Looking back at 30 years of research into holonic manufacturing systems, these explorations made a lasting scientific contribution to the overall architecture of intelligent manufacturing systems. Most notably, holonic architectures are defined in terms of their world-of-interest (Van Brussel et al., 1998). They do not have an information layer, a communication layer, etc. Instead, they have components that relate to real-world assets (e.g. machine tools) and activities (e.g. assembly). And, they mirror and track the structure of their world-of-interest, which allows them to scale and adapt accordingly.

This research has wandered around, at times learning from its mistakes, and progressively carved out an invariant structure while it translated and applied scientific insights from complex-adaptive systems theory (e.g. autocatalytic sets) and from bounded rationality (e.g. holons). This paper presents and discusses the outcome of these research efforts.

At the top level, the holonic structure distinguishes intelligent beings (or digital twins) from intelligent agents. These digital twins inherit the consistency from reality, which they mirror. They are intelligent beings when they reflect what exists in the world without imposing artificial limitations in this reality. Consequently, a conflict with a digital twin is a conflict with reality.

In contrast, intelligent agents typically transform NP-hard challenges into computations with low-polynomial complexity. Unavoidably, this involves arbitrariness (e.g. don’t care choices). Likewise, relying on case-specific properties, to ensure an outcome in polynomial time, usually renders the validity of an agent’s choices both short-lived and situation-dependent. Here, intelligent agents create conflicts by imposing limitations of their own making in their world-of-interest.

Real-world smart systems are aggregates comprising both intelligent beings and intelligent agents. They are performers. Inside these performers, digital twins may constitute the foundations, supporting walls, support beams and pillars because these intelligent beings are protected by their real-world counterpart. Further refining the top-level of this architecture, a holonic structure enables these digital twins to shadow their real-world counterpart whenever it changes, adapts and evolves.

In contrast, the artificial limitations, imposed by the intelligent agents, cannot be allowed to build up inertia, which would hamper the undoing of arbitrary or case-specific limitations. To this end, performers explicitly manage the rights over their assets. Revoking such rights from a limitation-imposing agent will free the assets. This will be at the cost of reduced services from the agent. When other service providers rely on this agent, their services may be affected as well; that’s how the inertia builds up and how harmful legacy is created. Thus, the services of digital twins are to be preferred over the services of an intelligent agent by developers of holonic manufacturing systems.

Finally, digital twins corresponding to the decision making in the world-of-interest (a non-physical asset) allow to mirror the world-of-interest in a predictive mode (in addition to track and trace). It allows to generate short-term forecasts while preserving the benefits of intelligent beings. These twins are the intentions of the decision-making intelligent agents. Evidently, when intentions change, the forecasts needs to be regenerated (i.e. tracking the corresponding reality by the twin). This advanced feature can be deployed in a number of configurations (cf. annex).

回顾30年的整体制造系统研究，这些探索为智能制造系统的整体架构做出了持久的科学贡献。最值得注意的是，完整的体系结构是根据它们的关注世界定义的（Van Brussel等，1998）。它们没有信息层，通信层等。相反，它们具有与实际资产（例如机床）和活动（例如装配体）相关的组件。并且，他们镜像并跟踪他们感兴趣的世界的结构，这使他们能够相应地扩展和适应。

这项研究四处流浪，有时会从错误中吸取教训，并逐步雕刻出不变的结构，同时翻译并应用了来自复杂自适应系统理论（例如自催化集合）和有限理性（例如holons）的科学见解。本文介绍并讨论了这些研究成果。

在顶层，整体结构将智能人（或数字孪生）与智能代理区分开。这些数字双胞胎继承了他们所反映的现实的一致性。当他们反映世界上存在的东西而没有在这种现实中强加人为的限制时，他们就是聪明的人。因此，与数字双胞胎的冲突就是与现实的冲突。

相反，智能代理通常将NP难题转换为具有低多项式复杂度的计算。不可避免地，这涉及到任意性（例如，无关紧要的选择）。同样，依靠特定于案例的属性来确保多项式时间内的结果，通常可以使代理选择的有效性既短暂又取决于情况。在这里，智能代理通过在自己感兴趣的世界中施加自己的制造限制来制造冲突。

现实世界中的智能系统是由智能人和智能代理组成的集合体。他们是表演者。在这些表演者内部，数字孪生可能会构成基础，支撑墙，支撑梁和支柱，因为这些聪明的生物受到其真实世界中同行的保护。整体结构进一步完善了该体系结构的顶层，从而使这些数字孪生能够在其变化，适应和发展的任何时候对现实世界中的对应对象进行遮蔽。

相反，不能允许由智能代理施加的人为限制来建立惯性，这将阻碍任意或特定于情况的限制的撤销。为此，表演者明确管理其资产的权利。从施加限制的代理商处撤销此类权利将释放资产。这将以减少代理的服务为代价。当其他服务提供商依赖此代理时，其服务也可能会受到影响；这就是惯性积累的方式以及有害遗产的产生方式。因此，全息制造系统的开发者将优先选择数字孪生的服务，而不是智能代理的服务。

最终，与感兴趣世界（一种非物质资产）中的决策相对应的数字双胞胎可以以一种预测模式（除了跟踪和追踪）来镜像感兴趣世界。它可以生成短期预测，同时保留智能生物的利益。这些双胞胎是决策智能代理的意图。显然，当意图改变时，需要重新生成预测（即，由双胞胎跟踪相应的现实）。可以以多种配置部署此高级功能（请参阅附件）。