An AI-based, decentralized approach to process-control and planning with massively parallel systems.

摘要

Process-control systems form an integral and important part of the present automated society. They are generally deployed for tasks that are hazardous, laborious, repetitive, and need quick response. Due to the nature of the tasks, it is very important that these systems be highly reliable and efficient. Accommodating all of these requirements makes such systems very large and complex. Hence, decentralization and distribution are the norm rather than an exception. Measures such as decentralization and decomposition into smaller subtasks imposes, a hierarchy on the system, which requires that robust, higher level tasks act as supervisors/coordinators. In spite of these measures, there are problems with such decomposed systems. In this dissertation we present a general model of process-control systems that will address the problems with the current approaches.; We replace the coordinating process by a massively parallel cellular automaton. The process-control system is reduced from its physical-world manifestation to a configuration-space and finally to a processor-state space. This transforms the process-control problem to that of path planning of a point object in the presence of obstacles. We develop a massively parallel, decentralized solution to the path planning problem. We present an interface between the massively parallel planning system and the physical process-control system that is responsible for transforming the physical-world information into processor-state variables and the generated plan back to lower-level instructions for the controllers. The massively parallel solution has several advantages such as simplicity, inherent fault-tolerance, and reusability.; We go on to discuss methods of fault-tolerance based on elimination of critical processors, and generalized trajectories that minimize synchronization. Further, we put forward a symbolic approximation of the system with regard to 2-terminal reliability employing partitioning envelopes. Some other issues considered are reduction of non-convex obstacles to convex obstacles using exterior visibility theorems and Weak Voronoi diagrams.; Robots can be considered a special class of process-control systems. We propose a Hybrid Architecture based on decentralized path planning. We argue that there are certain tasks where a completely decentralized, autonomous, multi-robot system may not provide a good solution. In such cases, the massively parallel decentralized planning system as a centralized planning entity, may still provide opportunity for centralized robotic systems.