Kernel Reinforcement Learning for sampling-efficient risk management of large-scale engineering systems

Mainstream methods for maintenance scheduling of multi-state systems (e.g. aircraft engines) often encounter challenges such as uncertainty accumulation, the need for extensive training data, and instability in the training process, particularly in life-cycle cost management. This paper introduces an innovative Kernel Reinforcement Learning (KRL) approach designed to enhance the reliability and safety of multi-state systems while significantly increasing decision-making efficiency. The policy and value functions are formulated non- parametrically to capture high-value episodes and datasets. KRL integrates probabilistic setups to imbue reinforcement learning with uncertainty, enhancing exploration of state-action spaces. Prior knowledge can be seamlessly integrated with the probabilistic framework to accelerate convergence. To address the memory issues associated with kernel methods when handling large datasets, the kernel matrix is dynamically updated with screened high-value datasets. Numerical evaluations on a k-out-of-n system, a coal mining transportation system, and an aircraft engine simulation demonstrate that the proposed KRL approach achieves faster convergence and reduced life-cycle costs compared to alternative methods. Specifically, KRL reduces the number of training episodes by 2-3 orders of magnitude, with a maximum cost reduction of 92%.