信息-物理-社会系统类研究的死结是什么？

What is the Deadlock in Research on Cyber-Physical-Social Systems？

首先简单介绍一下什么叫做信息-物理-社会系统（Cyber-Physical-Social Systems，简称CPSS）。这在学术圈也算是一个比较时髦的词，它实际上精炼地概括了未来很长一段时间人类社会相关地复杂系统都将处于的一个混合状态。这个系统由三个层次组成：1）信息层，主要涉及到在虚拟空间中的感知、交互、计算、决策；2）物理层，主要涉及到在物理世界中的、基于物理规则的系统演化或执行；3）社会层，主要关注这个复杂系统中人的行为以及互动。一个直观的例子便是由网联自动驾驶车辆参与的交通系统：通信与自动化决策是信息层，车辆与基础设施是物理层，而系统中的出行者（以及更广泛外围的参与者）则是社会层。基本可以说，当下以及未来很长一段时间的复杂系统都包含了这三个层次，因此对于CPSS的研究可以说是未来大量研究方向的一个大图景层次的归宿。

First, let's start with a simple introduction to what is called a Cyber-Physical-Social System (CPSS). This is a rather trendy term in academic circles, and it succinctly encapsulates the hybrid state that complex systems related to human society will exist in for a long time to come. This system is composed of three layers:

1. The Cyber Layer, which primarily involves perception, interaction, computation, and decision-making in the virtual space.

2. The Physical Layer, which primarily involves the evolution or execution of systems in the physical world, based on physical laws.

3. The Social Layer, which focuses on human behavior and interactions within this complex system.

An intuitive example is a transportation system involving connected autonomous vehicles: communication and automated decision-making are the cyber layer, the vehicles and infrastructure are the physical layer, and the travelers (as well as the broader peripheral participants) within the system constitute the social layer. It can be said that complex systems, both now and for the foreseeable future, will all encompass these three layers. Therefore, research on CPSS can be seen as the ultimate, big-picture destination for numerous future research directions.

在这个宏观图景之中，最困难的无疑是社会层。其根本原因在于：人的行为是最复杂、最难准确刻画的，而信息层与物理层的描述都有足够精确的理论工具。不过，描述社会层的理论工具也不是没有，其中绝大部分都来自于经济学。经过上百年的发展，经济学已经发展出了一整套数理模型来刻画人的行为选择，而其刻画的原则大体上万变不离其宗：人有某种内置的偏好结构，这种偏好可以被整理为一个效用函数；而人在系统中的行为，则在于通过某种方式来最大化这个效用函数。（这里，为了刻画人可能存在的非理性因素，我用了“某种方式”这个定语；最后的结果不一定是对效用函数的精确最优化。）在数据充足、又有强力人工智能模型辅助的当下，人类已经有能力以前所未有的精度刻画人的群体行为；即使是面对可能还没有数据的新场景，通过大量数据训练过的大模型也有可能产生准确的迁移能力，使得我们可以对新场景下的人类行为有足够准确的表述。

Within this macroscopic picture, the most difficult layer is undoubtedly the social one. The fundamental reason is that human behavior is the most complex and the most difficult to characterize accurately, whereas the cyber and physical layers have sufficiently precise theoretical tools for their description. However, it's not that theoretical tools for describing the social layer don't exist; the vast majority of them come from economics. After centuries of development, economics has established a comprehensive set of mathematical models to characterize human behavioral choices, and the underlying principle of this characterization has, by and large, remained constant: people have a certain built-in preference structure, which can be formulated as a utility function; and an individual's behavior within the system consists of maximizing this utility function in some way. (Here, to account for potential irrational factors in humans, I use the qualifier "in some way"; the final outcome is not necessarily a precise optimization of the utility function.) 

In the current era of abundant data, aided by powerful artificial intelligence models, humanity is now capable of modeling collective human behavior with unprecedented accuracy. Even when facing new scenarios where data may not yet exist, large models trained on massive datasets can potentially exhibit accurate transfer capabilities, allowing us to formulate sufficiently accurate representations of human behavior in these new contexts. 

因此，CPSS类的研究并没有什么技术上的死结，无非就是如何通过更少的计算资源、更巧妙的建模方法来得到更准确的模型。这类在未来必然会走向主流的研究，真正的困难在更深刻的层次。这一困难恰恰就在于上述的“万变不离其宗”的效用函数方法、以及在其基础之上所衍生出来的一系列后续方法。

Therefore, CPSS-related research does not face any insurmountable technical hurdles; it is merely a matter of how to obtain more accurate models through fewer computational resources and more ingenious modeling methods. For this type of research, which is bound to become mainstream in the future, the real difficulty lies at a more profound level. This difficulty lies precisely in the aforementioned utility function method—which has remained constant in its core principles—and the series of subsequent methods derived from it. 

说到这里，就不得不提到经济学发展史上被讨论了无数次的“理性人假设”。经济学为了简化研究，往往会假设研究中的人都是所谓的经济理性人 -- 就是刚才描述的那种一切都是为了最大化一个效用函数的人。理性人假设后来在经济学圈子内部被不断攻击，因为显而易见的是，人并不是完全精确的计算机器，其行为很多时候并不会去绝对性地求解一个函数优化问题。为了弥补这一问题，后续的经济学发展捣鼓出了很多补丁，其中比较著名的就是所谓的“有限理性假设”，即通过给理性人加上某种更混乱地修正项来捕捉人的复杂行为。信息的不完全性、人的计算能力的有限性、乃至制度经济学中的人可以有限地改变自己地目标函数中的各项权重（依然是根据模型环境作出确定的反应），亦是对整个系统的另一些补丁。这些补丁毫无疑问逐步增强了经济学模型解释并预测现实的能力。

At this point, we must mention the "rational agent hypothesis," which has been discussed countless times in the history of economics. To simplify their research, economists often assume that the people in their studies are so-called "economic rational agents"—precisely the kind of individuals described earlier whose every action is aimed at maximizing a utility function. The rational agent hypothesis later came under constant attack within the field of economics because, quite obviously, humans are not perfectly precise calculating machines, and their behavior often does not involve absolutely solving a function optimization problem.

To compensate for this problem, subsequent developments in economics devised many "patches." A notable one is the so-called "bounded rationality hypothesis," which captures complex human behavior by adding some form of more disordered correction term to the rational agent. Incomplete information, limited computational capacity of humans, or the limited capability of humans to change the weights of their objective functions in institutional economics (which is still a deterministic reaction to the model environment), also serve as other patches to the overall system. Undoubtedly, these patches have progressively enhanced the ability of economic models to explain and predict reality.

但本文要讨论的并不是上面这些被说了很多年的技术性问题。我们要讨论的是一个更少被触及的、同时也是更深刻的问题：人类应该是“理性人”吗？

However, what this article intends to discuss are not these technical problems mentioned above, which have been talked about for many years. We want to discuss a question that is more rarely touched upon, and at the same time, more profound: Should human beings be "rational agents"? 

从经济学建模的角度，我们当然可以说：这只是一种对现实的描述，研究者总能够用足够准确的方式来刻画人的行为。然而，从另一个角度来看：所有经济学模型、以及用类似方法去建立的社会模型，其中的“人”从根本上都是“限定性玩家”。这些人被他们的“上帝” -- 即建模者 -- 所导入的效用函数所严格约束，他们所有的行为都必须遵循由这个效用函数所定义的目标。当然，如果建模者愿意，可以在模型中给他们提供近乎无穷多种选择，从而让这些人看起来有近乎无限的自由。但是，这些人却从来不具备一种可能是人类所拥有的最高层次的能力：定义自己的效用函数的能力。用更“去经济学”的话来说，模型里的这些人从来都无法决定自己是否要参与这一套游戏、自己用怎样的心态去参与这一套游戏、以及自己是不是要去创造一个新游戏。

From the perspective of economic modeling, we can certainly say: this is merely a description of reality, and researchers can always characterize human behavior in a sufficiently accurate way. However, from another perspective, in all economic models—and in social models built using similar methods—the "human" is fundamentally a "constrained player." These individuals are strictly bound by the utility function instilled in them by their "God"—that is, the modeler. All their behaviors must follow the objectives defined by this utility function.

Of course, if the modeler so desires, they can provide these individuals with a nearly infinite variety of choices within the model, making them appear to have almost limitless freedom. However, these individuals never possess what is possibly the highest-level ability that humans have: the ability to define their own utility function. To put it in less "economic" terms, the individuals within the model can never decide for themselves whether to participate in this game, what mindset to adopt while playing it, or whether to create a new game entirely.

这个问题肯定不是多数经济学家、甚至是广义的社会学家们喜欢的问题。他们所进行的任何研究，最终都需要得到一些前人未能发现或重视的结论；而为了得到这些结论，就必然要对其笔下的人的自由意志有严格的限制。在历史上，为了学科的发展，进行这样的妥协也是完全有必要的。然而，从现在开始往后，整个游戏规则已经在悄然发生变化：我们已经并不满足于对已有的系统进行刻画，我们更需要借助人工智能的强大能力，对未来的CPS系统进行设计。而一旦我们开始设计这个系统，根本的矛盾就出现了：如果我们为了得到某种“最优”的设计而限制了CPS系统中的“人”的最高层次的自由意志，那么我们设计出来的任何系统都必将是一个经过精心装潢的监狱。在这个监狱中，人必须遵循某一些被定死的价值观念（被效用函数所限制）来进行选择，自主决定自己的“人生裁判”是不被允许的。

This is certainly not a question that most economists, or even social scientists in a broader sense, are fond of. Any research they conduct ultimately needs to yield conclusions that previous researchers had not discovered or valued; and to obtain these conclusions, it is necessary to place strict limitations on the free will of the individuals in their models. Historically, for the sake of the discipline's development, making such a compromise was entirely necessary.

However, from this point forward, the rules of the game have been quietly changing: we are no longer satisfied with merely characterizing existing systems; we need to leverage the powerful capabilities of artificial intelligence to design future CPS systems. And once we begin to design these systems, a fundamental contradiction emerges: if, in order to achieve some kind of "optimal" design, we restrict the highest level of free will of the "humans" within the CPS, then any system we create will inevitably be a beautifully decorated prison. In this prison, people must make choices by following certain fixed values (constrained by the utility function); being the ultimate arbiter of one's own life is not permitted.

从某种程度上，人类过去几十年的、各个学科都普遍采用的这一种建模范式，已经悄然地把我们的人类社会导向了这样的陷阱之中。

To some extent, this modeling paradigm, universally adopted across various disciplines over the past few decades, has quietly led our human society into such a trap. 

对我而言，这一根本性的问题，迫切需要在未来的CPS系统相关研究中逐渐引起重视。必须承认，在社会层的建模中引入真正的“自由意志个体”将会带来巨大的技术上的挑战；过去数十乃至上百年的整个方法论的根基都需要被推倒重来。对于如何刻画这样的自由意志个体、以及刻画自由意志个体所带来的一些涉及到决定论的根本性困境*，我目前也是没有很好的答案的。但或许正因为大多数人都缺乏很好的答案，这一问题可能才会是未来的CPS相关研究中最关键的、也是最值得投入精力的问题。

For me, this fundamental issue urgently needs to be given increasing attention in future research related to CPS systems. It must be admitted that introducing true "free-will agents" into the modeling of the social layer will present enormous technical challenges; the foundational methodology of the past several decades, or even centuries, would need to be overturned and rebuilt from scratch. As for how to characterize such free-will agents, and the fundamental dilemmas related to determinism that arise from doing so*, I do not currently have good answers myself. But perhaps it is precisely because few has good answers that this issue may be the most critical, and the most worthy of effort, in future CPS-related research. 

对于从事相关研究的学者来说，这也必将是一个痛苦的过程 -- 这无异于对自己之前的所学所用的扬弃。但我想，面对人工智能即将重塑大量的人类社会运作范式的可见未来，跳出已有的舒适圈去思考更根本的问题，才是应对这一巨变的最佳态度。

For scholars engaged in related fields, this will undoubtedly be a painful process—it is tantamount to abandoning and transcending what they have previously learned and used. But I believe that in the face of a foreseeable future where artificial intelligence is set to reshape numerous paradigms of human society's operation, stepping out of one's existing comfort zone to contemplate more fundamental questions is the best attitude for confronting this monumental change.

*注：决定论的根本困境指的是，如果建模系统中存在自由意志个体，那么系统最终的演化必然是不可预测的；而一旦系统不可预测，那么我们就无法通过一些明确的指标来评估系统的“优劣”。从这个意义上，如何研究自由意志个体参与的CPS系统，需要一种全新的底层哲学。

• Note: The fundamental dilemma of determinism refers to the fact that if a modeling system contains free-will agents, the system's ultimate evolution is necessarily unpredictable. And once a system is unpredictable, we can no longer evaluate its quality ("goodness or badness") using clear metrics. In this sense, studying CPS systems in which free-will agents participate requires an entirely new underlying philosophy.

扩展阅读（Extended reading）：
1. 制度经济学（Institutional economics）
2. 元博弈（Meta-games）