Communication constraints and latency in Networked Control Systems João P. Hespanha Center for Control Engineering and Computation University of California Santa Barbara In collaboration with Antonio Ortega (USC) and Yonggang Xu (UCSB)
Outline 1. Feedback control over a digital channel with limited bit-rate: process decoder digital channel encoder 2. Decentralized cooperative control with limited message-rate and delays process process process node node node communication network
Control with finite bit-rate reconstructed output y q (t) control input u(t) disturbance d(t) process noise n(t) output y(t) decoder digital channel b k {0,1} encoder b 0 b 1 b 2 b 3 b 4 t 0 t 1 t 2 t 3 t 4 t average bit-rate Motivation: Control of systems with sensors and actuators far from each other, connected by a digital network. E.g., control of an autonomous flying vehicle using measurements from a camera on the ground.
Control with finite bit-rate reconstructed output y q (t) control input u(t) disturbance d(t) process noise n(t) output y(t) decoder digital channel b k {0,1} encoder b 0 b 1 b 2 b 3 b 4 t 0 t 1 t 2 t 3 t 4 t average bit-rate Questions: 1. What is the minimum bit-rate for which stabilization (boundedness) is possible? 2. How to divide the bits among the distinct components of the output? 3. How to choose quantization intervals?
Minimum bit-rate y q (t) u(t) no noise/disturbance y(t) decoder b k {0,1} encoder average bit-rate Theorem: Stabilization is not possible with average bit-rate smaller than continuous-time process λ i eigenvalues of A discrete-time process [Tatikonda & Mitter]
Proof outline decoder b 0 b 1 b 2 b 3 b 4 encoder t 0 t 1 t 2 t 3 t 4 t A stable x decays to zero without control minimum bit-rate is zero (just set u ú 0) A has asympt. stable component of x in goes inv. subspace to zero without control can reduce A to its unstable inv. subspace we will assume that all eigenvalues of A have real part 0
Proof outline y q (t) decoder b 0 b 1 b 2 b 3 b 4 encoder y(t) = x(t) t 0 t 1 t 2 t 3 t 4 t 0 set to which x(t 0 ) is known to belong after bit b 0 is received 1 set to which x(t 1 ) is known to belong after bit b 1 is received 1 x(t 1 ) 0 x(t 0 ) x(t 2 ) 2
Proof outline y q (t) decoder b 0 b 1 b 2 b 3 b 4 encoder y(t) t 0 t 1 t 2 t 3 t 4 t 0 set to which x(t 0 ) is known to belong after bit b 0 is received 1 set to which x(t 1 ) is known to belong after bit b 1 is received Case 1: For some k, k has a single element could take x to zero (even in finite time) after t k diameter of k Case 2: As k, ρ( k ) 0 could take x to zero as k Case 3: As k, ρ( k ) unbounded no matter what control we use, x(t k ) is unbounded
Proof outline y q (t) decoder encoder y(t) b 0 b 1 b 2 b 3 b 4 t 0 t 1 t 2 t 3 t 4 t 0 set to which x(t 0 ) is known to belong after bit b 0 is received 1 set to which x(t 1 ) is known to belong after bit b 1 is received Case 1: For some k, k has a single element could take x to zero (even in finite time) after t k diameter of k Case 2: As k, ρ( k ) 0 could take x to zero as k Case 3: As k, ρ( k ) unbounded no matter what control we use, x(t k ) is unbounded Main idea: bit rate r min µ( k ) unbounded ρ( k ) unbounded volume of k diameter of k Stabilization not possible
Proof outline b k b k+1 t k t k+1 t k before b k+1 is received, it is only known that u k volume of k+1 volume of k
Proof outline b k b k+1 t k t k+1 t k before b k+1 is received, it is only known that u k volume of k+1 Q: Which coding would make µ( k+1 ) as small as possible? A: Divide k+1 into two sets of equal volume & use bit b k+1 to locate x(t k+1 ) in one of them volume of k coding that minimizes volume
Proof outline b k b k+1 t k t k+1 t At best with = only for coding that minimizes volume iterating explodes if average bit-rate is smaller than
Minimum bit-rate y q (t) u(t) no noise/disturbance y(t) decoder b k {0,1} encoder average bit-rate Theorem: Stabilization is not possible with average bit-rate smaller than continuous-time process λ i eigenvalues of A discrete-time process [Tatikonda & Mitter]
y q (t) Minimum bit-rate u(t) no noise/disturbance y(t) decoder b k {0,1} encoder average bit-rate Theorem: Assume A is diagonalizable 1. It is possible to keep the state bounded with any average bit rate larger or equal to r min 2. It is possible to make the state converge to zero with any average bit rate strictly larger than r min previous bound is tight But minimum volume coding may not work because unbounded volume unbounded diameter bounded volume ; bounded diameter
y q (t) Encoding/decoding schemes d(t) n(t) u(t) y(t) decoder w k {1,,N} T s (fixed) sampling interval encoder bit-rate Assume: closed-loop is stable for transparent encoding/decoding, i.e., A + B K asymptotically stable Questions: 1. How to design the encoder/decoder pair to make the closed-loop stable? 2. How much larger than r min does the bit-rate need to be for stability?
y q (t) State-prediction coding d(t) n(t) u(t) y(t) decoder w k {1,,N} encoder Inspired by Differential Pulse Code Modulation (DPCM) 1. Encoder and decoder maintain consistent estimates of the state, based only on the quantized information sent to the decoder. 2. At sampling times, the difference between the measured state and its estimate (based on previously transmitted data) is quantized and transmitted digitally. (hopefully state estimation error has smaller dynamic range than state itself) 3. Upon transmission, the state-estimates are corrected using the quantized error.
y q (t) State-prediction coding d(t) n(t) u(t) y(t) decoder w k {1,,N} encoder 1. Encoder and decoder maintain consistent estimates of the state, based only on the quantized information sent to the decoder. process estimate
y q (t) State-prediction coding d(t) n(t) u(t) y(t) decoder w k {1,,N} encoder 1. Encoder and decoder maintain consistent estimates of the state, based only on the quantized information sent to the decoder. smaller dynamic range 2. At sampling times, the difference between the measured state and its estimate is quantized and transmitted digitally. Q y T s A2D
y q (t) State-prediction coding d(t) n(t) u(t) y(t) decoder w k {1,,N} encoder 1. Encoder and decoder maintain consistent estimates of the state, based only on the quantized information sent to the decoder. 2. At sampling times, the difference between the measured state and its estimate is quantized and transmitted digitally. 3. Upon transmission, the state-estimates are corrected using the quantized error: without quantization (Q = identity) and noise, we would have
Encoder/decoder pair y Q A2D D2A encoder decoder 1. Encoder and decoder maintain consistent estimates of the state, based only on the quantized information sent to the decoder. 2. At sampling times, the difference between the measured state and its estimate is quantized and transmitted digitally 3. Upon transmission, the state-estimates are corrected using the quantized error.
Fixed-step quantization For simplicity assume A diagonal with real eigenvalues quantize the vector by using a scalar quantizer on each of its components If A was not diagonal one would precede the componentwise quantization by a diagonalizing linear transformation (see [JH, Ortega, Vasudevan, 02] for case of A with complex eigenvalues) i K i levels (equidistributed) i saturation level for the ith component quantizer K i # of quantization levels used for the ith component of # of words needed bit-rate
Fixed-step quantization quantizer saturation level Theorem: The state of the closed-loop system will remain bounded provided that # of quant. levels bit allocation & η i, δ i constants that depend on upper bounds on the noise/disturbance λ i large e λi Ts large K i large many bits needed for ith eigenspace
Fixed-step quantization quantizer saturation level Theorem: The state of the closed-loop system will remain bounded provided that # of quant. levels bit allocation & η i, δ i constants that depend on upper bounds on the noise/disturbance λ i large e λi Ts large K i large many bits needed for ith eigenspace required bit-rate bit-rate # of symbols bit-rate approximately r min when i À η i +δ i (but course quantization leads to large x even without noise/disturbance)
Variable-step quantization For simplicity assume A diagonal with real eigenvalues quantize the vector by using a scalar quantizer on each of its components If A was not diagonal one would precede the componentwise quantization by a diagonalizing linear transformation (see paper for case of A with complex eigenvalues) i K i levels (equidistributed) i (k) saturation level for the ith component quantizer at sampling time t k K i # of quantization levels used for the ith component of # of words needed bit-rate
Variable-step quantization Theorem: The state of the closed-loop system will remain bounded provided that η i, δ i constants that depend on upper bounds on the noise/disturbance bit allocation λ i large e λi Ts large K i large many bits needed for ith eigenspace required bit-rate bit-rate minimum rate can be achieved!!!
Conclusions process decoder digital channel encoder There exists a minimum rate below which stabilization is not possible We proposed encoder/decoder pairs (inspired by DPCM) that can achieve rates arbitrarily close to the minimal Variable-step quantization allows one to achieve the minimum bit rate Performance/robustness vs. bit-rate tradeoffs are still poorly understood Need to investigate problem in stochastic setting (Will entropy-like coding work with lower rates?)
Distributed Control physical couplings process process process spatially distributed process to be controlled (e.g., autonomous vehicles) sensing actuation node node node couplings supported by a communication network minimize communication (stealth, bandwidth) study the effect of nonideal communication (delays, drops, blackouts)
Scenarios Rendezvous in minimum-time or using minimum-energy (in spite of disturbances) Group of autonomous agents cooperate in searching for a target (perhaps mobile search & pursuit)
Communication minimization node node node couplings supported by a communication network current focus previous case The every bit-counts paradigm Goal: Design each to minimize the number of bits/second that need to be exchanged between nodes (quantization, compression, ) Domain: Media with little capacity and low-overhead protocols (bit at-a-time) E.g., underwater acoustic comm. between a small number of nodes. The cost-per-message paradigm Goal: Design each to minimize the number of message exchanges between nodes (scheduling, estimation, ) Domain: Media shared by a large number of nodes with nontrivial media access control (MAC) protocol (packet at-a-time) E.g., 802.11 wireless comm. between a large number of nodes.
Prototype problem process process process node node node In this talk: decoupled linear processes (with stochastic disturbance d) coupled quadratic control objective notation: x + x(k+1)
Prototype problem process process process node node node In this talk: decoupled linear processes (with stochastic disturbance d) coupled quadratic control objective E.g., rendez-vous of two vehicles notation: x + x(k+1)
Prototype problem process process process node node node In this talk: decoupled linear processes (with stochastic disturbance d) Minimum-cost solution (centralized) Completely decentralized solution
Communication performance trade-off control cost optimal communication achievable cost/comm. pairs nonachievable cost/comm. pairs minimum-cost (centralized) communication minimum comm. needed for solvability (zero, when decentralized solution yields finite cost)
Centralized architecture ith local process ith local constant communication between local and external process(es) jth external process Closed-loop system for simplicity here we assume only two processes
Estimator-based distributed architecture ith local process [Yook & Tilbury, Montestruque & Antsaklis, Xu & JH] ith local continuous open-loop estimator with discrete updates from the network ith local estimator for jth external process jth external process Closed-loop system additive perturbation w.r.t centralized equations
Communication logic ith local estimator for jth external process fusion logic network sched. logic jth external process How to fuse data? With no noise & delay (for now ) x j (k) received from network at time k best-estimate based on data received When to send data? Scheduling logic action Options periodically a j (k) = {k divisible by T?}, feedback policy a j (k) = F (x j (k), ) optimal T N
Optimal Scheduling Logic ith local estimator for jth external process fusion logic network sched. logic d is Gaussian i.i.d. with zero mean and covariance Σ jth external process Goals: minimize the estimation error minimize cost-penalty w.r.t. centralized minimize the number of transitions minimize communication bandwidth average L-2 norm average transmission rate relative weight of two criteria (will lead to Pareto-optimal solution)
Dynamic Programming solution undiscounted average-cost problem Theorem dynamic programming (DP) operator 1. There exists J * R and bounded h * : R n R such that 2. J * is the optimal cost and is achieved by the (deterministic) static policy 3. h can be found by value iteration
Dynamic Programming solution Proof outline: 1. e(k) is Markov and its transition distribution satisfies an Ergodic property (requires a mild restriction on the set of admissible policies omitted here) 2. T is a span-contraction [Hernandez-Lerma 96] 3. Result follows using standard arguments based on Banach s Fixed-Point Theorem for semi-norms. Theorem dynamic programming (DP) operator 1. There exists J * R and bounded h * : R n R such that 2. J * is the optimal cost and is achieved by the (deterministic) static policy 3. h can be found by value iteration
Example (2-dim) local process e 2 10 optimal scheduling λ= 100 5 a(k) = 0 a(k) = 1 0-5 not ellipses! -10-10 -5 0 5 e 10 1
Example (2-dim) local process e 2 10 optimal scheduling λ= 100 5 0-5 λ= 10 a(k) = 0 a(k) = 1 large weight in comm. cost large error threshold only communicate when error is very large -10-10 -5 0 5 e 10 1
Communication with latency ith local estimator for jth external process fusion logic network sched. logic jth external process delay of τ sched. logic sends x j (k) fusion logic recs. x j (k) x j (k) is incorporated into estimate k k+τ k+τ+1 τ= 0 in previous case
Example (1-dim) 10 2 10 1 10 0 10-1 estimation error variance τ = 0 optimal comm. vs. estim. error trade-off τ = 1 τ = 2 τ = 3 communication rate.1.25.5 optimal scheduling with network latency same error variance requires more bandwidth estimation error, given all information enroute
Conclusions process process process node node node communication network We constructed communications logics that minimize communication (measured in messages sending rate) We considered networks with (fixed) latency Study the effect of packet losses (especially important in wireless networks) Nonlinear processes Coupled control/communication-logic design Papers available at http://www.ece.ucsb.edu/~hespanha