Hierarchical Control of Semi-autonomous Teams Under Uncertainty

The rapid progress in embedded hardware and software makes plausible ever more ambitious distributed, multi-layer, multi-objective, adaptive control systems. However, adequate design methodologies and design support lag far behind. Consequently, today most of the cost in system development is spent on ad hoc, prohibitively expensive systems integration and validation techniques that rely almost exclusively on testing the entire system.
Systematic design of hierarchical architectures and design of controllers for individual agents at all levels of the hierarchy address this bottleneck. Our efforts are focused on building a solid analytical foundation based on hybrid systems, a practical set of software design tools that support the construction, integration, safety, and performance analysis, online adaptation and offline functional evolution of multi-unmanned air vehicle (UAV) hierarchical control systems.
The control of every large system is organized in a distributed hierarchy for deeper understanding facilitated by the hierarchical structure, reduction in complexity of communication and computation, modularity and adaptability, robustness, and scalability. So the question is not whether it is a good idea to control large systems this way. The interesting questions are: How do we describe such systems in ways that make meaningful distinctions among different hierarchical, distributed control organizations?; What approaches help to assess system performance?; and, What tools and techniques aid in the design of good control organizations?
To describe distributed, hierarchical systems in a formal language, its syntax must be able to express their essential aspects. A "distributed system" comprises several components or subsystems distinguished from each other by function, location, or just identity. Thus we may have components that function as sensors, actuators, controllers, vehicles, path planners, etc. This is functional differentiation. Or we may merely have a collection of functionally identical agents distinguished by name or location (identity).
Deterministic control strategies are vulnerable to attacks exploiting their regularity and predictability. For better robustness, we use randomized control strategies. To achieve the control objectives, the control strategies will compete with randomized strategies modeling probabilistic disturbances and faults. We give probabilistic performance bounds on UAV performance using viscosity solutions for the resulting games. We generalize results on reachability objectives in discrete multi-agent games to the liveness case and the hybrid case and their combination. For multi-modal and multi-agent systems, we have developed probabilistic estimates of safe behavior, and tools for the analysis of reliability and performance of distributed multi-agent systems operating in probabilistic and malicious environments. In general, this allows us to shift from worst-case behavior to mean behavior estimates of control algorithms.