Sparsely-Packetized Predictive Control by Orthogonal Matching Pursuit

We study packetized predictive control, known to be robust against packet dropouts in networked systems. To obtain sparse packets for rate-limited networks, we design control packets via an l0 optimization, which can be effectively solved by orthogonal matching pursuit. Our formulation ensures asymptotic stability of the control loop in the presence of bounded packet dropouts. 1. Packetized Predictive Control. Let us consider the following discretetime, LTI plant model with a scalar input: x(k + 1) = Ax(k) +Bu(k), k ∈ N0, x(0) = x0, (1.1) where x(k) ∈ R and u(k) ∈ R for k ∈ N0. We assume that (A,B) is reachable. We are interested in a networked control architecture where the controller communicates with the plant actuator through an erasure channel, see Fig. 1.1. This channel introduces bounded packet-dropouts. In packetized predictive control (PPC), as described, e.g., in [4, 2] at each time instant k, the controller uses the state x(k) of the plant (1.1) to calculate and send a control packet of the form u ( x(k) ) , [ u0 ( x(k) ) , u1 ( x(k) ) , . . . , uN−1 ( x(k) )]⊤ ∈ R (1.2) to the plant input node. To achieve robustness against packet dropouts, buffering is used. To be more specific, suppose that at time instant k the data packet u(x(k)) defined in (1.2) is successfully received at the plant input side. Then, this packet is stored in a buffer, overwriting its previous contents. If the next packet, u(x(k+1)), is dropped, then the plant input u(k+1) is set to u1(x(k)), the second element of u(x(k)). The subsequent elements of u(x(k)) are then successively used until some control packet u(x(k+ l)), l ≥ 2, is successfully received, i.e., no dropout occurs. 2. Sparse Control Packet Design for Asymptotic Stability. In the PPC method discussed above, control packets u(x(k)) are transmitted at each time k ∈ N0 through an erasure channel. It is often the case that the bandwidth of the channel is limited, and hence one has to compress control packets to a smaller data size. To achieve an efficient compression of packets, we adopt a technique developed in compressed sensing [1] that considers the sparsity of a signal (or a vector). A sparse vector contains many 0-valued elements and can be highly compressed by only coding ∗This research was supported in part under the MEXT (Japan) Grant-in-Aid for Young Scientists (B) No. 22760317, and also Australian Research Council’s Discovery Projects funding scheme (project number DP0988601). †Graduate School of Informatics, Kyoto University, Kyoto, 606-8501, Japan; e-mail: [email protected] ‡School of Electrical Engineering & Computer Science, The University of Newcastle, NSW 2308, Australia; e-mail:[email protected] §Department of Electronic Systems, Aalborg University, Denmark; e-mail: [email protected] Controller Buffer Plant x(k) u(x(k)) u(k) Fig. 1.1. Networked Control System with PPC its few nonzero components. Based on this notion, in [2] we presented a sparsityinducing l/l optimization for PPC which gives practical stability, i.e., existence of a bounded invariant set. In the present work, we embellish the approach of [2] by showing how to find sparse control packets u(x(k)) ensuring that the networked system with bounded packet dropouts is asymptotically stable. For this, we propose to design the control packets via the following sparsity-promoting optimization: u(x(k)) , argmin u∈RN ∥u∥0 subject to ∥xN |k∥P + N−1 ∑ i=0 ∥xi|k∥Q ≤ x(k)⊤Wx(k), (2.1) where ∥u∥0 is the number of the nonzero elements in u = [u0, u1, . . . , uN−1] and x0|k = x, xi+1|k = Axi|k +Bui, i = 0, 1, . . . , N − 1. (2.2) In (2.2), N is the horizon length, taken here equal to the buffer size. The matrix Q > 0 is a design parameter which allows one to shape the response of the plant state components; P > 0 and W > 0 are chosen such that the feedback system is asymptotically stable, as indicated in Theorem 2.1 below. To state our subsequent results, we introduce the matrices Φ, Υ and Q̄ by Φ ,  B 0 . . . 0 AB B . . . 0 .. .. . . . .. AN−1B AN−2B . . . B  , Υ ,  A A .. A  , Q̄ , blockdiag{Q, . . . , Q } {{ } N−1 , P}. Theorem 2.1 (Asymptotic Stability). Assume that the number of consecutive packet-dropouts is uniformly bounded by N . Suppose the matrices P , Q, and W are chosen by the following procedure: i. Choose Q > 0 arbitrarily. ii. Solve the following Riccati equation to obtain P > 0: P = A⊤PA−A⊤PB(B⊤PB)−1B⊤PA+Q. (2.3) iii. Compute the constants ρ ∈ [0, 1) and c > 0 via ρ , 1− λmin(QP), c , 1− ρ 1− ρ max i λmax { Φi PΦi(Φ ⊤Q̄Φ)−1 } , where Φi is the i-th column of matrix Φ. iv. Choose a matrix E such that 0 < E < (1− ρ)P/c. v. Set W := P −Q+ E. Then for any choice of the control vector from the feasible set of the optimization (2.1) (which includes the sparsest control packets u(x(k))), the networked control system is asymptotically stable, that is, limk→∞ x(k) = 0. 0 10 20 30 40 50 60 70 80 90 100 10 −30 10 −20 10 −10 10 0 2−norm of the state x(k)