Iterative Learning Control Based on Random Variance Reduction Gradient Method

Abstract

AbstractTraditional iterative learning control (ILC) algorithms usually assume that full system information and operation data can be utilized. However, due to the uncertainty and complexity of actual systems, it is difficult to access full system information and operation data accurately and completely. In this chapter, a novel ILC scheme based on stochastic variance reduced gradient (SVRG) is proposed. This scheme is not only suitable for resolving the incomplete information problem, but also converges efficiently under both strongly convex and non-strongly convex control objectives. To demonstrate the advantages, this chapter studied two scenarios, i.e., random error data dropout and model-free data-driven approach, and proposed two SVRG-based ILC algorithms for these two scenarios, respectively. It is theoretically demonstrated and experimentally verified that the proposed SVRG-based ILC scheme converges faster than both the full gradient and stochastic gradient methods for the two involved scenarios.