Emerging robust and data‐driven control methods for uncertain learning systems

Abstract

Learning systems represent a particularly important class of practical data-driven systems that adapt to their environment based on the environment's response to the system's action. In some settings traditional robust and adaptive control techniques have been verified to be effective when plants are large scale and/or have very complex dynamics (e.g., industrial processes, power grid networks, and transportation systems). However, often in these plants the problems of parameter mismatch and unmodeled dynamics are encountered, and thus the techniques of robust or adaptive control methods in the framework of modern control may no longer be sufficient. As an alternative to robust or adaptive methods, various learning paradigms have been established, for example, reinforcement learning, deep learning, artificial neural networks, and iterative learning, among others. These learning paradigms construct controllers or control signals directly with the data collected and stored when the system operates, typically without the need of identifying system models. However, even such “model-independent” control systems must make assumptions about the system dynamics. One such common assumption is that the system dynamics are linear. Another is that they are time-invariant (autonomous). When the dynamics do not meet these assumptions, traditional learning paradigms may also fail. Despite the success of learning-based methods, finding suitable control frameworks for learning systems when there is uncertainty in the assumptions related to the system dynamics is still an open problem. The aim of this special issue is to collect recent research results that address issues involved in data-driven control methods and related algorithms for learning-based systems that operate under uncertainty. The special issue consists of fourteen papers covering several key areas of activity in which progress has been achieved. Many of the papers also include results of simulation or experimental tests of the presented control methods. Next we give a brief description of each of the fourteen papers. We have organized the papers thematically into four categories: (1) iterative learning control (ILC), (2) neural networks and machine learning, (3) identification and control, and (4) applications. The first five papers describe research on the control method of ILC, which is a process developed especially for applications where the same operation needs to be repeated over a limited period of time. In the first of these papers, Ding and Li give results on an adaptive ILC algorithm for nonlinear continuous nonparameterized systems. This adaptive ILC algorithm, which can adjust the adaptive parameters in both the iteration domain and the time domain, is proposed to track different reference trajectories repetitively over a finite time interval, and also to provide a unified adaptive control strategy for nonlinear continuous nonparameterized systems that have asymmetric control gain matrices for trajectory tracking in different domains. In the next paper, Chen and Chu address the fact that in multiagent collaborative tracking tasks, the system often encounters constraints in practice that cause challenging difficulties in the case of handling large-scale systems. They propose a novel constrained ILC design and further develop a decentralized implementation of the resulting ILC algorithm using the alternating direction method of multipliers, allowing the design to scale up to handle large-scale and varying system dynamics. The next three papers consider the special problem of ILC when there are quantization and intersample behavior effects. In the third paper, Ohnishi et al. develop a framework for a state-tracking ILC that mitigates the oscillatory intersample behavior that is often encountered in output-tracking ILC. The paper addresses stability in the iteration domain, introducing the idea of a robustness filter, designed in the frequency domain. The last two ILC pa