学——博弈论、微观经济学、市场设计——这些内容超出本指南的范围，但它们与计算的融合正是本节主题。
背景知识：算法（§2.6）和复杂性（§3.3）提供计算约束；博弈论、微观经济学和市场设计提供这种关系的经济学一侧。
这种关系：策略与计算相遇之处
算法博弈论：策略互动的计算问题
博弈论由冯·诺伊曼和摩根斯坦（1944）、纳什（1950）等人发展出来，是对策略互动的数学研究：当一个主体的结果取决于他人的选择时，理性主体应该如何行动。它后来成为经济学的核心，为分析竞争、谈判、拍卖，以及任何主体利益相互作用的情境提供框架。算法博弈论则追问经济学没有追问的计算问题：博弈论所描述的均衡真的能被计算出来吗？能以多高效率计算？一旦把计算纳入考虑，理论会发生什么变化？
事实证明，这些问题会反过来影响理论本身。博弈论的核心解概念是纳什均衡：一组策略，在其中没有任何主体能够通过单方面改变自己的策略而改善自身结果。纳什证明了每个有限博弈至少存在一个纳什均衡，可能是混合策略均衡。但纳什的证明是非构造性的：它证明均衡存在，却没有说明如何找到均衡。计算机科学提出的问题——计算一个纳什均衡到底有多难——经过数十年后，由一项结果回答：计算纳什均衡对于一个叫作 PPAD 的复杂性类是完全的，也就是 PPAD-complete（Daskalakis、Goldberg 和 Papadimitriou，2009），而这个问题通常被认为是难以处理的。这对经济学有一个很强的含义：如果计算均衡本身不可行，那么把均衡概念作为理性主体行为预测的力量就会减弱，因为不能假设主体能够计算出没有高效算法可以计算的东西。一个解概念的计算复杂性，会影响它作为经济理论的有效性——这就是计算机科学结果重塑经济学概念的例子。
无政府代价（price of anarchy）由 Koutsoupias 和 Papadimitriou（1999）提出，是另一种非常具有算法特征的贡献。它衡量自利行为的结果，也就是某个均衡，相比一个善意的中央规划者所能达到的最优结果差多少。对于一个给定系统——例如用户自私选择路径的网络，或主体按自身利益行动的市场——无政府代价量化的是去中心化自利行为相对于集中式优化的成本。这是一个计算机科学问题，因为它用优化基准衡量均衡质量；同时它又应用于经济系统，并产生了一整套结果，用来刻画什么时候自利行为接近最优，也就是无政府代价低；什么时候自利行为会造成灾难，也就是无政府代价高。它是理解系统何时可以自组织、何时需要干预的工具；而这对于设计任何有策略性主体互动的系统都很重要。
机制设计：设计游戏规则的工程
机制设计有时被称为“反向博弈论”（reverse game theory），这个反转准确捕捉了它的特征：它不是在既定规则下分析主体如何行动，而是设计规则，使主体的自利行为产生我们希望的结果。它是这一边界上的工程学科：给定一个目标，例如高效分配资源、诱导真实信息、最大化收入，然后设计机制，也就是互动规则，使理性且自利的主体在追求自身利益时产生这个目标。它与计算机科学的融合非常自然，因为机制越来越多地表现为算法；而设计机制需要同时进行经济学分析，也就是它会创造什么激励；以及计算分析，也就是它能否被高效计算。
机制设计的基础成就是设计激励相容机制（incentive-compatible mechanisms）：在这种机制中，诚实行为符合每个主体的自身利益，因此机制可以依赖主体揭示真实偏好，而不是操纵系统。维克里拍卖（Vickrey auction，也称第二价格密封投标拍卖，Vickrey 1961）是经典例子：在这种拍卖中，最高出价者获胜，但支付第二高出价；每个竞拍者的最优策略都是报出自己的真实估值，因为偏离真实出价不会带来好处，反而可能造成损失。机制被设计成使说真话成为占优策略，从而让拍卖结果可预测且有效率。Vickrey-Clarke-Groves 机制，即 VCG 机制，把这一思想推广到一大类场景中，为设计能实现有效率结果的激励相容机制提供了一般方法。这些结果属于经济学，因为它们讨论激励与均衡；但它们在大规模场景中的应用属于计算机科学，因为这些机制必须被计算出来，而且常常要在大量主体和大量可能结果上实时计算。
互联网让机制设计变得极其重要，因为历史上最大的经济机制如今都是运行在互联网平台上的算法。搜索广告拍卖——搜索引擎出售广告位置的机制——分配着数千亿美元，并且每天运行数十亿次；它的设计，也就是广义第二价格拍卖及其分析和改进，是大规模机制设计，而设计是否正确具有巨大的经济后果。匹配市场——例如把住院医师分配给医院的 National Resident Matching Program，它使用 Gale-Shapley 延迟接受算法；把学生分配给公立学校；把器官分配给受捐者——则是在不使用金钱的分配问题中应用机制设计。它们的设计依托为 Roth 和 Shapley 赢得 2012 年诺贝尔经济学奖的匹配理论，并决定分配是否稳定、高效且公平。这些都是经济机制，其设计需要计算机科学的计算思维，并作为计算机科学的算法来实现；而其影响则位于经济生活中最重大的层面。
计算作为经济理性的约束
计算机科学对经济学最深的贡献是概念性的：它指出计算是理性的约束条件，而这挑战了古典经济理论核心中的完全理性理想化。古典经济学把主体建模为完全理性者——能够根据自己的偏好和信息计算出最优行动。计算机科学则追问：这个最优行动是否真的能被计算出来？答案常常是否定的：计算不可处理，而一个必须解决不可处理问题的完全理性主体，只是一个虚构物。
这种挑战之所以有力量，正是因为计算机科学能把它严格化。Herbert Simon 等经济学家曾从一般层面提出有限理性：真实主体拥有有限计算资源，无法实现古典理论假定的完美优化。复杂性理论可以让这种有限理性变得精确：如果寻找最优行动是 NP-hard 的，那么任何拥有合理计算资源的主体，无论是人还是机器，都无法可靠地找到它；而一个假设他们能够做到这一点的理论，就是在假设不可能之事。经济主体被假定要解决的问题的计算复杂性，成为经济理论能否合理主张某些事情的约束条件。这是计算机科学在重塑经济学基础——不是增加工具，而是用复杂性理论提供的严格性，挑战经济学的核心理想化。
这种重塑还进一步深入到我们应如何理解和设计经济系统。如果主体受到计算约束，那么机制就应该被设计成让主体容易以最优方式参与；一个让最优策略很难计算的机制，即使理论性质再好，也不是好机制。最优策略的简单性因此成为设计标准。如果均衡很难计算，那么系统达到或无法达到均衡的动态过程就变得重要，而不仅仅是均衡本身。计算视角改变了经济学追问的问题：不只是均衡是什么，还包括它能否被达到、达到需要多长时间、主体能否计算出自己在其中的部分。这是计算关切对经济思想的真实重塑，也是一个领域改变另一个领域基础的最清楚案例之一。
对话也反向发生：经济学为计算机科学提供了推理自利主体系统的框架。随着计算机系统越来越多地涉及许多利益不同的参与方，这个框架变得不可或缺。节点可能自利的分布式系统、用户竞争资源的网络、参与者受经济激励驱动的区块链、主体目标彼此不同的多智能体 AI 系统——这些都是计算机系统，但其正确设计需要对激励进行经济学分析。设计这类系统的计算机科学家必须像经济学家一样思考：这个设计会创造什么激励？于是，关于策略互动的经济学框架，已经成为计算机科学家的工具箱的一部分。在这一边界处，两个领域已经真正相互依赖。
这种视角会改变什么
看清计算机科学与经济学的关系，会改变实践者对策略性主体互动系统的理解。
第一种改变，是认识到涉及多个自利参与方的系统需要经济学思维。内化这一点的实践者会看见：一个包含自利节点的分布式系统、一个拥有策略性用户的平台、一个存在竞争流量的网络、一个多智能体 AI 系统，不只是技术系统，也是经济系统；它的行为取决于设计所创造的激励。他们不只会问“这在技术上能不能工作？”，还会问“这会创造什么激励？策略性主体会如何回应？”——而这个问题决定了当参与者追求自身利益时，系统是否会按设计意图运行。
第二种改变，是获得机制设计式的系统设计视角。理解机制设计的实践者，会把涉及策略性主体的系统设计理解为激励设计：不是假设主体会按期望行动，而是设计规则，使自利行为产生期望结果。这是一种有力的重构——从控制行为，转向对齐激励——它适用于任何参与者拥有自身利益的系统，从拍卖设计，到分布式系统中的激励设计。
第三种改变，是校准对理性和计算的理解。理解计算会约束理性的实践者，会带着适当怀疑看待“理性主体”的假设：真实主体受到计算限制，最优策略可能不可处理，一个系统或理论如果假设完全理性，就是在假设某种不可能的东西。这种校准既适用于理解经济系统，也就是完全理性假设可能失效的地方；也适用于设计机制，也就是机制应该让受限主体能够容易地良好参与。
第四种改变，是认识到问题的重要性与双向对话。实践者如果看见历史上最大的经济机制如今都是算法——搜索广告拍卖、匹配市场、平台定价都是超大规模机制设计——就会理解这个边界并非纯学术，而是具有真实后果，支配着巨量资源的分配。他们也会看见真正的双向对话：计算正在重塑经济理论的基础，而经济学则为计算机科学处理策略性主体系统提供所需框架。两个领域已经相互依赖；在其边界处工作的实践者，必须同时动用二者。
资源
基础与桥接文本
这个融合领域的标准文本是 Algorithmic Game Theory（Nisan、Roughgarden、Tardos 和 Vazirani 编，Cambridge，2007，网上免费）。这是一部定义该领域的论文集，由该领域的奠基者撰写章节，覆盖均衡计算、无政府代价、机制设计及应用。它是这个边界领域的综合参考书，由创造这个领域的人写成，并且可以免费获取。它假定读者具有数学成熟度，以及 §2.6 和 §3.3 所需的算法背景。
Tim Roughgarden 的 Twenty Lectures on Algorithmic Game Theory（Cambridge，2016）是更易进入的入口，基于他的 Stanford 课程，覆盖机制设计、拍卖、无政府代价和均衡计算，讲法具有课程教学的清晰性，而不是单纯作为参考书。配套的讲课视频可在 YouTube 免费观看；书和视频合在一起，是具有算法背景的学习者进入这个领域的最佳路径。
对于机制设计和市场设计，Roth 的 Who Gets What — and Why（2015）是一本易读读物，由该领域的设计者之一、诺贝尔奖得主撰写，解释市场如何被设计得稳定、高效和公平，并通过住院医师、学校、器官等真实应用让问题的重要性变得具体。对于经济学基础，博弈论教材，例如 Osborne 的 An Introduction to Game Theory，或更高级的 Fudenberg 和 Tirole，可以提供算法研究所依托的经济学理论。
| 资源 | 作用 | 类型 |
|---|---|---|
| Nisan、Roughgarden、Tardos & Vazirani，Algorithmic Game Theory（免费） | 定义该领域的参考书 | 参考 |
| Roughgarden，Twenty Lectures on Algorithmic Game Theory | 易读入口 | 入门 |
| Roth，Who Gets What — and Why | 市场设计的易读处理 | 入门 |
| Osborne，An Introduction to Game Theory | 博弈论基础 | 深入 |
关键结果与应用
关于均衡的计算复杂性，Daskalakis-Goldberg-Papadimitriou 关于纳什均衡 PPAD 完全性的工作是里程碑结果；Papadimitriou 的综述则让这个结果更容易进入。关于无政府代价，Roughgarden 的 Selfish Routing and the Price of Anarchy（2005）是由该概念共同发展者写成的深入处理。
关于支撑互联网广告的拍卖理论，Krishna 的 Auction Theory（第 2 版，2009）是标准经济学教材；而关于广义第二价格拍卖的文献（Edelman、Ostrovsky 和 Schwarz，2007；Varian，2007）分析了支撑互联网运转的具体机制。关于匹配，Roth 和 Sotomayor 的 Two-Sided Matching 是基础性处理；Gale-Shapley 延迟接受算法（1962，源头论文）则适合直接阅读原文。
Roughgarden 的 Stanford 课程（CS364A Algorithmic Game Theory，含免费讲义和视频）是面向计算方向学习者系统进入该领域的最佳结构化路径。
| 资源 | 作用 | 类型 |
|---|---|---|
| Roughgarden CS364A 课程（免费） | 该领域的结构化路径 | 入门 |
| Mechanism Design for Social Good（免费） | 公益机制设计研究共同体 | 参考 |
| ACM EC 论文集 / SIGecom 资源（通常免费） | 经济学与计算研究共同体 | 参考 |
| Daskalakis 等，纳什均衡的 PPAD 完全性 | 均衡复杂性的里程碑结果 | 深入 |
| Roughgarden，Selfish Routing and the Price of Anarchy | 无政府代价的深入处理 | 深入 |
| Krishna，Auction Theory（第 2 版） | 拍卖理论基础 | 深入 |
| Gale & Shapley，延迟接受算法（1962） | 匹配理论的基础论文 | 深入 |
陷阱
| 陷阱 | 为什么会误导 | 更好的回应 |
|---|---|---|
| 为策略性主体设计系统，却没有经济学思维 | 如果一个系统的参与者拥有自己的利益——平台、带有自利节点的分布式系统、市场、多智能体系统——那么只考虑技术正确性，而不考虑它创造的激励，系统就会被操纵。参与者会利用规则奖励的一切，而这未必是设计者真正希望的东西。当参与者具有策略性时，技术正确性并不能保证良好行为。 | 使用机制设计思维：分析系统创造了什么激励，策略性主体会如何回应，并设计规则，使自利行为产生期望结果。追问机制奖励了什么行为，以及策略性主体是否能利用它。系统必须为实际会使用它的主体设计——也就是自利主体——而不是为合作主体设计。 |
| 假设主体完全理性 | 经济系统的设计和分析常常假设主体完全理性，能够计算自己的最优策略。计算机科学表明，这常常是不可能的：最优策略可能不可处理；一个无法计算最优策略的主体就不会采取它。如果一个机制的良好性质依赖于主体找到一个计算上困难的最优策略，那么它在实践中不会实现那些性质。 | 考虑主体的计算边界。设计机制时，应使最优策略容易找到；最优策略的简单性是一项设计标准。分析系统时，也应承认主体受计算约束，未必能达到完全理性所预测的均衡。主体问题的计算复杂性是分析的一部分，不是可以被理想化抹掉的细节。 |
| 把均衡看作会自动实现 | 博弈论分析会识别均衡，因此很容易假设系统会到达均衡。但均衡可能很难计算，例如纳什均衡是 PPAD-complete；即使均衡可计算，真实系统的动态过程也未必会到达它。均衡是一个不动点，不是系统一定会抵达那里的保证。 | 考察动态过程，而不只是考察均衡。追问均衡是否能被计算，系统动态是否真的收敛到它，以及收敛需要多长时间。对于可能无法达到均衡的系统，真正发生的是非均衡状态下的行为，而这也必须被理解。均衡存在，并不保证系统会实现它。 |