Learning from similar systems and online data-driven LQR using iterative randomised data compression

원문 초록

The problem of data-driven recursive computation of receding horizon LQR control through a randomized combination of online/current and historical/recorded data is considered. It is assumed that large amounts of historical input-output data from a system, which is similar but not identical to the current system under consideration, is available. This (possibly large) data set is compressed through a novel randomized subspace algorithm to directly synthesize an initial solution of the standard LQR problem, which however is sub-optimal due to the inaccuracy of the historical model. The first instance of this input is used to actuate the current system and the corresponding instantaneous output is used to iteratively re-solve the LQR problem through a computationally inexpensive randomized rank-one update of the old compressed data. The first instance of the re-computed input is applied to the system at the next instant, output recorded and the entire procedure is repeated at each subsequent instant. As more current data becomes available, the algorithm learns automatically from the new data while simultaneously controlling the system in near optimal manner. The proposed algorithm is computationally inexpensive due to the initial and repeated compression of old and newly available data. Moreover, the simultaneous learning and control makes this algorithm particularly suited for adapting to unknown, poorly modeled and time-varying systems without any explicit exploration stage. Simulations demonstrate the effectiveness of the proposed algorithm vs popular exploration/exploitation approaches to LQR control.