Pipeline-Oriented Learning and Control Orchestration for Large-Scale Distributed Actuation Platforms

Abstract

The use of learning-enabled control is increasingly becoming a component of large-scale distributed actuation infrastructures, including industrial automation systems, smart infrastructure or cyber-physical environments. Nevertheless, current mono or end-to-end learning-control systems are limited in scalability, have high latency and low fault resistance when they are applied over distributed assets. In this paper, a pipeline oriented learning and control orchestration technique is proposed, which decomposes sensing, learning, control, and actuation into separately executable pipeline stages to be coordinated by a distributed orchestration layer. Asynchronous execution, dynamic location of the stages and multi-pipeline coordination can be achieved with the described architecture where learning and control components can scale and develop independently without compromising the real time actuation restrictions. A prototype implementation is tested with large-scale distributed workloads and should be compared with centralised and using a fixed pipeline baselines. The experimental results show the close affinity to scale, major cuts in end-to-end delay, and novel resource use and resilience when assuming variable demand and node outages. These results identify effectiveness of pipeline-based orchestration as a scalable and resilient architectural design model of learning-enabled distributed actuation systems.