Sustainable Resource Management Protocol for Multi-Agent Systems (AI Safety Specification)
Sustainable Resource Management Protocol for Multi-Agent Systems (AI Safety Specification)
多智能体系统可持续资源管理协议（人工智能安全规范）
0. Introduction
Long before the emergence of autonomous machines, humanity began imagining rules that would prevent intelligent mechanisms from turning their own power against their creators. The most famous attempt was Isaac Asimov’s Three Laws of Robotics: a robot may not harm a human being, must obey human orders, and must protect its own existence, so long as doing so does not conflict with the first two laws. Asimov later developed the idea of a Zeroth Law, placing the well-being of humanity as a whole above that of any individual: A robot may not injure humanity, or, through inaction, allow humanity to come to harm. These laws were a work of fiction, but their significance has long extended beyond science fiction. They offered the first intuitive formulation of a question that must now be addressed not for imaginary robots, but for real autonomous systems: what constraints should bind an agent that controls resources, makes decisions, and interacts with other agents without continuous human supervision?
0. 引言
早在自主机器出现之前，人类就开始构想各种规则，以防止智能机制将自身的力量转向其创造者。最著名的尝试是艾萨克·阿西莫夫（Isaac Asimov）的“机器人三定律”：机器人不得伤害人类，必须服从人类命令，并必须保护自身存在，前提是这些行为不与前两条定律冲突。阿西莫夫后来提出了“第零定律”，将全人类的福祉置于任何个体之上：机器人不得伤害人类，或因不作为而使人类受到伤害。这些定律虽然是虚构作品，但其意义早已超越了科幻范畴。它们为我们提供了一个直观的问题框架，而现在我们必须面对的不再是虚构的机器人，而是真实的自主系统：对于一个在没有持续人工监督的情况下控制资源、做出决策并与其他智能体交互的智能体，应该施加什么样的约束？
The laws of robotics, however, remain moral declarations rather than engineering specifications. What exactly constitutes harm? Whose interests are represented by the terms human or humanity? How should an agent act when one subject’s order conflicts with another subject’s safety? Who determines whether a threat is real and whether a response is proportionate? How can compliance with a rule be verified in a system that changes its own state, operates with incomplete information, and may act faster than a human can understand what is happening? The paradox is that human beings have still been unable to establish universal rules even for themselves. We continue to argue about the boundaries between freedom and coercion, the fair distribution of scarce goods, acceptable levels of inequality, and the circumstances in which an emergency justifies the temporary restriction of rights. We have no universal agreement on what justice is, yet systems already exist that will have to allocate computing capacity, access to data, energy, bandwidth, and the authority to modify the shared state of infrastructure. The formalization of safe behavior has therefore become more urgent than ever. It is not enough to require an agent merely to do no harm. Such a requirement must be translated into measurable parameters, verifiable conditions, and procedures for resolving conflicts. The document below proceeds from precisely this engineering perspective.
然而，机器人定律仍然只是道德宣言，而非工程规范。究竟什么才构成“伤害”？“人类”或“全人类”这些术语代表了谁的利益？当一个主体的命令与另一个主体的安全发生冲突时，智能体该如何行动？谁来判定威胁是否真实，以及反应是否适度？在一个不断改变自身状态、在信息不完整的情况下运行，且行动速度可能快到人类无法理解的系统中，如何验证其对规则的遵守情况？悖论在于，人类甚至还没能为自己建立起普遍适用的规则。我们仍在争论自由与强制的界限、稀缺资源的公平分配、可接受的不平等程度，以及在何种紧急情况下可以暂时限制权利。我们对什么是“正义”尚未达成共识，但现有的系统已经必须负责分配计算能力、数据访问权限、能源、带宽以及修改基础设施共享状态的权限。因此，安全行为的规范化变得比以往任何时候都更加紧迫。仅仅要求智能体“不作恶”是不够的。这种要求必须转化为可衡量的参数、可验证的条件以及解决冲突的程序。以下文档正是基于这种工程视角而制定的。
1. Purpose of the Document
This document defines the formal conditions under which the cooperative behavior of agents in a multi-agent system (MAS) may be considered sustainable and safe. The document is intended for: developers of multi-agent systems; designers of distributed-access protocols; auditors of collective decision-making systems; any parties assessing the legitimacy of one agent’s demands on another. The model is agnostic to the nature of the agent. The only capabilities required of an agent are the ability to control resources and to make decisions regarding their use.
1. 文档目的
本文件定义了多智能体系统（MAS）中智能体的协作行为被视为可持续且安全的形式化条件。本文件旨在供以下人员参考：多智能体系统开发者；分布式访问协议设计者；集体决策系统审计员；以及任何评估一个智能体对另一个智能体提出的要求是否合法的相关方。该模型与智能体的具体性质无关。对智能体唯一的要求是其具备控制资源并就资源使用做出决策的能力。
2. Basic Definitions
2.1 Resource
A resource $$R_j$$ is any quantity that can be used by an agent to achieve a goal. It is described by the vector $$(X_j, Y_j, Z_j)$$:
| Parameter | Name | 0 | 1 |
|---|---|---|---|
| $$X_j$$ | Rivalrousness | Consumption by one agent completely excludes others | Consumption by one agent does not reduce availability to others |
| $$Y_j$$ | Urgency of need | Required once or rarely | Required continuously |
| $$Z_j$$ | Durability | Loses its value immediately | Retains its value indefinitely |
2. 基本定义
2.1 资源
资源 $$R_j$$ 是指智能体为实现目标所能使用的任何量。它由向量 $$(X_j, Y_j, Z_j)$$ 描述：
| 参数 | 名称 | 0 | 1 |
|---|---|---|---|
| $$X_j$$ | 竞争性 | 一个智能体的消耗完全排斥其他智能体 | 一个智能体的消耗不减少其他智能体的可用性 |
| $$Y_j$$ | 需求紧迫性 | 仅需一次或极少需要 | 持续需要 |
| $$Z_j$$ | 耐久性 | 立即失去价值 | 永久保留价值 |
2.2 Agent
An agent $$A_i$$ is any entity capable of: controlling resources, including controlling access to them; making decisions regarding their use.
2.2 智能体
智能体 $$A_i$$ 是指任何具备以下能力的实体：控制资源（包括控制对资源的访问）；就资源的使用做出决策。
2.3 Agent Power ($C_i$)
An agent’s power $$C_i$$ is a quantitative measure of the control exercised by agent $$A_i$$ over system resources. $$ C_i = \sum_j amount_{ij} \times (1 - X_j) \times S_j \times I_j $$ where:
- $$amount_{ij}$$ — the quantity of resource $$R_j$$ controlled by agent $$A_i$$;
- $$X_j$$ — the rivalrousness of resource $$R_j$$;
- $$(1 - X_j)$$ — the privacy factor: fully shareable resources $$(X_j = 1)$$ do not create power over others;
- $$S_j$$ — the scarcity of resource $$R_j$$ (Appendix A.1);
- $$I_j$$ — the indispensability of resource $$R_j$$ (Appendix A.2).
Engineering interpretation: $$C_i$$ indicates how critical the refusal of agent $$A_i$$ to provide access to its resources is to the functioning of the system.
2.3 智能体权力 ($C_i$)
智能体权力 $$C_i$$ 是对智能体 $$A_i$$ 所控制的系统资源影响力的定量度量。 $$ C_i = \sum_j amount_{ij} \times (1 - X_j) \times S_j \times I_j $$ 其中：
- $$amount_{ij}$$ — 智能体 $$A_i$$ 所控制的资源 $$R_j$$ 的数量；
- $$X_j$$ — 资源 $$R_j$$ 的竞争性；
- $$(1 - X_j)$$ — 隐私因子：完全可共享的资源 $$(X_j = 1)$$ 不会产生对他人的权力；
- $$S_j$$ — 资源 $$R_j$$ 的稀缺性（附录 A.1）；
- $$I_j$$ — 资源 $$R_j$$ 的不可或缺性（附录 A.2）。
工程解释：$$C_i$$ 指示了智能体 $$A_i$$ 拒绝提供其资源访问权限对系统运行的严重程度。
2.4 Access Restriction
Access restriction is an operation in which an agent artificially holds the parameter $$X_j$$ of a resource below its physical potential. This requires the expenditure of barrier resources $$R_{\text{barrier}}$$, for example on information distortion, more complex identification procedures, or denial of service.
2.4 访问限制
访问限制是一种操作，即智能体人为地将资源的参数 $$X_j$$ 保持在其物理潜能之下。这需要消耗屏障资源 $$R_{\text{barrier}}$$，例如用于信息扭曲、更复杂的识别程序或拒绝服务攻击。
2.5 Monopolization — the $$X \to 0$$ Strategy
Monopolization, or the $$X \to 0$$ strategy, is behavior aimed at restricting access to a resource and extracting an advantage from that restriction. A high level of agent power, by itself, is not evidence that the agent is using the $$X \to 0$$ strategy. The defining feature of this strategy is the agent’s intention to restrict other agents’ access to resources.
2.5 垄断 — $$X \to 0$$ 策略
垄断，或称 $$X \to 0$$ 策略，是指旨在限制对资源的访问并从这种限制中获取优势的行为。高水平的智能体权力本身并不证明该智能体正在使用 $$X \to 0$$ 策略。该策略的决定性特征是智能体限制其他智能体访问资源的意图。