Transfer Learning for Iterative Learning Control Using Gaussian Processes

Iterative Learning Control (ILC) is a technique used to improve the transient tracking performance of a repetitively operated system, such as a batch process. However, the assumption of iteration invariance of the process conditions limits the extension of ILC to similar processes and often demands process modelling and controller design from scratch. Data-driven modelling strategies are highly effective but they suffer from large model development cost during training. This is undesirable since operating conditions of industrial batch processes can switch rapidly to meet the requirements of the market leaving little time and resources for model development. In this paper, we present a novel method of transferring optimally tuned ILC across processes using the modern ideas of transfer learning (TL). Specifically, we develop a Gaussian process model using ideas of forward adaptive transfer learning (FAT) to facilitate knowledge transfer between processes. A major advantage of the proposed method is that optimal ILC designed for current process can be transferred to a new process using only limited training data. The proposed method is demonstrated on the Cott-Batch Reactor, showing substantial promise for the proposed TL-based FAT-GP.