第 1 章 多智能体系统导论 / Chapter 1: Introduction to Multi-Agent Systems

第 1 章 多智能体系统导论 / Chapter 1: Introduction to Multi-Agent Systems

什么是多智能体系统（Multi-Agent System, MAS），它为何在短短三十年间从分布式人工智能（Distributed Artificial Intelligence, DAI）的一个分支，跃升为现代人工智能研究的核心范式之一？要回答这个问题，我们必须追溯 Agent 概念的历史脉络，理解智能体区别于普通软件程序的核心属性，把握 MAS 相对单智能体系统的结构性优势，并考察其在机器人、互联网、金融与智慧城市等领域的真实落地。本章作为全书的导论，将沿着"历史—定义—对比—应用—研究问题—本书结构"的逻辑链条，为读者建立关于多智能体系统的整体认知框架，并自然过渡到第 2 章对智能体内部架构的深入讨论。

What is a multi-agent system (MAS), and why has it evolved from a branch of Distributed Artificial Intelligence (DAI) into one of the central paradigms of modern AI research within just three decades? To answer this question, we must trace the historical development of the Agent concept, understand the core properties that distinguish agents from ordinary software programs, grasp the structural advantages of MAS over single-agent systems, and examine its real-world deployment in robotics, the Internet, finance, and smart cities. This chapter, serving as the introduction to the entire book, follows the logical chain of "history—definition—comparison—applications—research problems—book structure" to establish a comprehensive cognitive framework for MAS and naturally leads into Chapter 2's deeper discussion of internal agent architectures.

1.1 分布式人工智能的起源与 Agent 范式形成 / Origins of DAI and the Emergence of the Agent Paradigm

多智能体系统并非凭空出现的概念，它扎根于分布式人工智能五十年的研究积累之中。理解这段历史，有助于我们看清 MAS 当下形态的形成逻辑，并预判其未来走向。

MAS did not emerge from nowhere; it is rooted in fifty years of research in Distributed Artificial Intelligence. Understanding this history helps us see how the current form of MAS was shaped and allows us to anticipate its future trajectory.

1970 年代——DAI 的思想萌芽 / The 1970s: Conceptual Seeds of DAI

分布式人工智能的真正起点，可以追溯到 1970 年代对"分布式问题求解"（Distributed Problem Solving, DPS）的探索。1975 年，MIT 的 Marvin Minsky 在其著名论文《A Framework for Representing Knowledge》中提出了"心智社会"（Society of Mind）理论，主张智能不是某个单一模块的产物，而是大量简单"代理"相互作用的涌现结果。Minsky 写道："You can build a mind from many little parts, each mindless by itself." 这一思想虽然偏哲学化，但它第一次系统地把"多个相互作用的智能单元"作为智能的根本来源，为后续 Agent 理论埋下了种子。

The true beginning of DAI can be traced to the 1970s and the exploration of Distributed Problem Solving (DPS). In 1975, Marvin Minsky at MIT proposed his "Society of Mind" theory in "A Framework for Representing Knowledge," arguing that intelligence is not the product of a single module but the emergent result of many simple "agents" interacting with one another. He famously wrote, "You can build a mind from many little parts, each mindless by itself." Although largely philosophical, this view was the first to systematically propose "many interacting intelligent units" as the fundamental source of intelligence, planting the seed for later Agent theory.

同期，Carl Hewitt 在 1973 年提出的 Actor 模型（Actor Model）则为分布式计算提供了第一个严格的数学基础。Actor 模型把并发计算的基本单元定义为"Actor"，每个 Actor 拥有独立状态、通过异步消息通信、能并发执行。该模型后来深刻影响了面向 Agent 的软件工程方向。

In parallel, Carl Hewitt proposed the Actor Model in 1973, providing the first rigorous mathematical foundation for distributed computing. The Actor Model defines the basic unit of concurrent computation as an "Actor"; each actor has independent state, communicates via asynchronous messages, and executes concurrently. This model subsequently influenced Agent-oriented software engineering profoundly.

1980 年代——Agent 概念的形式化 / The 1980s: Formalization of the Agent Concept

1980 年代是 Agent 概念走向形式化的关键十年。1986 年，Yoav Shoham 在 Stanford 的博士论文中首次将面向 Agent 的编程范式（Agent-Oriented Programming, AOP）作为独立研究方向提出，主张直接以"信念—愿望—意图"（Belief-Desire-Intention, BDI）作为基本编程原语。几乎同时，Michael Wooldridge、Nick Jennings 等人在伦敦大学国王学院开始建立 Agent 的形式化语义基础，他们的工作后来成为 MAS 教科书的经典内容。

The 1980s were the crucial decade in which the Agent concept was formalized. In 1986, Yoav Shoham's PhD thesis at Stanford first proposed Agent-Oriented Programming (AOP) as an independent research direction, advocating the use of Belief-Desire-Intention (BDI) directly as basic programming primitives. Around the same time, Michael Wooldridge and Nick Jennings at King's College London began establishing the formal semantic foundations of agents—work that later became classic textbook content.

1990 年代——MAS 形成独立学科 / The 1990s: MAS as an Independent Discipline

1990 年代见证了 MAS 作为独立学科的成型。1993 年，European AgentLink 项目的启动标志着欧洲学界开始系统性投入 Agent 研究；1995 年，第一届国际 MAS 会议（International Conference on Multi-Agent Systems, ICMAS）召开；1996 年，Wooldridge 与 Jennings 在《Knowledge Engineering Review》发表论文《Intelligent Agents: Theory and Practice》，系统总结了 Agent 与 MAS 的理论框架，被引用超过一万次，成为该领域的奠基文献之一。同年，IEEE 的 FIPA（Foundation for Intelligent Physical Agents）也开始制定 Agent 通信语言（Agent Communication Language, ACL）的国际标准，1997 年发布的 FIPA-ACL 至今仍是 Agent 通信的事实规范。

The 1990s witnessed MAS becoming an independent discipline. The launch of the European AgentLink project in 1993 signaled systematic investment in Agent research; the first International Conference on Multi-Agent Systems (ICMAS) was held in 1995; and in 1996, Wooldridge and Jennings published "Intelligent Agents: Theory and Practice" in the *Knowledge Engineering Review*, systematically summarizing the theoretical framework of agents and MAS—a paper cited over ten thousand times, and a foundational work in the field. That same year, IEEE's FIPA began developing international standards for Agent Communication Language (ACL); the 1997 FIPA-ACL specification remains the de facto standard for Agent communication to this day.

2000 年代——工程化与产业渗透 / The 2000s: Engineering and Industrial Penetration

进入 21 世纪，MAS 开始从理论走向大规模工程应用。2004 年，DARPA 在美国国防高级研究计划局的支持下举办了多届 RoboCup 救援仿真比赛，推动 MAS 在灾害响应场景中的研究。2006 年，亚马逊在其电商平台上推出"亚马逊市场"（Amazon Marketplace）级别的大规模多买家—多卖家自动化交易系统，其底层算法大量借鉴了 MAS 中的拍卖与协商机制。同期，Swarm Robotics 的研究在欧盟 Swarm-bots 项目中取得突破，验证了大规模机器人协作的可行性。

In the 21st century, MAS began transitioning from theory to large-scale engineering applications. With DARPA support, multiple RoboCup Rescue simulation competitions drove MAS research for disaster response scenarios starting in 2004. In 2006, Amazon launched its Marketplace platform involving massive multi-buyer/multi-seller automated transactions, with underlying algorithms drawing heavily from auctions and negotiation mechanisms in MAS. Concurrently, swarm robotics achieved breakthroughs through the EU Swarm-bots project, validating the feasibility of large-scale robotic cooperation.

2010 年代至今——深度学习与 MARL 的复兴 / From the 2010s Onward: The Renaissance of Deep Learning and MARL

2010 年代后期，深度强化学习的爆发为 MAS 注入了新的活力。多智能体强化学习（Multi-Agent Reinforcement Learning, MARL）成为研究热点，OpenAI、DeepMind 等机构在多人游戏（StarCraft、Dota 2）、机器人协作等场景中展示了 MAS 的强大能力。2019 年，DeepMind 的 AlphaStar 在 StarCraft II 中达到人类职业选手水平，其背后是复杂的多智能体协同策略。2020 年代以来，大语言模型（Large Language Model, LLM）与多智能体系统的结合开启了 Agentic AI 新范式，多个 LLM Agent 协同完成复杂任务成为研究前沿。

In the late 2010s, the explosion of deep reinforcement learning injected new vitality into MAS. Multi-Agent Reinforcement Learning (MARL) became a research hotspot, with institutions such as OpenAI and DeepMind demonstrating powerful MAS capabilities in multi-player games (StarCraft, Dota 2) and robotic collaboration. In 2019, DeepMind's AlphaStar achieved human professional-player level in StarCraft II, built upon complex multi-agent cooperative strategies. Since the 2020s, the combination of Large Language Models (LLMs) and MAS has opened a new Agentic AI paradigm, in which multiple LLM-based agents cooperate to complete complex tasks—a frontier research direction.

1.2 智能体的定义与三大核心属性 / Defining Agents: Definitions and Core Properties

历史脉络梳理完毕后，我们现在回到一个更基础的问题：究竟什么是智能体？什么样的软件实体才算真正的 Agent？本节将给出 Agent 的形式化定义与三大核心属性，为后续讨论奠定术语基础。

Having traced the historical development, we now return to a more fundamental question: what exactly is an agent? What kind of software entity qualifies as a true Agent? This section provides the formal definition of Agent and its three core properties, laying the terminological foundation for the following discussion.

智能体的弱定义与强定义 / Weak and Strong Notions of Agency

Wooldridge 在其经典教材《An Introduction to Mul