Model-Free Learning for Massive MIMO Systems: Stochastic Approximation Adjoint Iterative Learning Control

摘要

Learning can substantially increase the performance of control systems that perform repeating tasks. The aim of this letter is to develop an efficient iterative learning control algorithm for MIMO systems with a large number of inputs and outputs that does not require model knowledge. The gradient of the control criterion is obtained through dedicated experiments on the system. Using a judiciously selected randomization technique, an unbiased estimate of the gradient is obtained from a single dedicated experiment, resulting in fast convergence of a Robbins-Monro type stochastic gradient descent algorithm. Analysis shows that the approach is superior to earlier deterministic approaches and to related SPSA-type algorithms. The approach is illustrated on a multivariable example.