Designing Complex Engineering Systems using Multi-Agent Coordination Approaches

设计越来越多的复杂系统，如最先进的飞机、先进的电力系统、无人驾驶飞行器和自动驾驶汽车，需要许多专家的知识。但是，尽管每个专家可能能够有效地设计一个特定的子系统，这些“最优”子系统的组合通常不会导致一个“好的”系统级设计。这是因为这些子系统之间的交互对复杂系统的性能至关重要，并且随着系统变得越来越复杂，捕获交互变得越来越困难。这个项目的长期目标是通过导出设计师-子系统目标函数来解决这个设计-分解问题，当所有子系统设计师实现了他们的目标时，总体设计目标也就实现了。传统上，两种截然不同的方法被用于设计复杂的工程系统：(i)集中式设计方法，在这种方法中，所有潜在的系统状态和由设计决策产生的行为的影响必须精确建模。（ii）中介设计方法，提供外部强加的中间设计目标，以确定竞争设计之间的权衡。中介方法避免了集中式方法的建模困难，但代价是昂贵的（时间和金钱上的）中介机制。该项目的方法是这两种方法的混合，提供了一种方法，在这种方法中，可以在不需要详细的交互模型或任意设计目标的情况下协调决策。该项目将建立的一个关键见解是，复杂系统设计从根本上类似于多智能体协调问题，它位于博弈论、机器学习和控制理论的交叉点。在这两种情况下，较低级别（子系统或代理）的决策和该级别上的交互会导致全局行为（复杂系统或多代理系统）。在这个项目中，设计师-子系统对将与代理保持一致。相应的设计子系统目标将基于系统级目标衍生出来，同时隐式地考虑子系统交互。这些目标将取代传统的中间设计目标与原则目标，如果满足，确保所需的整体系统性能。该项目旨在表明，基于目标函数的方法，以前成功地应用于复杂的多智能体系统的操作，可以在一个修改的形式也可以作为复杂工程系统的设计方法。

The design of a growing class of complex systems, such as state-of-the-art aircraft, advanced power systems, unmanned aerial vehicles, and autonomous automobiles, requires the knowledge of many experts. But though each expert may be able to design a particular subsystem efficiently, the combination of those "optimal" subsystems does not generally lead to a "good" system-level design. This is because the interactions among these subsystems are critical to the performance of the complex system, and capturing the interactions becomes increasingly difficult as the systems become more and more complex. The long-term goal of this project is to address this design-decomposition problem by deriving designer-subsystem objective functions in such a way that, when all subsystem designers achieve their objectives, the overall design goals are also achieved. Traditionally, two broadly different approaches are used to design complex engineered systems: (i) a centralized design approach where the impacts of all potential system states and behaviors resulting from design decisions must be accurately modeled. (ii) a mediated design approach where externally imposed intermediate design targets are provided to determine the tradeoffs among competing designs. The mediated approach avoids the modeling difficulties of the centralized approach, but at the cost of expensive (in time and money) mediation mechanisms. The project's approach is a hybrid of the two approaches, providing a method in which decisions can be reconciled without the need for either detailed interaction models or arbitrary design targets. A key insight on which this project will build is that complex systems design is fundamentally similar to the multiagent coordination problem, that lies at the intersection of game theory, machine learning, and control theory. In both instances, the decisions at the lower level (subsystems or agents) and the interactions at that level, lead to global behavior (complex system or multiagent system.) In this project, designer-subsystem pairs will be aligned with agents. Corresponding designer-subsystem objectives will then be derived based on system-level objectives while implicitly accounting for subsystem interactions. These objectives will replace the traditional intermediate design targets with principled objectives that, if met, ensure desired overall system performance. This project will aim to show that an approach based on objective functions, previously successfully applied to the operation of complex multiagent systems, can in a modified form also serve as an approach for the design of complex engineered systems.