Predictive Cost

Adaptive Control
A NUMERICAL INVESTIGATION OF PERSISTENCY,
CONSISTENCY, AND EXIGENCY

ZARMEEN AZMAT

TAM W. NGUYEN, SYED ASEEM UL ISLAM,

DENNIS S. BERNSTEIN, and ILYA V. KOLMANOVSKY

A
mong the multitude of modern control methods, (LQR) and linear quadratic Gaussian (LQG), and nonlinear
model predictive control (MPC) is one of the most control methods, such as feedback linearization and sliding
successful [1]–[4]. As noted in “Summary,” this mode control. Although MPC is computationally intensive,
success is largely due to the ability of MPC to it is more broadly applicable than Hamilton–Jacobi–Bellman-
respect constraints on controls and enforce con- based control and more suitable for feedback control than the
straints on outputs, both of which are difficult to handle with minimum principle. In many cases, the constrained optimi-
linear control methods, such as linear quadratic regulator zation problem for receding-horizon optimization is convex,
which facilitates computational efficiency [5].
Digital Object Identifier 10.1109/MCS.2021.3107647 MPC uses a model of the plant with constrained reced-
Date of current version: 12 November 2021 ing-horizon optimization to compute a sequence of

64 IEEE CONTROL SYSTEMS » DECEMBER 2021 1066-033X/21©2021IEEE

future control inputs. The first element of the optimized
sequence is implemented, and the remaining elements are Summary

M
discarded. At a first glance, this strategy seems ad hoc and odel predictive control, which is based on receding-
wasteful. However, receding-horizon optimization is horizon optimization, is a widely used modern control
known to be asymptotically stabilizing for a sufficiently technique with numerous successful application in diverse
long prediction horizon or with terminal constraints, areas. Much of this success is due to the ability of reced-
and, if the initial state is feasible and the model matches ing-horizon optimization to enforce state and control con-
the plant, it is recursively feasible with respect to con- straints, which are crucial in many applications of control.
straints [6]–[9]. To avoid the need for an observer, predictive cost adaptive
MPC handles convex control and output constraints ef- control (PCAC) uses the block observable canonical form,
ficiently compared to conventional feedback control meth- whose state consists of past values of the control inputs and
ods that treat control constraints as input nonlinearities. In measured outputs. PCAC also uses recursive least squares
particular, linear dynamics with control magnitude and (RLS) with variable-rate forgetting for online identification.
move-size (rate) saturation comprise a Hammerstein system The article describes the algorithmic details of PCAC
[10]–[12], whose importance is reflected by the vast litera- and numerically investigates its performance through a col-
ture on bounded control and antiwindup strategies [13]– lection of numerical examples that highlight various control
[15]. Within the context of MPC, however, magnitude and challenges, such as model-order uncertainty, sensor noise,
move-size control constraints are treated as optimization prediction horizon, stabilization, magnitude and move-size
constraints rather than explicit nonlinearities, which simpli- saturation, and stabilization. The numerical examples are
fies the treatment of this problem. used to probe the performance of PCAC in terms of persis-
From the perspective of classical control and modern tency, consistency, and exigency. Since, unlike dual con-
optimal state-space control, MPC is a paradigm shift. For trol, PCAC does not employ a separate control perturbation
example, conventional feedback methods use integral ac- to enhance persistency, the focus is on self-generated per-
tion to asymptotically follow step commands and reject sistency during transient operation. For closed-loop iden-
step disturbances, and it is common practice to embed inte- tification using RLS, sensor noise gives rise to bias in the
grators within LQR and LQG controllers. MPC, however, identified model, and the goal is to determine the effect of
performs numerical optimization at each step, and thus no the lack of consistency. Finally, the numerical examples are
integrator per se appears [8]. Without integral action, MPC used to assess exigency, which is the extent to which the
lacks one of the key strengths of classical control, namely, online identification emphasizes model characteristics that
the ability to follow setpoint commands and reject step dis- are most relevant to meeting performance objectives.
turbances without knowledge of the dc gain of the plant.
On the other hand, by eschewing integral action, MPC is
not degraded by integrator windup. This dependence is mitigated by robust MPC techniques [31],
Since MPC uses model-based prediction to determine [32], minimax techniques for minimizing the loss of a worst-case
future control inputs, it follows that these predictions require scenario, and robust tube-based MPC [33]–[35].
knowledge of future commands and disturbances. Future com- As an alternative to robust MPC, extensions of MPC to in-
mands are sometimes known; this is the aim of preview control clude online identification and learning are considered in [36]–
and trajectory-tracking methods [16], [17]. On the other hand, [39]. These techniques can potentially overcome the worst-case
except for disturbance-feedforward control architectures [18], considerations of robust MPC by allowing MPC to learn the
disturbances are rarely known, especially in the future. The true plant dynamics and disturbances. For system identifica-
lack of future knowledge of commands and disturbances thus tion, the commands, disturbances, and control inputs must
represents a potential obstacle for MPC. provide sufficient persistency to facilitate identification. Concur-
Although MPC is often based on linear models, this tech- rent learning with MPC can be viewed as a form of indirect
nique is also applicable to nonlinear systems [19]. An addition- adaptive control [40]. Within this context, the role of persistency
al challenge is the case of output feedback, where not all of the is a longstanding issue [41]. More generally, indirect adaptive
plant states are measured. In this case, state estimation can be control can be viewed as a specialized version of data-driven
used to provide estimates of unmeasured states, where the control [42]–[48].
state estimates serve as ersatz states for the receding-horizon Beyond persistency, since online identification and learning
optimization [20]–[24]. Transient errors in the state estimates, occur during closed-loop operation, the control input is corre-
however, can impede the ability to enforce constraints on un- lated with the measurements due to disturbances and sensor
measured states. Output-feedback MPC that does not rely on a noise. When recursive least squares (RLS) is used for closed-
state estimator is developed in [25]–[30] and [8, Ch. 5]. loop identification, this correlation may obstruct consistency and,
In addition to the lack of knowledge about future commands thus, lead to asymptotic bias in the parameter estimates [49]–
and disturbances, a potential weakness of MPC is the need for a [51]. Alternative identification methods, such as the prediction
sufficiently accurate model for effective predictive optimization. error method (PEM) [52] and instrumental variables (IV) [53],

DECEMBER 2021 « IEEE CONTROL SYSTEMS 65

provide consistency despite signal correlation. Using numerical The present article describes predictive cost adaptive con-
examples to examine asymptotic bias, RLS, PEM, and IV are ap- trol (PCAC), which incorporates the two features described
plied to open-loop identification, closed-loop identification us- previously. No attempt is made in this article to derive stability
ing a fixed-gain linear controller, and MPC control, where the or performance guarantees for this approach. Instead, the goal
identified plant model is updated at each step. is to systematically investigate PCAC through a collection of
Since MPC is an inherently discrete-time control technique, numerical examples. These examples are chosen to highlight
its application to plants with continuous-time dynamics entails the stability and performance of PCAC for a diverse collection
a sampled-data plant with analog-to-digital sampling and dig- of plants and scenarios. These scenarios include the effect of
ital-to-analog input reconstruction. As in any sampled-data model order; sensor noise; unknown step, harmonic, and
controller implementation, the sampling rate must be chosen to broadband disturbances; stabilization; sampled-data effects;
minimize aliasing and folding effects [54], [55]. Within this con- magnitude and move-size control constraints; output con-
text, MPC provides direct digital control of discretized plants straints; and abrupt and gradual changes in the plant.
without the need to discretize continuous-time controllers. The The numerical examples in this article provide a venue for
present article considers examples with sampled-data dynam- investigating the interplay between identification and control, a
ics to assess the intersample response [56]–[58]. longstanding problem in control theory [64]–[66]. This inter-
The present article focuses on MPC for constrained linear play is addressed by dual control, where the objective is to de-
plants with two nonstandard features. First, only a limited num- termine probing signals that enhance persistency and, thus, the
ber of states of the plant are assumed to be measured; this is the speed and accuracy of the concurrent identification [67]–[69]. In
case of output feedback rather than full-state feedback. Instead of contrast to dual control, PCAC does not require a separate
using an observer to provide estimates of unmeasured states, the probing input, such as the dither signal used by extremum-
present article takes advantage of the block observable canonical seeking control to estimate gradients [70]. Instead, the present
form (BOCF) [59], which is a state-space realization whose state article investigates the phenomenon of self-generated persistency,
is a function of measured inputs and outputs. This realization where PCAC automatically increases the persistency of the
thus removes the need to build an observer to estimate unmea- control signal in response to the closed-loop performance.
sured states. Within this framework, a key assumption is the Within the context of PCAC, the interplay between identifi-
availability of measurements of all constrained outputs. cation and control is embodied by exigency, which refers to the
The second feature of the present article is the use of online ability of the identification to prioritize features of the identified
identification to construct and refine a model for the constrained model that impact closed-loop performance. A key goal of the
receding-horizon optimization. Concurrent identification is numerical examples is, thus, to explore manifestations of per-
performed with RLS [60], [61] with the additional benefit of vari- sistency, consistency, and exigency within PCAC. This article
able-rate forgetting (VRF) [62], [63]. extends the preliminary investigation of PCAC given in [71]
and [72] through refinements of the control algorithm; more
extensive numerical examples; and an investigation of persis-
TABLE 1 The key takeaways from each example. MIMO:
multiple-input, multiple-output; VRF: variable-rate tency, consistency, and exigency. Table 1 lists the key takeaways
forgetting. from each numerical example.
The next section describes the control architecture within
Example Key Takeaways which this article is framed. Next, elements of PCAC are de-
1 • Robustness to model order scribed, which include input–output models, the BOCF for an
• Self-generated persistency input–output predictive model, and quadratic programming
• Exigency
(QP). Linear single-input, single-output (SISO) Examples 1–5
2 • Stabilization
are then presented, followed by linear multiple-input, multiple-
• Self-generated persistency
output (MIMO) Examples 6 and 7, which are followed by linear
3 • Stabilization of unstably stabilizable plants
time-varying Examples 8–10. Conclusions and directions for
4 • Harmonic and broadband disturbance rejection
future research are then presented.
• Exigency
5 • Output constraints
• Role of relaxation weight
NOTATION
The n # n identity matrix is denoted by I n, the m # n matrix of
6 • Unmatched broadband disturbance rejection
for MIMO plants ones by 1 m # n, the m # n matrix of zeros by 0 m # n, and the Kro-
7 • Stabilization and output constraints for MIMO necker product by 7 . The infinity norm of a vector
9
plants x = [x 1 f x n]T ! R n is defined by ||x||3 = max (|x 1|, f,|x n|) .
8 • Importance of VRF for abruptly changing systems
9 • Tuning VRF in the presence of sensor noise CONTROL ARCHITECTURE AND OBJECTIVES
10 • Enforcement of constraints for abruptly changing The examples in this article consider the control architecture
systems shown in Figure 1. In some cases, the continuous-time plant G
is specified as a transfer function or state-space model. In these

66 IEEE CONTROL SYSTEMS » DECEMBER 2021

cases, G is discretized according to the sample-and-hold oper- ing these signals, PCAC produces the discrete-time control
ations, and, to capture intersample behavior, the Matlab func- u k ! R m at each step k.
tion ode45 is used to simulate the continuous-time dynamics
between sampling times. In other cases, the dynamics of the PREDICTIVE COST ADAPTIVE CONTROL
plant are specified as a discrete-time transfer function or in-
put–output model. Input–Output Models
In Figure 1, let a representation of G be given by For online system identification and prediction, consider
MIMO input–output models of the form
xo (t) = Ax (t) + Bu (t) + D 1 w (t), (1)
nt nt
y (t) = Cx (t) + Du (t), (2) yt k = - / Ft i yt k - i + / G
t i u k - i, (8)
i=1 i=0

where x (t) ! R lx is the state, w (t) ! R lw is the disturbance, where k $ 0 is the time step, nt $ 1 is the data window for esti-
y (t) ! R p is the output of G, A ! R lx # lx, B ! R lx # m, C ! R p # lx, mation, u k ! R m is the control, yt k ! R p is the model output,
D ! R p # m, and D 1 ! R lx # lw. The disturbance w(t) is matched if t 0, f, G
and Ft 1, f, Ft nt ! R p # p and G t nt ! R p # m are the coefficient
there exists Ur ! R m # m such that D 1 = BUr ; otherwise, the dis- matrices to be estimated. Based on the structure of (8), the
turbance is unmatched. Note that, if B = D 1, then the distur- 1-step predicted output is given by
bance is matched, and Ur = I. The disturbance w(t) is assumed nt nt
to be piecewise constant within each interval kTs to (k + 1) Ts, y 1|k = - / Ft i, k + 1 y k + 1 - i + G
t 0, k + 1 u 1|k + / G
t i, k + 1 u k + 1 - i, (9)
i=1 i =1
at the value w k, where Ts 2 0. In the case where w(t) is stochas-
tic, the standard deviation of w k is specified. The output y(t) of where u 1|k ! R m is the 1-step computed input, and Ft 1, k + 1, f,
G is corrupted by sensor noise v(t). The sample operation yields t 0, k + 1, f, G
Ft nt , k + 1 ! R p # p and G t nt , k + 1 ! R p # m are the estimated
9 9
y k = y (kTs) + v k, where v k = v (kTs) ! R p is the sampled sensor coefficient matrices at step k. As explained below, the coefficient
noise, and Ts is the sampling time. For each example, the statis- t 0, k + 1, f, G
matrices Ft 1, k + 1, f, Ft nt , k + 1, G t nt , k + 1 are available at step
tics of the sampled sensor noise v k are specified. k. Similarly, using the 1-step predicted output y 1|k and the
The tracking output y t,k ! R p t is defined by 2-step computed control u 2|k, the 2-step predicted output is
9
given by
y t,k = C t y k, (3)
nt
where C t ! R p t # p. The performance objective is to have y t,k fol- y 2|k = - Ft 1, k + 1 y 1|k - / Ft i, k + 1 y k + 2 - i + G
t 0, k + 1 u 2|k
i=2
low a commanded trajectory rk ! R p t, whose future values may nt 
or may not be known, and, thus, the command preview may or t 1, k + 1 u 1|k + / G
+G t i, k + 1 u k + 2 - i . (10)
i=2
may not be available.
In addition to the performance objective, the constrained Hence, for all j $ 1, the j-step predicted output is given by
output y c,k ! R p c is defined by j -1 nt j -1

9 y j|k = - / Ft i, k + 1 y j - i|k - / Ft i, k + 1 y k + j - i + / Gt i, k + 1 u j - i|k

y c,k = C c y k, (4) i=1 i= j i=0
nt

where C c ! R p c # p. The objective is to enforce the inequali- +/G
t i, k + 1 u k + j - i, (11)
i= j
ty constraint
Cy c,k + D # 0 n c # 1, (5) where the first term is zero for j = 1, and the second and fourth
terms are zero for all j 2 nt .
where C ! R n c # p c, and D ! R n c. Note that (5), where “#” is in- PCAC combines online identification with receding-hori-
terpreted componentwise, defines a convex set. zon optimization, particularly:
In physical systems, the control is constrained in both mag- »» Identification: RLS is used to estimate the coefficients of
nitude and rate. The magnitude control constraint has the form the model (8).
u min # u k # u max, (6)
v (t )
where u min ! R m is the vector of the minimum control magni- w (t )
rk Ts
tudes, and u max ! R m is the vector of the maximum control y(t ) yk
yt,k uk u (t ) G
magnitudes. In addition, the move-size control constraint has Ct PCAC ZOH
yc,k
the form Cc
Tu min # u k - u k - 1 # Tu max, (7) PCAC: Predictivefollowing
Cost Adaptive Control
FIGURE 1 Command and disturbance rejection under
ZOH: Zero-Order Hold
sampled-data adaptive control. The objective is to follow com-
where Tu min ! R m is the vector of minimum control move siz-
mands r k to the tracking output y t,k = C t y k while enforcing con-
es, and Tu max ! R m is the vector of maximum control move straints on the constrained output y c,k = C c y k, control u k, and
sizes. As shown in Figure 1, the inputs to PCAC are the com- control move-size u k - u k - 1 . All sample-and-hold operations are
mand rk, tracking output y t,k, and constrained output y c,k . Us- synchronous.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 67

 9
»» Receding-horizon optimization: Using the identified model where, for all k $ 0, t k = P kj = 0 m -j 1 ! R, m k ! (0, 1] is the forget-
t t
(8), QP is used to perform receding-horizon optimization ting factor, P0 ! R [np (m + p) + mp] # [np (m + p) + mp] is positive definite,
tnp (m + p) + mp
to determine the next control input. and i 0 ! R is the initial estimate of the coefficient vec-
tor. The performance variable z k (it) ! R p is defined by
Recursive Least Squares With Variable-Rate nt nt
z k (it) = y k + / Ft i y k - i - / G
9 t i u k - i, (13)
Forgetting for Online Identification
i =1 i=0
For online identification, RLS is used to estimate the coefficients
where the vector it ! R
t (m + p) + mp
np
of the input–output model (8). To do this, RLS minimizes the of coefficients to be estimated
cumulative cost is defined by

it = vec 6F t nt @ . (14)
9 t0 f G
t 1 f Ft nt G
k
z i (i) z i (it) + 1 (it - i 0)T P 0-1 (it - i 0), (12)
ti T t
J k (it) = / t
i=0
tk tk Defining the regressor matrix z k ! R p # [np (m + p) + mp] by

Open- and Closed-Loop Identification Using the Prediction

Error Method and Instrumental Variables

T his sidebar compares the accuracy of the prediction error shown. IV is implemented using the Matlab function iv4 with a
method (PEM) and instrumental variables (IV) with recur- second-order identification model.
sive least squares (RLS) in open- and closed-loop frameworks.
All of the examples in this sidebar consider the asymptotically Example S1: Open-Loop Identification Using Recursive
stable nonminimum-phase (NMP) plant Least Squares and the Prediction Error Method
Let u k, w k, and v k be zero-mean, Gaussian white noise
4 (z - 1.2) sequences with a standard deviation of one. Figure S1
G (z) = , (S1)
z 2 - z + 0.8 shows the error of the estimate of each coefficient of (S1)
versus the time step. Note that all of the PEM coefficient
with matched disturbance w k . The output y k of (S1) is cor- errors approach zero, which indicates consistency. How-
rupted by sensor noise v k, and the noisy output is used for ever, the RLS estimates are biased, which shows that RLS
identification. The input u k of (S1) is assumed to be known and lacks consistency. 
is also used for identification.
For RLS, let nt = 2 and P0 = 1000. PEM is implemented Example S2: Open-Loop Identification Using Recursive
using the Matlab function recursiveBJ with initial parameter Least Squares and Instrumental Variables
covariance 1000, a second-order identification model, and a Let u k, w k, and v k be as in Example S1. Figure S2 shows the
zeroth-order noise model. Numerical tests with higher-order error of the estimate of each coefficient of (S1) versus the time
noise models yielded less accurate results and, thus, are not step. Note that all of the IV coefficient errors approach zero,

100 100 100 100

10–2 10–2
|θ1,k – θ1,k|

|θ2,k – θ2,k|

|θ1,k – θ1,k|

|θ2,k – θ2,k|

10–2 10–2
10–4 10–4
" " " "

10–6 10–6 10–4

10–8 10–8 10–4
10–10 10–10 10–6
100 100 100 100
|θ3,k – θ3,k|

|θ4,k – θ4,k|

|θ3,k – θ3,k|

|θ4,k – θ4,k|

10–2 10–2
" " " 10–2 "
10–4 10–4 10–2
10–6 10–6 10–4
RLS RLS
10–8 10–8 PEM 10–6 10–4 IV
10–10 0 10–10 0
10 102 104 106 10 102 104 106 100 101 102 103104 105 100 101 102 103104 105
k (Step) k (Step) k (Step) k (Step)

FIGURE S1 Example S1: Open-loop identification accuracy of FIGURE S2 Example S2: Open-loop identification accuracy of
the prediction error method (PEM) and recursive least squares recursive least squares (RLS) and instrumental variables (IV).
(RLS). Note that PEM coefficient errors approach zero. Note that the IV coefficient errors approach zero.

68 IEEE CONTROL SYSTEMS » DECEMBER 2021

 z k = 6-y k - 1 f -y k - nt u k f u k - nt @ 7 I p, (15)
9 T T T T
Recursive Least Squares Online Identification
it follows that the performance variable (13) can be rewritten as L k = m -k 1 Pk, (18)
z k (it) = y k - z k it. (16) P k + 1 = L k - L k z Tk (I p + z k L k z kT)-1 z k L k , (19)
T
Note that, with (16), the cost function (12) is strictly convex and i k + 1 = i k + Pk + 1 z k ( y k - z k i k). (20)
quadratic and, thus, has a unique global minimizer. The unique
global minimizer Note that i k + 1 computed using (20) is available at step k.
t 0, k + 1, f, G
Thus, Ft 1, k + 1, f, Ft nt , k + 1, G t nt , k + 1 are available at step k.
i k + 1 = argmin it J k (it) = vec 6F t nt , k + 1@
9 t 0, k + 1 f G
t 1, k + 1 f Ft nt , k + 1 G The step-dependent parameter m k is the forgetting factor. In
(17) t
! R np (m + p) + mp the case where m k is constant, RLS uses constant-rate forgetting
is computed by RLS as: (CRF); otherwise, RLS uses VRF [62]. For VRF, m k is given by

100
100 104 104
|θ1,k – θ1,k|

|θ2,k – θ2,k|

|θ1,k – θ1,k| 102

|θ2,k – θ2,k|
"
10–1
" 102
10–1 " "
100 100
10–2 10–2 10–2
10–2 10–4
10–4
10–6
104 104
|θ3,k – θ3,k|

|θ4,k – θ4,k|

|θ3,k – θ3,k|

|θ4,k – θ4,k|
" " 100 " 102 " 102
100
100 100
10–1 10–2 10–2
RLS IV
10–1 PEM 10–4 10–4 RLS
10–2
100 102 104 106 100 102 104 106 100 101 102 103 104 105 100 101 102 103 104 105
k (Step) k (Step) k (Step) k (Step)

FIGURE S3 Example S3: Closed-loop identification accuracy of

recursive least squares (RLS) and the prediction error method FIGURE S4 Example S4: Closed-loop identification accuracy of
(PEM). Note that the PEM coefficient errors are uniformly recursive least squares (RLS) and instrumental variables (IV).
smaller than the RLS coefficient errors. Note that IV produces large estimation errors.

which indicates consistency. However, the RLS estimates are Example S4: Closed-Loop Identification Using Recursive
biased, which shows that RLS lacks consistency.  Least Squares and Instrumental Variables
Let u k, w k, and v k be as in Example S3. Figure S4 shows the
Example S3: Closed-Loop Identification error of the estimate of each coefficient of (S1) versus the time
Using Recursive Least Squares and the step. Note that the RLS estimates are biased, and IV produces
Prediction Error Method large estimation errors. 
Let w k and v k be zero-mean, Gaussian white noise sequenc- These numerical examples, although limited, demonstrate
es with a standard deviation of 0.1. For closed-loop identifi- that PEM is more accurate than RLS for both open-loop and
cation examples, the input u k is given by a linear quadratic closed-loop identification. Furthermore, although IV showed
Gaussian controller designed using the Matlab command lqg clear advantages over RLS for open-loop identification, its per-
with weights Q xu = Q wv = I 3. Figure S3 shows the error of the formance for closed-loop identification is poor, at least for the
estimate of each coefficient of (S1) versus the time step. Note instruments implemented using the Matlab command iv4. Note
that, after convergence, the PEM coefficient errors are uni- that these closed-loop identification examples are based on a
formly smaller than the RLS coefficient errors; the improve- linear time-invariant controller. The performance of PEM within
ment ranges from 0.71 to 9.05. However, all of the PEM and the context of adaptive control, where the controller is linear time
RLS estimates are biased, which shows that neither method varying, is considered in “Closed-Loop Identification Under Pre-
is consistent for this example.  dictive Cost Adaptive Control With the Prediction Error Method.”

DECEMBER 2021 « IEEE CONTROL SYSTEMS 69

 1
1 + hg (z k - x d, f, z k) 1 6g (z k - x d, f, z k)@
mk = , (21) tor of (22) approximates the short-term average standard
deviation of z k . Consequently, the case g(z k - x d, f, z k) 2 0 im-
where 1 : R " " 0, 1 , is the unit step function, where 1 (x) = 0 plies that the short-term average standard deviation of z k is
for all x 1 0 and 1 (x) = 1 for x $ 0, and greater than the long-term average standard deviation of z k .
k
The function g (z k - x d, f, z k) used in VRF suspends forgetting
1 / z Ti z i when the short-term average standard deviation of z k drops
9 x n i = k - xn
g(z k - x d, f, z k) = k
- 1. (22) below the long-term average standard deviation of z k . This
1 / T
z z technique prevents forgetting in RLS-based online identifica-
i i
x d i = k - xd
tion due to zero-mean sensor noise with a constant standard
In (21) and (22), h $ 0 and 0 1 x n 1 x d are numerator and de- deviation rather than due to the magnitude of the noise-free
9
nominator window lengths, respectively. Define g(0, f, 0) = 0. identification error.
If the sequence z k - x d, f, z k is zero-mean noise, then the nu-
merator and denominator of (22) approximate the average Receding-Horizon Optimization
standard deviation of the noise over the intervals [k - x n, k] Note that (11) can be applied recursively to predict the out-
and [k - x d, k], respectively. In particular, by choosing x d & x n, put over the prediction horizon. A simpler approach that
it follows that the denominator of (22) approximates the long- avoids recursion is to use the BOCF state-space realization of
term average standard deviation of z k, whereas the numera- (8) given by

Closed-Loop Identification Under Predictive Cost Adaptive

Control With the Prediction Error Method

T his sidebar investigates the performance of predictive cost output y k is corrupted by sensor noise v k, and the noisy output
adaptive control (PCAC) when the prediction error method is used for identification. The control input u k, which is given by
(PEM) is used instead of recursive least squares (RLS). Nu- receding-horizon optimization, is also used for identification.
merical examples are presented to compare the performance For PCAC/PEM, PEM-based identification is combined
of PCAC/PEM and PCAC/RLS. Note that, unlike the numerical with receding-horizon optimization. PEM is implemented using
examples considered in “Open- and Closed-Loop Identification the Matlab function recursiveBJ with a zeroth-order noise
Using the Prediction Error Method and Instrumental Variables,” model, model order nt , and initial parameter covariance P0.
where the controller is linear time invariant, the PCAC control- Furthermore, the identification model in PEM is initialized at
ler is adaptive. The examples in this sidebar involve discrete- i 0. Both PCAC/PEM and PCAC/RLS use u min = - 1, u max = 1,
time plants with matched disturbances w k . For each plant, the r = 50I , - 1, Pr = 50, R = 10I ,,
Tu min = - 1, Tu max = 1, , = 50, Q

5
100
|θ1,k – θ1,k|

|θ2,k – θ2,k|

" 100 "

0 10–2
yk

10–1
10–4
–5 PCAC/PEM 10–2
1 PCAC/RLS
100
0.5
|θ3,k – θ3,k|

|θ4,k – θ4,k|

" 100 "

0 10–2
uk

–0.5 10–1
PCAC/PEM
10–4 PCAC/RLS
–1
10–2
0 2,000 4,000 6,000 8,000 10,000 100 102 104 100 102 104
k (Step) k (Step) k (Step)

FIGURE S5 Example S5: Adaptive control of the asymptotically

stable plant (S1) using predictive cost adaptive control (PCAC)/ FIGURE S6 Example S5: Closed-loop identification accuracy of
prediction error method (PEM) and PCAC/recursive least the prediction error method (PEM) and recursive least squares
squares (RLS). PCAC/PEM and PCAC/RLS have 2.03- and (RLS) operating within predictive cost adaptive control (PCAC)/
1.70-dB suppression relative to the open-loop response, PEM and PCAC/RLS. Note that neither algorithm is uniformly
respectively. more accurate.

70 IEEE CONTROL SYSTEMS » DECEMBER 2021

 C = 6I p 0 p # p g 0 p # p@, D
9 t 9 9 t
x 1|k = A kxt k + Bt k u k, (23) tk= G 0, k + 1 . (28)
y k = C k xt k + Dt k u k, (24)
The model structure (8) and its realization (27), (28), which ex-
9
where x 1|k ! R np is the 1-step predicted state, xt k = [xt T1, k f xt Tnt , k]T !
t
tends the SISO construction given in [59], provides the basis for
tnp
R , and the output prediction described below. Note that (25) and (26)
depend on past values of the measurement y k and control u k
9 t 0, k + 1 u k, (25)
xt 1, k = y k - G rather than the 1-step predicted output y 1|k given by (9) and
1-step computed input u 1|k given in the next section.
nt - i + 1 nt - i + 1
9
xt i, k = - / Ft i + j - 1, k + 1 y k - j + / t i + j - 1, k + 1 u k - j, i = 2, f, nt ,
G Using (23), the 1-step predicted output given by (9) can be
j =1 j =1 rewritten as
(26)
R t V t k u 1|k . (29)
S-F1, k + 1 I p g g 0 p # pW Rt t 0, k + 1VW y 1|k = Cx 1|k + D
S h SG 1, k + 1 - Ft 1, k + 1 G
0p # p j h W S t t t W
9S
t k= W 9 G 2, k + 1 - F2, k + 1 G 0, k + 1
A S h h j j 0 p # pW, Bt k = S W, Similarly, defining the i-step predicted state
S h S h W
h j Ip W SG t 0, k + 1W
t nt , k + 1 - Ft nt , k + 1 G
SS t W
-Fnt , k + 1 0 p # p g g 0 p # pW T X
T X 9 t t
(27) x i|k = A k x i - 1|k + B k u i - 1|k, i $ 2, (30)

100
2000 100
10–1
|θ1,k – θ1,k|

|θ2,k – θ2,k|
" "
10–2
0 10–1
10–3
10–2 10–4 PCAC/PEM
–2000 PCAC/PEM 102 PCAC/RLS
PCAC/RLS 102
1
101
|θ3,k – θ3,k|

|θ4,k – θ4,k|

" 101 "

100
0
uk

100
10–1
–1 10–1 10–2
0 20 40 60 80 100 100 101 102 100 101 102
k (Step) k (Step) k (Step)

FIGURE S8 Example S6: Closed-loop identification accuracy of

FIGURE S7 Example S6: Adaptive control of the unstable plant the prediction error method (PEM) and recursive least squares
(S2) using predictive cost adaptive control (PCAC)/prediction (RLS) operating within predictive cost adaptive control (PCAC)/
error method (PEM) and PCAC/recursive least squares (RLS). PEM and PCAC/RLS. Note that neither algorithm is uniformly
Note that PCAC/PEM fails to stabilize the plant. more accurate.

t = 2, i 0 = 10 -2 $ 1 2nt # 1, P0 = 10 3 I 2nt , and a strictly proper

m = 1, n Example S6: An Unstable Plant
identification model. Let G(z) be given by

4 (z - 1.2)
Example S5: An Asymptotically Stable Plant G (z) = , (S2)
z 2 - z + 1.2
Let G(z) be given by (S1), and let w k and v k be as in Example
S3. Figure S5 shows y k and u k for PCAC/PEM and PCAC/RLS. which is unstable, and let w k and v k be as in Example S3.
Figure S6 shows the error of each coefficient of (S1) versus the Figure S7 shows y k and u k for PCAC/PEM and PCAC/RLS.
time step obtained with PCAC/PEM and PCAC/RLS. Note that Figure S8 shows the error of each coefficient of (S2) versus the
all of the PEM and RLS estimates are biased, which shows that, time step obtained with PEM and RLS operating within PCAC/
for this example, neither method is consistent. However, RLS is PEM and RLS in PCAC. Note that PCAC/PEM fails to stabilize
uniformly faster in reducing the identification errors.  the plant. 

DECEMBER 2021 « IEEE CONTROL SYSTEMS 71

 and Ct k, , ! R , p # np and Tt k, , ! R ,p # ,m are the observability and
t
the i-step predicted output is given by
block Toeplitz matrices
t k u i|k, i $ 2. (31)
y i|k = Cx i|k + D
R t V
S Dk 0 p#m g g g g 0 p#mW
Note that the i-step prediction at step k uses the current esti- SH t k, 1 t
Dk g g g g 0 p#mW
R C V S
mates At k, Bt k, and D
t k at each intermediate stage of the predic- S W t t t k g g g 0 p# mW
t S H k, 2 H k, 1 D W
t 9 S CA k W t 9
tion horizon. It follows that Ck, , = S t t
, Tk, , = S H k, 3 H k, 2 H k, 1 Dt t k g g 0 p#mW,
h W S t t t W
SS t , - 1WW S H k, 4 H k, 3 H k, 2 j j h W
Y1|k, , = Ct k, , x 1|k + Tt k, , U 1|k, ,, (32) TCA k X S h h h j j j 0 p#mW
where SS t t k, , - 2 H
t k, , - 3 g H t k, 2 H t k WW
t k, 1 D
H k, , -1 H
T X
y 1|k u 1|k (34)
Y1|k, , = > h H ! R , p, U 1|k, , = > h H ! R ,m, (33)
9 9

t k, i ! R p # m is defined by
where, for all i = 1, f, , - 1, H
y ,|k u ,|k

3
1.5 rk rk
2
1 1
yk

yk
0.5 0
0 –1
6
2 4
2
uk

uk

1
0
0 –2
0
log ∆IR log yk – rk

log ∆IR log yk – rk

0
–2
–2
–4
–4
0
0
–1
–0.5
–2
–1 –3
–1 0
log ∆DG

log ∆DG

–2 –2
–3
–4 –4
–5 –6
0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Im (z)

Im (z)

Plant Zero Plant Zero

0 Model Pole 0 Model Pole
Model Zero Model Zero
–0.1 –0.1
–1 –0.5 0 0.5 1 1.5 –1 –0.5 0 0.5 1 1.5
Re (z) Re (z)
(a) (b)

FIGURE 2 Example 1: Command following for the discrete-time asymptotically stable single-input, single-output plant (46) using nt = n = 2
t 0,k = 0. (a) The single-step command is rk / 1. The output y k approaches the step command
with a strictly proper model, that is, with G
with a decreasing command-following error, which is 5.6e-6 at k = 60. Note that, although the impulse-response error TIR stops
decreasing after the initial transient, the dc gain estimation error approaches zero, which indicates exigency within the closed-loop
identification; the identification of the dc gain is prioritized by recursive least squares to decrease the command-following error. The
bottom-most plot compares the poles and zero of the identified model at k = 60 to the poles and zero of the plant. (b) The multistep
command rk is given by (47). For each step command, y k approaches the command with a decreasing command-following error, which
is 1.6e-4 at k = 19, 1.2e-4 at k = 39, and 1.6e-4 at k = 60. Note that, although TIR stops decreasing after each transient, the dc gain
estimation error approaches zero. The bottom-most plot compares the poles and zero of the identified model at k = 60 to the poles and
zero of the plant. DDG: the absolute value of the difference between the dc gain of the plant and model; TIR: the L 2-norm difference
between the 30-step impulse response of the plant and model.

72 IEEE CONTROL SYSTEMS » DECEMBER 2021

 9
t k, i = t ik- 1 Bt k . (35)
H CA
Quadratic Programming-Based Receding-Horizon
Optimization
9
Let R k, , = [r Tk + 1 f r Tk + ,]T ! R ,p t be the vector of , future com-
9 9
mands; define C t, , = I , 7 C t ! R , p #, , let Yt, 1|k, , = C t, , Y1|k, , ,
t p min (Yt, 1|k, , - R k, ,)T Q (Yt, 1|k, , - R k, ,) + (TU 1|k, ,)T RTU 1|k, , (37)
U 1|k, ,

where Y1|k, , is given by (32), be the ,- step propagated tracking-

output vector; and define the sequence of differences of the subject to
computed control inputs by
C , Y1|k, , + D , # 0 ,n c, (38)
U min # U 1|k, , # U max, (39)
9
TU 1|k, , = [(u 1|k - u k)T(u 2|k - u 1|k)T f (u ,|k - u , - 1|k)T]T ! R ,m. (36)
TU min # TU 1|k, , # TU max , (40)

where Q = 80 p Q#r p 0 p Pr# p B ! R , p t # , p t is the positive definite output

9
To facilitate the implementation of linear inequality constraints t t
t t
9
over the horizon ,, define C , = I , 7 (CC c) ! R ,n c # , p and r ! R (, - 1) p t # (, - 1) p t is the positive definite cost-to-go
weight; Q
9
D , = 1 , # 1 7 D ! R ,n c . With this notation, QP-based receding- output weight; Pr ! R p t # p t is the positive definite terminal
horizon optimization is given by: output weight; R ! R ,m # ,m is the positive definite control

3
1.5 rk rk
2
1 1
yk

yk

0.5 0
0 –1
6
2 4
2
uk

uk

1
0
0 –2
0
log ∆IR log yk – rk

log ∆IR log yk – rk

0
–2
–4 –2
–6 –4
0
0
–1
–0.5
–2
–1 –3
0 0
log ∆DG

log ∆DG

–2 –2
–4 –4
–6 –6
0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Im (z)

Im (z)

Plant Zero Plant Zero

0 Model Pole 0 Model Pole
Model Zero Model Zero
–0.1 –0.1
–1 –0.5 0 0.5 1 1.5 –1 –0.5 0 0.5 1 1.5
Re (z) Re (z)
(a) (b)

FIGURE 3 Example 1: Command following for the discrete-time asymptotically stable single-input, single-output plant (46) using
nt = 3 2 n = 2 with a strictly proper model. (a) The single-step command is rk / 1. The output y k approaches the step command with a
decreasing command-following error, which is 4.7e-7 at k = 60. (b) The multistep command rk is given by (47). For each step command,
y k approaches the command with a decreasing command-following error, which is 2e-4 at k = 19, 1.3e-4 at k = 39, and 1.7e-4 at k = 60.
Note that, in (b), the estimates of the poles and zeros, including the pole-zero cancellation, are more accurate than in (a) due to the
persistency arising from the changing step command. TDG: the absolute value of the difference between the dc gain of the plant and
model; TIR: the L 2-norm difference between the 30-step impulse response of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 73

 9 9
move-size weight; U min = 1 , # 1 7 u min ! R ,m, U max = 1 , # 1 7
9 9 Quadratic Programming-Based Receding-Horizon
u max ! R ,m, DU min = 1 , # 1 7 Tu min ! R ,m , and DU max = 1 , # 1 7
Optimization With Output-Constraint Relaxation
Tu max ! R ,m. Since R is positive definite, QP-based reced-
ing-horizon optimization is a strictly convex optimiza- min (Yt, 1|k, , - R k, ,)T Q (Yt, 1|k, , - R k, ,)
U 1|k, ,, f
tion problem. 
+ (TU 1|k, ,)T RTU 1|k, , + f TSf (41)
It may occur in practice that the constraint (38) on the
predicted output cannot be satisfied for all values of the subject to
control input that satisfy (6) and (7). In this case, QP-based C , Y1|k, , + D , # f, (42)
receding-horizon optimization is infeasible. To overcome
U min # U 1|k, , # U max, (43)
this problem, a standard technique is to introduce a slack
variable f ! R ,n c to relax the constraint (38). With this TU min # TU 1|k, , # TU max, (44)
modification, the QP-based receding-horizon optimiza- 0 ,n c # 1 # f, (45)
tion becomes:

3
1.5 rk rk
2
1 1
yk

yk
0.5 0
0 –1
6
2 4
2
uk

uk

1
0
0 –2
0
log ∆IR log yk – rk

log ∆IR log yk – rk

0
–1
–1
–2
–2
–3

0.2 3
0 2
–0.2 1
–0.4 0
0 0
log ∆DG

log ∆DG

–1 –1
–2 –2
–3
0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Im (z)

Im (z)

Plant Zero Plant Zero

0 Model Pole 0 Model Pole
–0.1 –0.1
–1 –0.5 0 0.5 1 1.5 –1 –0.5 0 0.5 1 1.5
Re (z) Re (z)
(a) (b)

FIGURE 4 Example 1: Command following for the discrete-time asymptotically stable single-input, single-output plant (46) using
nt = 1 1 n = 2 with a strictly proper model. (a) The single-step command is rk / 1. The output y k approaches the step command with a
decreasing command-following error, which is 7.8e-4 at k = 60. (b) The multistep command rk is given by (47). For each step command,
y k approaches the command with a decreasing command-following error, which is 5.3e-3 at k = 19, 5.7e-3 at k = 39, and 1.8e-2 at
k = 60. Note that the accuracies of the impulse response and the poles and zeros of the identified model in both (a) and (b) are poor,
due to the fact that nt 1 n. Nevertheless, in both (a) and (b), the estimate of the dc gain is sufficiently accurate to allow predictive cost
adaptive control to approach the step command. TDG: the absolute value of the difference between the dc gain of the plant and model;
TIR: the L 2-norm difference between the 30-step impulse response of the plant and model.

74 IEEE CONTROL SYSTEMS » DECEMBER 2021

where S ! R ,n c # ,n c is the positive definite constraint relaxation examples, R k, , = 1 , # 1 7 rk, which implies that future com-
weight. Since R and S are positive definite, (41)–(45) is a strictly mands over the prediction horizon are assumed to be equal to
convex optimization problem. the current command.
The accelerated dual gradient-projection algorithm is used Discrete-time and sampled-data plants are considered in
to solve (37)–(40) and (41)–(45), where the previously comput- this section. For all sampled-data examples, in this and later
ed Lagrange multipliers provide a warm start for the next itera- sections, the analog-to-digital conversion is an instantaneous
tion. For real-time implementation, the control computed sampler, and the digital-to-analog conversion is a zero-order
between k and k + 1 is implemented at step k + 1. hold (ZOH) device. PCAC is run between sampling times, and
the continuous-time dynamics are integrated by ode45 to cap-
LINEAR SINGLE-INPUT, SINGLE-OUTPUT EXAMPLES ture intersample behavior. All commands are discrete-time
This section considers linear time-invariant SISO plants, where signals, all deterministic disturbances are discretized at the in-
y t, k = y c, k = y k . None of the examples in this article consider tegration step size, and all stochastic disturbances are modeled
command preview. In particular, for all command-following as constant between sampling times.

2 rk rk
2
1
yk

yk
1
0 0
4 dk dk
4
"

"
2 dk dk
uk

uk

2
0 0
0
log dk – dk log ∆IR log yk – rk

log dk – dk log ∆IR log yk – rk

0
–1 –1
–2 –2
–3 –3
0.2 0.2
0 0
–0.2 –0.2
–0.4 –0.4
–0.6 –0.6
" 0 " 0
–1 –1
–2 –2

0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Im (z)

Im (z)

Plant Zero Plant Zero

0 Model Pole 0 Model Pole
Model Zero Model Zero
–0.1 –0.1
–1 –0.5 0 0.5 1 1.5 –1 –0.5 0 0.5 1 1.5 2
Re (z) Re (z)
(a) (b)

FIGURE 5 Example 1: Command following and disturbance rejection for the discrete-time asymptotically stable single-input, single-output
plant (46) using rk / 1 and nt = n = 2 with a strictly proper model. (a) The single-step disturbance is d k / 0.8. The output y k approaches
the step command with a decreasing command-following error, which is 3e-3 at k = 60. Note that, although TIR stops decreasing after
the initial transient, recursive least squares (RLS) continues to refine the estimate dtk of the disturbance for the remainder of the simula-
tion. This refinement indicates exigency within the closed-loop identification, where the estimation of dtk is prioritized to decrease the
command-following error. The bottom-most plot shows a pole-zero cancellation in the identified model at k = 60. (b) The multistep distur-
bance d k is given by (52). For each step disturbance, y k approaches the command with a decreasing command-following error, which is
1.3e-2 at k = 19, 2.4e-2 at k = 39, and 8.8e-2 at k = 60. Note that, although TIR stops decreasing within a few steps after each change
in the disturbance, RLS continues to refine the estimate dtk of the disturbance while dk is constant. Note that, in both (a) and (b), the step
disturbances induce self-generated persistency, which facilitates the ability of RLS to effectively construct an estimate dtk of the distur-
bance d k, which, in turn, facilitates disturbance rejection. The bottom-most plot shows a pole-zero cancellation in the identified model at
k = 60, due to the step disturbance. TIR: the L 2-norm difference between the 30-step impulse response of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 75

 All of the subsequent examples involving PCAC are based of model order nt . Consider the discrete-time input–output
on RLS for closed-loop identification. “Open- and Closed-Loop SISO plant
Identification Using the Prediction Error Method and Instru-
mental Variables” compares the accuracy of PEM and IV with y k = -0.5y k - 1 + 0.1y k - 2 + u k - 1 - 0.4u k - 2 . (46)
RLS for open-loop identification as well as for closed-loop iden-
tification with a linear time-invariant controller. In addition, Let u min = -10, u max = 10, Tu min = -10, Tu max = 10, and v k = 0.
“Closed-Loop Identification Under Predictive Cost Adaptive No output constraint is considered in this example. The plant
Control With the Prediction Error Method” compares the per- (46) is initialized with y -1 = y -2 = 0 and u 0 = u -1 = u -2 = 0.
formance of PCAC using PEM and RLS for PCAC control, Note that the order of the plant is n = 2. At each time step,
where the identified plant model is updated at each step. PCAC uses (37)–(40) with , = 5, Q r = 2I , - 1, Pr = 5, and R = I , .
-2 3
Let m = 1, 0 i = 10 $ 1 2n # 1 and 0 = 10 I 2nt .
t , P
Example 1: A Discrete-Time Asymptotically Figures 2–4 show the response of PCAC for nt = 2, nt = 3,
Stable Minimum-Phase Plant and nt = 1, respectively, using a strictly proper model, that is,
The goal of this example is to investigate the ability of PCAC to using G t 0,k = 0. The absolute value of the command-following
follow step commands and reject step disturbances for an as- error, the L 2 - norm difference between the 30-step impulse re-
ymptotically stable minimum-phase plant for various choices sponse of the plant and model (TIR), the absolute value of the

2 rk 2 rk
yk

yk

1 1
0 0
4 dk 4 dk
"
"

2 dk 2 dk
uk

uk

0 0
0 0
log dk – dk log ∆IR log yk – rk

log dk – dk log ∆IR log yk – rk

–1 –1
–2
–2
–3
0 0
–0.5 –0.5
–1
–1
0
" " 0
–1 –1
–2 –2
–3 –3
0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Im (z)

Im (z)

Plant Zero Plant Zero

0 Model Pole 0 Model Pole
Model Zero Model Zero
–0.1 –0.1
–1 –0.5 0 0.5 1 1.5 2 –1 –0.5 0 0.5 1 1.5 2
Re (z) Re (z)
(a) (b)

FIGURE 6 Example 1: Command following and disturbance rejection for the discrete-time asymptotically stable single-input, single-out-
put plant (46) using rk / 1 and nt = 3 2 n = 2 with a strictly proper model. (a) The single-step disturbance is d k / 0.8. The output y k
approaches the step command with a decreasing command-following error, which is 1.3e-3 at k = 60. (b) The multistep disturbance d k
is given by (52). For each step disturbance, y k approaches the command with a decreasing command-following error, which is 7.4e-3
at k = 19, 2e-2 at k = 39, and 7.1e-2 at k = 60. Note that, although the accuracies of the poles and zeros of the identified model in both
(a) and (b) are poor, the estimate of the disturbance is sufficiently accurate to allow predictive cost adaptive control to approach the
command and reject the disturbance. TIR: the L 2-norm difference between the 30-step impulse response of the plant and model.

76 IEEE CONTROL SYSTEMS » DECEMBER 2021

difference between the dc gain of the plant and model (TDG), n n
y k = - / Fi y k - i + / G i (u k - i + d k - i), (48)
and the pole-zero location of the plant and model identified at i =1 i=1
k = 60 are computed as diagnostics. Figures 2(a), 3(a), and 4(a)
show the response of PCAC for the constant command rk / 1, where n = 2, F1 = 0.5, F 2 = -0.1, G 1 = 1, and G 2 = -0.4. In the
and Figures 2(b), 3(b), and 4(b) show the response of PCAC for case of a constant matched disturbance dt k ! R , the plant out-
the three-step command put is given by
n n n
1, 0 # k 1 20, y k = - / Fi y k - i + / G i u k - i + / G i dt k, (49)
rk = * -1, 20 # k 1 40, (47)
i =1 i =1 i =1

3, k $ 40. and for a given model order nt , the output of the model identi-
fied by RLS at step k is given by
Figures 2–4 show that, for asymptotically stable minimum- nt nt
phase plants, PCAC follows step commands and is robust to y k = - / Ft i, k y k - i + / G
t i, k u k - i . (50)
i =1 i =1
the choice of model order nt .
Next, a matched disturbance d k ! R is applied to the plant Using (50) in (49), the disturbance dt k estimated by RLS is
(46), for which the plant output is given by given by

1.5 rk 2 rk
1
1
yk

yk

0.5
0 0
4 dk 4 dk
"

"

2 dk 2 dk
uk

uk

0 0
log dk – dk log ∆IR log yk – rk

log dk – dk log ∆IR log yk – rk

0 0
–0.5 –0.5
–1 –1
–1.5
–1.5
0.4 0.4
0.2 0.2

0 0
" 0 " 0
–1 –1
–2 –2
–3 –3
0 20 40 60 80 100 0 20 40 60 80 100
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Im (z)

Im (z)

Plant Zero Plant Zero

0 Model Pole 0 Model Pole
–0.1 –0.1
–1 –0.5 0 0.5 1 1.5 –1 –0.5 0 0.5 1 1.5
Re (z) Re (z)
(a) (b)

FIGURE 7 Example 1: Command following and disturbance rejection for the discrete-time asymptotically stable single-input, single-out-
put plant (46) using rk / 1 and nt = 1 1 n = 2 with a strictly proper model. (a) The single-step disturbance is d k / 0.8. The output y k
approaches the step command with a decreasing command-following error, which is 2.9e-2 at k = 100. (b) The multistep disturbance
d k is given by (52). For each step disturbance, y k approaches the command with a decreasing command-following error, which is 1.2e-1
at k = 19, 2.1e-1 at k = 39, and 1.7e-1 at k = 100. Note that, in both (a) and (b), y k approaches the command more slowly than in the
cases nt = 2 and nt = 3 considered in Figures 5 and 6, respectively. TIR: the L 2-norm difference between the 30-step impulse response
of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 77

 is applied. Figures 5–7 show that, for the asymptotically stable
t i, k u k - i + / Fi y k - i - / G i u k - iG / G i-1.
dt k = =- / Ft i, k y k - i + / G
nt nt n n n

i=1 i =1 i =1 i =1 i =1
minimum-phase plant (46), PCAC rejects step disturbances
(51) with nt $ n. Note that self-generated persistency is evident in
the control input arising from the step-command changes and
Figures 5–7 show the response of PCAC for nt = 2, nt = 3, step disturbances. Furthermore, it is seen that, although the
and nt = 1, respectively, with constant command rk / 1, impulse-response error TIR stops decreasing after each tran-
where the error of the disturbance estimate dt k is comput- sient, RLS keeps refining the dc gain of the model for the re-
ed. In Figures 5(a), 6(a), and 7(a), the constant disturbance mainder of the simulation. This indicates exigency within
d k / 0.8 is applied, and in Figures 5(b), 6(b), and 7(b), the identification, as RLS prioritizes the identification of the dc gain
three-step disturbance over the impulse response to achieve command following. G
Example 1 showed that PCAC can follow step commands
0.8, 0 # k 1 20, and reject step disturbances. This ability is further investigated
d k = * -0.4, 20 # k 1 40, (52) within the context of fixed-gain integral control in “Following
1.2, k $ 40 Step Commands Without an Integrator.”

Following Step Commands Without an Integrator

S ince receding-horizon optimization determines control in- pared with linear time-invariant integral control as well as with
puts without using an explicit feedback controller, it is de- PCAC using a fixed model with erroneous dc gain. To do this,
sirable to compare the performance of predictive cost adaptive consider the discrete-time plants
control (PCAC) with the performance of linear time-invariant z - 0.5
G (z) = , (S3)
controllers. In particular, within the context of linear time-invari- (z - 0.2) (z - 0.8)
ant control, asymptotic step-command following is achievable 0.1 (z - 0.5)
G (z) = , (S4)
if and only if the controller or plant includes an integrator. To (z - 0.2) (z - 0.8)
5 (z - 0.5)
investigate asymptotic step-command following under PCAC, G (z) = , (S5)
(z - 0.2) (z - 0.8)
the performance of PCAC using online identification is com-
0.1953 (z - 0.2)
G (z) = , (S6)
(z - 0.5) (z - 0.9)
rk uk yk
Gc (z) G (z) which are asymptotically stable and have dc gains 3.125, 0.3125,
− 15.625, and 3.125, respectively. Note that (S3)–(S5) have the
same poles and zeros and that (S3) and (S6) have the same dc
gains but different poles and zeros. Disturbance and sensor noise
FIGURE S9 Basic servo loop with step command rk and integral are not considered in these examples. A strictly proper model is
controller G c. t 0,k = 0 is enforced.
used for identification in PCAC; that is, G

4 Example S7: Linear Time-Invariant Integral Control

rk
2 Consider the basic servo loop shown in Figure S9 with the in-
yk

0 tegral controller
–2 G c (z) = 0.1 . (S7)
3 z-1
2 Figure S10 shows the closed-loop response with the controller
uk

1 (S7) and plants (S3)–(S6). The closed-loop dynamics are as-

0 ymptotically stable for G(z) given by (S3), (S4), and (S6), and,
Log |yk – rk|

0
S3 thus, the command-following error converges to zero. On the
–2 S4 other hand, the closed-loop dynamics are unstable for G(z)
–4 S5
S6 given by (S5), and, thus, y k diverges. 
0 10 20 30 40 50 60 70 80 90 100
k (Step) Example S8: Predictive Cost Adaptive
Control With Fixed Models
FIGURE S10 Example S7: Linear time-invariant integral control for Let the true plant G(z) be given by (S3). PCAC is applied with the fixed
G(z) given by (S4)–(S6) with the controller (S7). For plants (S3),
models (S3)–(S6) with u min = - 1, u max = 1, Tu min = - 1, Tu max = 1,
(S4), and (S6), the closed-loop system is asymptotically stable,
, = 50, Q r = I , - 1, Pr = 1, and R = I ,. Figure S11 shows that the
and, thus, asymptotic step-command following is achieved. For
the plant (S5), the closed-loop system is unstable. command-following error approaches zero in the cases where

78 IEEE CONTROL SYSTEMS » DECEMBER 2021

Example 2: A Discrete-Time Unstable plant (53) is initialized with y -1 = y -2 = 0 and u 0 = u -1 =
Nonminimum-Phase Plant u -2 = 0. At each time step, PCAC uses (37)–(40) with , = 20,
The goal of this example is to investigate the ability of PCAC to Qr = 2I , - 1, Pr = 5, and R = I , . Let m = 1 and P0 = 10 3 I 2nt .
follow step commands for various choices of initial models; Figure 8 shows the response of PCAC for nt = n = 2
various levels of zero-mean, Gaussian white sensor noise; and using a strictly proper model, where i 0 = a 60 0 0 1@ T,
various model orders nt . Furthermore, this example compares a ! " -10, -1, - 0.1, 0.1, 1, 10 , . Note that the initial model is a
the identification accuracy of RLS within PCAC and in the case finite-impulse response (FIR) model. Figure 8(a) uses the con-
where the closed-loop data are from an LQG controller with a stant command rk / 1, and Figure 8(b) uses the three-step com-
built-in integrator. Consider the discrete-time input–output mand rk given by (47).
SISO plant Figure 9 shows the response of PCAC for rk / 1 and
nt = n = 2 using a strictly proper model with a randomly cho-
y k = 1.4y k - 1 - 0.3y k - 2 + u k - 1 - 1.3u k - 2 . (53) sen initial estimate i 0 . In particular, in Figure 9, parts(a) and (b)
each shows 100 responses of PCAC with i 0 sampled from a
Let u min = -50, u max = 50, Tu min = -10, Tu max = 10, and zero-mean, Gaussian distribution with standard deviations of
v k = 0. No output constraint is considered in this example. The one and two, respectively.

3
2 rk 1
yk

1 0.5
yk

0 rk
0
1 1
0.5
uk

0.5 0
uk

S4
S5
Log |∆DG| Log |yk – rk|

0 0 S6
–1 10–2.12n × 1

"
0
Log |yk – rk|

–5 S3 –2
S4 –3
–10 S5
S6 1
–15 0
0 5 10 15 20 25 30 35 40 45 50
–1
k (Step) –2
FIGURE S11 Example S8: Step-command following using pre- 0 5 10 15 20 25 30
dictive cost adaptive control with fixed models (S3)–(S6). k (Step)
Asymptotic command following is achieved in the cases (S3)
and (S6), where the dc gain of the model is the same as the dc
gain of the plant. Note that asymptotic command following is FIGURE S12 Example S9: Step-command following using pre-
achieved in the case of (S6) despite the fact that all of the poles dictive cost adaptive control with the initial models (S4)–(S6)
and zeros of the fixed model are incorrect relative to the true as well as the initial model given by i 0 = 10 -2 1 2nt # 1. For all four
plant (S3). For (S4) and (S5), where the dc gain of the model is initial models, the command-following error approaches zero.
not equal to the dc gain of the plant, the command-following TDG: the absolute value of the difference between the dc gain
error remains bounded but does not approach zero. of the plant and model.

PCAC uses the fixed models (S3) and (S6), both of which have Pr = 1, R = I ,, m = 1, nt = 2, and P0 = 10 3 I 2nt for the initial models
the correct dc gain. However, in the case where PCAC uses (S4)–(S6) as well as the initial model given by i 0 = 10 -2 1 2nt # 1.
the fixed models (S4) and (S5), both of which have an incor- Figure S12 shows that, for all four initial models, the
rect dc gain, the command-following error remains bounded command-following error approaches zero. Furthermore,
but does not approach zero.  the dc gain of the identified model approaches the dc
gain of (S3). 
Example S9: Predictive Cost Adaptive Control With Various These examples demonstrate that using (TU 1; k, ,) T RTU 1; k, ,
Initial Models That Are Different From the True Plant in (41) does not yield integral action, and, in addition, that esti-
Let the true plant G(z) be given by (S3). PCAC is applied mation of the dc gain of the plant is necessary for asymptotic
r = I , - 1,
with u min = - 1, u max = 1, Tu min = - 1, Tu max = 1, , = 50, Q step-command following.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 79

 Next, let v k be a zero-mean, Gaussian white noise sequence. LQG controller with an integrator. This implies that self-
In particular, in Figure 10, parts(a) and (b) each shows the re- generated persistency facilitates improvement in the iden-
sponse of PCAC using rk / 1, the strictly proper FIR model tification accuracy. G
i 0 = 60 (2nt - 1) # 1 1@ with n
T T
t = 2, nt = 3, and nt = 4, for v k with
standard deviations of 0.05 and 0.15, respectively. Figure 11 Example 3: An Unstably Stabilizable Plant
shows the accuracy of the model identified by RLS as the stan- The goal of this example is to investigate the ability of PCAC to
dard deviation of v k increases for nt ! " 1, 2, f, 6 , with u k giv- stabilize unstably stabilizable plants. Consider the discrete-
en by receding-horizon optimization and LQG with a built-in time unstable SISO plant
integrator. The accuracy of the model is determined by TIR at
k = 140 averaged over 100 simulation runs. (z - 1.1) (z - 0.4)
G (z) = . (54)
Figures 8–10 show that, for the unstable nonminimum- (z - 1.2) (z 2 + 1.2z + 0.57)
phase (NMP) plant (53), the output approaches the step
commands despite the presence of bias in the estimation of Note that the order of the plant is n = 3, and its relative degree
the model coefficients. Figure 11 shows TIR , and, thus, the is one. The damping ratio and natural frequency of the complex
predictive error of the identified model is smaller with poles of (54) are approximately 0.1 and 0.8r rad/step, respec-
RLS within PCAC than with RLS that uses data from an tively. Since there is an unstable pole on the right side of an

20 rk 10 rk
0 0
α = –10 α = –10
yk

yk

–20 –10
α = –1 α = –1
–40 α = –0.1 α = –0.1
–20
α = 0.1 α = 0.1
20 α=1 10 α=1
α = 10 α = 10
0
uk

uk

0
–20 –10
log yk – rk

log yk – rk

0 1
0
–2
–1
–4

15 15
log ∆IR

log ∆IR

10 10
5 5
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 Plant Pole 0.1 Plant Pole

Plant Zero Plant Zero
Im (z)

Im (z)

0 Model Pole 0 Model Pole

Model Zero Model Zero
–0.1 –0.1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Re (z) Re (z)
(a) (b)

FIGURE 8 Example 2: Command following for the discrete-time unstable nonminimum-phase single-input, single-output plant (53) using
rk / 1 and nt = n = 2 with a strictly proper finite-impulse response model initialized with i 0 = a [0 0 0 1] T, a ! {- 10, - 1, - 0.1, 0.1, 1, 10}.
(a) The single-step command is rk / 1. The output y k approaches the step command with a decreasing command-following error, which
is in the interval [2.1e-3, 1.0e-2] at k = 60. Note that, for all a, TIR approaches asymptotic values for k $ 25. The bottom-most plot com-
pares the poles and zero of the identified model at k = 60 to the poles and zero of the plant. (b) The multistep command rk is given by
(47). For a ! - 10, y k approaches the three-step command with a decreasing command-following error, which is in the interval [4.3e-1, 1.1]
at k = 19, [3.3e-2, 6.7e-2] at k = 39 , and [8e-2, 1.2e-1] at k = 60. For a = - 10, y k approaches the last command with a decreasing com-
mand-following error, which is 1.3e-1 at k = 60. Note that, for all a, TIR approaches asymptotic values for k $ 25. The bottom-most plot
compares the poles and zero of the identified model at k = 60 to the poles and zero of the plant. TIR: the L 2-norm difference between
the 30-step impulse response of the plant and model.

80 IEEE CONTROL SYSTEMS » DECEMBER 2021

NMP zero, pole-zero interlacing fails in the open right-half 1, 0 # k 1 100,
plane, and, thus, the plant is unstably stabilizable. That is, it can rk = * -1, 100 # k 1 200, (55)
be stabilized by a discrete-time unstable linear time-invariant 2, k $ 200.
controller but cannot be stabilized by a discrete-time asymp-
totically stable linear time-invariant controller [73]. Next, consider the continuous-time SISO plant
Let u min = - 50, u max = 50, Tu min = -10, Tu max = 10, , = 50,
Qr = 4I , - 1, Pr = 4, R = I ,, m = 1, P0 = 10 3 I 2nt , and i 0 = 60 T(2nt - 1) # 1 1@T.
(s - 0.1) (s + 0.6)
A strictly proper model is used for identification in PCAC. G (s) = , (56)
(s - 0.2) (s 2 + 2g~ n s + ~ 2n)
No output constraint is considered in this example. The
plant is initialized with y -1 = -0.4, y -2 = 0.3, and u -1 =
u -2 = u -3 = 0. PCAC uses (37)–(40), where the standard de- where g = 0.1 and ~ n = 0.8r. Note that, as in the case of the
viation of the sensor noise v k is 0.02. In this and the following discrete-time plant (54), the continuous-time plant (56) can be
examples, identification and control do not commence until stabilized by a continuous-time unstable linear time-invariant
the regressor matrix z k given by (15) is populated with nt controller but cannot be stabilized by a continuous-time as-
measurements. Figure 12(a) shows the response of PCAC ymptotically stable linear time-invariant controller [73]. With
for the discrete-time plant (54) for nt = n = 3 using the three- the sample period Ts = 1 s, Figure 12 shows that the ZOH-
step command discretized plant has the same dynamics as the discrete-time

0 20
0
yk

yk

–5 –20
rk rk
–10 –40
2
20
0
–2 0
uk

uk

–4
–20
0
log yk – rk

log yk – rk

0
–2 –2
–4 –4
–6 –6

10 20
log ∆IR

log ∆IR

5 10
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
k (Step) k (Step)

0.1 PlantPole
Plant Pole 0.1 PlantPole
Plant Pole
PlantZero
Plant Zero PlantZero
Plant Zero
Im (z)

Im (z)

0 ModelPole
Pole 0 ModelPole
Pole
Model Model
ModelZero
Model Zero ModelZero
Model Zero
–0.1 –0.1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Re (z) Re (z)
(a) (b)

FIGURE 9 Example 2: Command following for the discrete-time unstable nonminimum-phase single-input, single-output plant (53) using
rk / 1 and nt = n = 2 with a strictly proper model, which is initialized with 100 Gaussian-distributed samples of i 0 with a standard devia-
tion of v. (a) For v = 1, the output y k approaches the step command with a decreasing command-following error, which is in the inter-
val [9.5e-4, 5.5e-3] at k = 60. (b) For v = 2, the output y k approaches the step command with a decreasing command-following error,
which is in the interval [6.7e-3, 2.1e-2] at k = 60. Note that, in both (a) and (b), although TIR approaches an asymptotic value for k $ 25,
indicating the presence of bias and, thus, lack of consistency in the estimate of the model, the poles and zero of the model are close to
the poles and zero of the plant at k = 60. TIR: the L 2-norm difference between the 30-step impulse response of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 81

plant (54) and, thus, is not stabilizable by a discrete-time Figure 12(b) shows the response of PCAC for the continuous-
asymptotically stable linear time-invariant controller. The plant time plant (56) using the same setup as for the discrete-time
is initialized with x (0) = 60.4 - 0.8 1.1@T, where x(0) is the ini- plant (54). Figure 12 shows that PCAC stabilizes the discrete-
tial state of the realization time (54) and sampled-data (56) unstably stabilizable plants.
Since the plant is unstable, an FIR initial model—all of whose
poles are zero—is a poor choice. Consequently, choosing a large
- 0.3027 -3.1 0.6317 1
xo = > 2 0 H x + >0H u, (57)
optimization horizon yields poor model predictions, which, in
0
turn, yield a large initial transient response. This example sug-
0 1 0 0
gests that, for unstable plants, it is advisable to avoid choosing
y = 61 0.25 - 0.03@ x. (58) large values of the prediction horizon ,.  G

20
5 rk rk
0
yn,k

yn,k
0
" " "
n=2

" " "

–20 n=2
–5 n=3 n=3
2 n=4 5 n=4
0 0
uk

–2 uk –5
–10
–4
log ∆IR log yn,k – rk

log ∆IR log yn,k – rk

1
0 0
–2 –1
–2
–4 –3

2 3
1 2
0 1
0
–1
0 20 40 60 80 100 120 0 20 40 60 80 100 120
k (Step) k (Step)

Re (z) Re (z)
–1 –0.5 0 0.5 1 1.5 2 2.5 –1 –0.5 0 0.5 1 1.5 2 2.5

0.5 0.5
Im (z)

Im (z)

0 0
–0.5 –0.5

0.01 0.01
Im (z)

Im (z)

0 0
–0.01 –0.01
1 1.1 1.2 1.3 1.4 1.5 1 1.1 1.2 1.3 1.4 1.5
Re (z) Re (z)

Plant Pole Plant Zero Plant Pole Plant Zero

"

"

"

"

Model Pole (n = 2) Model Zero (n = 2) Model Pole (n = 2) Model Zero (n = 2)

"

"

"

"

Model Pole (n = 3) Model Zero (n = 3) Model Pole (n = 3) Model Zero (n = 3)

"

"

"

"

Model Pole (n = 4) Model Zero (n = 4) Model Pole (n = 4) Model Zero (n = 4)

Unit Circle Unit Circle

(a) (b)

FIGURE 10 Example 2: Command following for the discrete-time unstable nonminimum-phase single-input, single-output plant (53) using
rk / 1; noisy measurement with a noise standard deviation of v; and finite-impulse response strictly proper initial model i 0 = [0 T(2nt - 1) # 1 1] T,
where nt = 2, nt = 3, and nt = 4. (a) For v = 0.05, in all cases, y k approaches the step command with a decreasing command-following
error. (b) For v = 0.15, y k approaches the step command with a decreasing command-following error for nt = 3 and nt = 4. For nt = 2, y k
does not approach the step command, and the command-following error is 7.9 at k = 120. Note that, in both (a) and (b), larger values of nt
yield smaller values of TIR. The bottom-most plots show that larger values of nt facilitate identification, and the poles and zero of the model
are closer to the poles and zero of the plant. TIR: the L 2-norm difference between the 30-step impulse response of the plant and model.

82 IEEE CONTROL SYSTEMS » DECEMBER 2021

Example 4: An Asymptotically Stable R-0.14 - 2.5 -0.2
- 2VW R0.5V
S S W
Continuous-Time Plant S 2 0 0
0W S0W
The goal of this example is to investigate the application of
xo = S
0 1 0W
0
x + S 0 W (u + d ), (61)
SS WW SS WW
PCAC for sampled-data control with harmonic and broadband T 0 0 1
0X T0X
disturbance rejection on a lightly damped plant. Consider the y = 60 0 0 1@ x. (62)
continuous-time SISO plant
Figures 13 and 14 show that, for the asymptotically stable
G (s) = 1 , (59)
(s 2 + 2g 1 ~ n, 1 s + ~ 2n, 1) (s 2 + 2g 2 ~ n, 2 s + ~ n2 , 2) lightly damped plant (59), PCAC rejects harmonic and broad-
band disturbances. Note that, in the presence of a harmonic
where g 1 = 0.05, ~ n,1 = 1, g 2 = 0.01, and ~ n,2 = 2. The data are disturbance, a pole-zero cancellation occurs at the frequency of
sampled with sample period Ts = 1 s. Note that the order of the the harmonic disturbance, which serves as an implicit internal
plant is n = 4. Let u min = - 2, u max = 2, Tu min = - 2, Tu max = 2, model of the unknown disturbance. This pole-zero cancellation
i 0 = 60 (2nt - 1) # 1 1@ rk / 0, m = 1, and P0 = 10 I 2nt . No output
T T 3
indicates exigency within the closed-loop identification, where
constraint is considered in this example. At each time step, RLS captures the frequency and amplitude of the harmonic dis-
PCAC solves (37)–(40) with , = 10, Q r = 50I , - 1, Pr = 50, and turbance to predict and cancel the harmonic disturbance for
R = 0.1I , . A strictly proper model is used for identification in receding-horizon optimization. Note that this pole-zero cancel-
PCAC. Let v k be a zero-mean, Gaussian white noise sequence lation requires that nt $ n + 2~ d, where ~ d is the number of
with a standard deviation of 0.005. pure harmonic tones in the disturbance d k . For broadband dis-
Figure 13 shows harmonic disturbance rejection, and turbances, as the order of the model increases, the frequency
Figure 14 shows broadband disturbance rejection using response of the identified model becomes more accurate, thus
PCAC with nt = n = 4, nt = 5, and nt = 6. The matched continu- manifesting exigency. G
ous-time harmonic disturbance is given by
Example 5: A Continuous-Time Unstable Plant
d (t) = A d sin (~ d t), (60)
The goal of this example is to investigate PCAC’s ability to sta-
where A d = 1 and ~ d = r/15. The discrete-time matched bilize an unstable sampled-data plant with a rigid-body mode
broadband disturbance is given by d k, where d k is a zero-mean, as well as enforce output and control constraints. Additionally,
Gaussian white noise sequence with a standard deviation of this example investigates the effect of the constraint slack
0.5. Figures 13(a) and 14(a) consider zero initial conditions for weight S. Consider the continuous-time SISO plant
(59), and Figures 13(b) and 14(b) consider nonzero initial condi-
tions, where x (0) = 610 -5 -8 300@T is used as the initial state 1
G (s) = . (63)
of the realization s 2 (s 2 + 2gs + 1)

10 30
n=2 n=2
" " " " "

" " " " "

25
8 n=3 n=3
n=4 20 n=4
n=5 n=5
6
log (∆IR)

n=6 15 n=6
∆IR

4 10
5
2
0
0 –5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
σ σ
(a) (b)

FIGURE 11 Example 2: Accuracy of the identified model versus the standard deviation v of the sensor noise for the discrete-time
unstable nonminimum-phase single-input, single-output plant (53) with rk / 1 and with the controllers given by predictive cost adaptive
control (PCAC) and linear quadratic Gaussian (LQG) with an integrator. The model accuracy is determined by TIR at k = 140 averaged
over 100 simulations. PCAC uses a strictly proper finite-impulse response initial model, and LQG with an integrator uses (53) to compute
the control gain with output and control weights Q = 2 and R = 1, respectively. (a) PCAC. For each v, larger values of nt yield smaller
values of TIR. Therefore, for each sensor-noise level, the identified model is more accurate for larger values of nt . (b) LQG with an inte-
grator. For each nt , larger values of v yield larger values of TIR. Note that, unlike (a), for all v $ 0.05, an increase in the model order nt
results in a less accurate model. In addition, for each value of nt , the models in (b) are less accurate than those in (a). TIR: the L 2-norm
difference between the 30-step impulse response of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 83

The data are sampled with sample period Ts = 1 s. Let u min = -10, Denote the ith component of u k, Tu max, Tu min, and u 1|k, by
u max = 10, Tu min = -5, Tu max = 5, m = 1, P0 = 10 3 I 2nt , , = 20, u k, i, Tu max, i Tu min, i, and u 1|k, i, respectively. In this example, for
Qr = 40I , - 1, Pr = 40, and R = 10I , . PCAC is initialized with each i = 1, f, m, the saturation
the strictly proper FIR model i 0 = 60 T(2nt - 1) # 1 1@ , and v k is a ze-
T

ro-mean, Gaussian white noise sequence with a standard de-

u k + 1, i =
viation of 0.001. Let the output constraint be given by (5),
min (max (u k, i + Tu max, i, u min, i), u max, i), u 1|k, i - u k, i 2 Tu max, i
where C = 61 - 1@T, and D = 6-20 -20@T. A realization of (63)
is given by
* min (max (u k,i + Tu min,i, u min,i), u max,i), u 1|k,i - u k,i 1 Tu min,i,
min (max (u 1|k, i, u min, i), u max, i), otherwise
(66)
R-2g -1 0 0VW R1V
S S W
S 1 0 0 0W S0W
xo = S x + S W u, (64)
0 1 0 0W 0 is applied.
SS WW SS WW
Figure 15 shows the response of PCAC for various values
T 0 0 1 0X T0X
of the constraint relaxation weight S. Figure 15(a) shows the
y = 60 0 0 1@ x. (65) response of PCAC using (37)–(40) as well as from (41)–(45)

4 rk 4 rk
2 2
0
yk

yk

0
–2
–4 –2
20 40
20
0
uk

uk

0
–20 –20
1
1
log yk – rk

log yk – rk

0 0
–1 –1
–2 –2
–3
–3
2
2
log ∆IR

log ∆IR

1 1
0 0
–1 –1

0 50 100 150 200 250 300 0 50 100 150 200 250 300
k (Step) t (s)

Plant Pole Plant Pole

0.5 0.5
Plant Zero Plant Zero
Im (z)

Im (z)

0 Model Pole 0 Model Pole

–0.5 Model Zero –0.5 Model Zero
Unit Circle Unit Circle
–1 –0.5 0 0.5 1 1.5 2 2.5 –1 –0.5 0 0.5 1 1.5 2 2.5
Re (z) Re (z)
(a) (b)

FIGURE 12 Example 3: Command following for the unstably stabilizable single-input, single-output plants (54) and (56) using nt = n = 3
and the three-step command (55). (a) For the discrete-time plant (54), the output y k approaches each step command with a decreasing
command-following error. The bottom-most plot compares the poles and zeros of the identified model at k = 300 to the poles and zeros
of the plant. It was observed (not shown) that the response time was larger than 100 steps for 32 # , # 40, and PCAC was unable to
follow the step commands for , # 31. For , 2 85, it was observed (not shown) that the output diverged. (b) For the continuous-time plant
(56), the output y(t) approaches each step command with a decreasing command-following error. The bottom-most plot compares the
poles and zeros of the identified model at t = 300 s to the poles and zeros of the exact discretization of the continuous-time plant (56).
Note that the discretized plant is unstably stabilizable. TIR: the L 2-norm difference between the 30-step impulse response of the plant
and model.

84 IEEE CONTROL SYSTEMS » DECEMBER 2021

with nt = n = 4, g = 0.7, x (0) = 61 -3 2 0.5@T, and the three- 0, 0 # k 1 120,
step command rk = * 19, 120 # k 1 150, (68)
- 25, k $ 150.
0, 0 # k 1 50,
rk = * 19, 50 # k 1 80, (67)
Figure 15 shows that, for the unstable plant (63), PCAC
-25, k $ 80.
stabilizes the plant, and the output approaches the feasible
Figure 15(b) shows the response of PCAC using (41)–(45) with commands. Note that the nonzero constraint relaxation weight
nt = 6 2 n = 4, g = 0.01, x (0) = [10 -5 8 3]T, and the three- reduces the overall constraint violation, especially in cases
step command where the combination of the commands and constraints is

2 2
0 0
Open Loop Open Loop

y (t )
y (t )

–2 " " " –2

" " "

n=4 n=4
n=5 n=5
2 d (t) 2 d (t)
n=6 n=6
0 0
uk

uk
–2 –2
0 100 200 300 400 500 0 100 200 300 400 500
t (s) t (s)

0 0
Gain (dB)

Gain (dB)

–20 –20
–40 –40
Plant Plant

" " "

n=4
" " "

0 n=4 0
n=5 n=5
Phase (°)

Phase (°)

–180 n=6 –180 n=6

–360 –360

0 π /4 π /2 3π /4 π 0 π /4 π /2 3π /4 π
θ (rad/Sample) θ (rad/Sample)
1 1
Im (z)

Im (z)

0 0

–1 –1
–1 –0.5 0 0.5 1 1.5 2 2.5 –1 –0.5 0 0.5 1 1.5 2 2.5
Re (z ) Re (z)

Plant Pole Plant Zero Plant Pole Plant Zero

"

"

"

"

Model Pole (n = 4) Model Zero (n = 4) Model Pole (n = 4) Model Zero (n = 4)

"

"

"

"

Model Pole (n = 5) Model Zero (n = 5) Model Pole (n = 5) Model Zero (n = 5)

"

"

"

"

Model Pole (n = 6) Model Zero (n = 6) Model Pole (n = 6) Model Zero (n = 6)

Unit Circle Unit Circle

(a) (b)

FIGURE 13 Example 4: Matched disturbance rejection for the continuous-time asymptotically stable single-input, single-output plant (59)
and harmonic disturbance (60) with nt = n = 4, nt = 5, and nt = 6. (a) Zero initial conditions. Note that larger values of nt yield a better
performance and that u k approximately cancels d(t) for nt = 6. Furthermore, larger values of nt yield a more accurate frequency response
of the identified models at t = 500 s. For nt = 6, the identified model has an approximate pole-zero cancellation at the disturbance fre-
quency i d = r/15 rad/sample, which is denoted by the vertical dash–dotted line; the spurious pole in the identified model allows
receding-horizon optimization to predict the future harmonic response of the plant, which facilitates the optimization of the future control
inputs. This pole-zero cancellation serves as an implicit internal model of the unknown disturbance. The bottom-most plot compares the
poles and zeros of the identified model at t = 500 s to the poles and zeros of the exact discretization of the plant. For nt = 6, the approx-
imate pole-zero cancellations on the unit circle occur at the frequency i d of the harmonic disturbance. (b) Nonzero initial conditions
x (0) = [10 - 5 - 8 300] T for the state-space realization (61), (62). Compared to the case of zero initial conditions, the transient
response is more pronounced, and the control saturation occurs over a longer interval of time. Moreover, since nonzero initial conditions
result in more persistency during the transient, the identified models in (b) are more accurate than those in (a).

DECEMBER 2021 « IEEE CONTROL SYSTEMS 85

 Although PEM and IV are more accurate than RLS/VRF for both
open- and closed-loop identification, these methods were found
to be less compatible with PCAC than RLS/VRF.

infeasible. Furthermore, the constraint violation is reduced as zu = Ex, (70)

the model accuracy improves. G
where w (t) ! R lw is an unmatched disturbance, yu (t) ! R l yu ,
LINEAR MULTIPLE-INPUT, MULTIPLE-OUTPUT zu (t) ! R l zu, A ! R lx # lx, B ! R lx # m, C ! R l yu # lx, D ! R l yu # m,
EXAMPLES D 1 ! R lx # lw , and E ! R l zu # lx. At each time step, PCAC uses the
sampled measurement, where y k = 6yu Tk zu Tk @ ! R l yu + l zu, and the
T
This section considers linear time-invariant MIMO plants with
the state-space realization (1) along with tracking output (3) and constrained output (4) are based on zu k .
Note that both yu k and zu k are noisy measurements, and zu k is
yu = Cx + Du, (69) used to define the performance metric.

1 200
y (t )

0
y (t )

0
Open Loop Open Loop
–1 –200

" "
n=4
" " "

n=4 –400
n=5 n=5
2 dk 2 dk

"
n=6 n=6
0 0
uk

uk

–2 –2
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
t (s) t (s)

0 0
Gain (dB)
Gain (dB)

–20 –20 Plant

Plant Open Loop
–40 Open Loop –40
" " "

n=4
" " "

n=4 n=5
360 0
n=5 n=6
180
Phase (°)

Phase (°)

n=6 –180
0
–180 –360
–360

0 0
log (PSD)

log (PSD)

–2 –2
–4 –4
–6 –6
0 π /4 π /2 3π /4 π 0 π /4 π /2 3π /4 π
θd (rad/Sample) θd (rad/Sample)
(a) (b)

FIGURE 14 Example 4: Matched disturbance rejection for the continuous-time asymptotically stable single-input, single-output plant (59)
and zero-mean, Gaussian white noise sequence d k with a standard deviation of 0.5 for nt = n = 4, nt = 5, and nt = 6. (a) Zero initial con-
ditions. Note that larger values of nt yield more accurate magnitudes of the frequency response of the identified models at t = 5000 s.
The bottom-most plot shows that, for all nt , predictive cost adaptive control (PCAC) reduces the two peaks of the open-loop power
spectral density (PSD) of y k, which indicates that PCAC rejects the broadband disturbance d k . (b) Nonzero initial conditions
x (0) = [10 - 5 - 8 300] T for the state-space realization (61) and (62). As in (a), the bottom-most plot shows that, for all nt , PCAC
reduces the two peaks of the open-loop PSD of y k . Note that, for all nt , the frequency responses of the identified models at t = 5000 s
are more accurate in (b) than in (a) due to the nonzero initial conditions.

86 IEEE CONTROL SYSTEMS » DECEMBER 2021

Example 6: An Asymptotically Stable Plant With R-0.1 0 V
S W
Nonminimum-Phase Transmission Zeros and S 0 0 W
Nonminimum-Phase Channel Zeros S 0 0 W
D1 = S W, E = 62 -2.1 0 0 0 0@ . (73)
The goal of this example is to investigate the applicability of S 0 0 W
PCAC MIMO plants. Consider the continuous-time plant S 0 0 W
S W
MIMO (1)–(70), where T 0 - 0.5X
R 0 1 0 0 0 0 V R0 0V
S W S W The transmission zeroes (TZs) and channel zeroes (CZs) of
S-1.4 -0.06 0.4 0.05 0 0 W S1 0W
S 0 0 0 1 0 0 W S0 0W (71)–(73) are shown in Figure 16. The data are sampled with
A=S
S 0.8 0.1 -1 -0.16 0.2 0.06 W
W, B = S W, (71)
S0 2W sample period Ts = 0.5 s. Let u min = 6-2 -1@T, u max = 62 1@T,
S 0 0 0 0 0 1 W S0 0W T u min = 6-1 -0.5@T, T u max = 61 0.5@T, m = 1, P0 = 10 3 I 2nt ,
S
0 0 1 0 . 3 1 0 . 3
W S W , = 10, Qr = 40I , - 1, Pr = 40, R = 610 0 @ 7 I ,, and rk / 0. Let v k
T - - X T0 0X 0 10
be a zero-mean, Gaussian white noise sequence with a
C=; E, D = 0 2 # 2, (72)
0 0 -3.1 3.3 0 0
0 0 0 0 0.31 -0.33 standard deviation of 0.02. The plant is initialized with

200 600
rk rk
400
200
y (t )

y (t )
0
0
No Slack –200 S = 0.1
–200 S = 0.1 S = 10
20 S = 10 S = 1000
S = 1000 10 S = 5000
10
0
uk

uk

0
–10
–10

40 8000
6000
εk∞

εk∞

20 4000
0 2000
0
log ∆IR

log ∆IR

2 2

0 0
–2
–2
0 20 40 60 80 100 120 0 50 100 150 200
t (s) t (s)

1 Plant Pole 1 Plant Pole

Plant Zero Plant Zero
Im (z )

Im (z )

0 Model Pole 0 Model Pole

Model Zero Model Zero
–1 Unit Circle –1 Unit Circle
–1 –0.5 0 0.5 1 1.5 2 –1 –0.5 0 0.5 1 1.5 2
Re (z) Re (z)
(a) (b)

FIGURE 15 Example 5: Command following with the output constraint - 20 # y k # 20 and continuous-time unstable single-input, single-
output plant (63) for various values of S. (a) Predictive cost adaptive control (PCAC) is applied using (37)–(40) and (41)–(45) with
nt = n = 4, g = 0.7, x (0) = [1 - 3 2 0.5] T, and the three-step command (67). At t = 9 s, (37)–(40) become infeasible, and the three-step
command cannot be followed. In contrast, with (41)–(45), y(t) approaches the first two step commands, and the slack is activated at
t = 8 s, t = 50 s, and t = 80 s, as can be seen in the third plot. Since the last step command is not achievable, y(t) moves further away
from the command as S increases, where the command-following error at t = 120 s is approximately 0.02 for S = 0.1, 1.00 for S = 10,
and 4.81 for S = 1000. Furthermore, note that the overshoot at t = 19 s violates the output constraint due to the poor initial model and
nonzero initial condition. The distance between y(t) and the constraint at t = 19 s for S = 0.1 is 41.47, which is reduced by 47% for
S = 10 and 89% for S = 1000. (b) PCAC is applied using (41)–(45) with nt = 6 2 n = 4, g = 0.01, x (0) = [10 - 5 8 3] T, and the three-
step command (68). Compared to (a), note that, for all S, the output-constraint violation is more severe during the first transient. How-
ever, as the model becomes more accurate, y(t) approaches the first and second step commands, where the constraint is enforced after
t = 106 s. For the third step command, as in (a), y(t) moves away from the step command as S increases. TIR: the L 2-norm difference
between the 30-step impulse response of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 87

x(0) = [1 -3 5 0.5 2 -1]T, and PCAC uses (37)–(40) with the R 0 1 0 0 0 0 V
S W
strictly proper FIR initial model i 0 = 60 np
t (m + p) # 1 1 mp # 1@ . No
T T
S-1.4 -0.06 0.4 0.05 0 0 W
output constraint is considered in this example, and the S 0 0 0 1 0 0 W
A=S W,
tracking output y t,k is given by (3) with C t = 60 0 1@; that is, S 0. 8 0.1 -1 -0.04 0.2 -0.06W
y t,k = zu k . S 0 0 0 0 0 1 W
S W
Figure 17 shows broadband disturbance rejection for nt = 3, T 0 0 1 -0.3 -1 0.3 X
tn = 4, and nt = 6. The discrete-time unmatched broadband dis-
E=; E.
2 -2.1 0 0 0 0
turbance is given by the random vector w k = 6w 1, k w 2, k@T ! R 2, 0 1 0 0 0 -1.1
(74)
where the components w 1, k and w 2, k are zero-mean, Gaussian
white noise sequences with standard deviations of 1.7 and 2.5, The TZ and CZ of (74) are shown in Figure 18. The data are
respectively. Figure 17 shows that, for the MIMO plant (71)– sampled with sample period Ts = 0.5 s. Let u min = 6-15 - 5@ T,
(73) with NMP TZ and NMP CZ, PCAC rejects an unmatched u max = 615 5@ T, Tu min = 6-5 -2@ T, Tu max = 65 2@ T, m = 1,
broadband disturbance. G P0 = 10 3 I 2nt , , = 40, Q r = 40I , - 1, Pr = 40, and R = 610 0 @ 7 I , .
0 10
Let k be a zero-mean, Gaussian white noise sequence with a
v
Example 7: An Unstable Plant With Nonminimum-Phase standard deviation of 0.02. The plant is initialized with
Transmission Zeros and Nonminimum-Phase Channel Zeros x (0) = 62 -3.3 4 -2.6 2 -1.3@ T, and PCAC uses (41)–(45)
The goal of this example is to investigate the ability of PCAC to w i t h t h e s t r i c t l y p ro p e r F I R i n i t i a l m o d e l i 0 =
stabilize MIMO plants while enforcing output constraints. 60 T(nt (m + p) p - mp) # 1 1 mp # 1@ T. The tracking output y t,k is given by (3)
Consider the continuous-time MIMO plant (1)–(70), where B is with C t = 61 0@, and the constrained output is given by (4) with
given by (71), C and D are given by (72), D 1 = 0 6 # 2, and C c = 60 1@; that is, y t, k = zu 1, k , and y c = zu 2,k . The output con-
straint is given by (5), where C = 61 -1@ T, and D = 6-10 -10@ T.
Figure 19 shows the response of PCAC for nt = 2 and vari-
ous values of S using the three-step command
0.5 0, 0 # k 1 40,
rk = * 10, 40 # k 1 80, (75)
Im (s)

0 -20, k $ 80.

Figure 18 shows that, for the unstable MIMO plant (71), (72),
–0.5 and (74) with NMP TZ and NMP CZ, PCAC stabilizes the plant
–1 0 1 2 3
Re (s) while asymptotically satisfying the output constraints, and the
output approaches the step commands. G
TZ From u to y~
TZ From u to z~
LINEAR TIME-VARYING EXAMPLES
(a) This section considers linear time-varying SISO plants, where
2 y t, k = y c, k = y k .

1 Example 8: Abrupt Change From an Asymptotically

Stable Plant to an Unstable Nonminimum-Phase Plant
Im (s)

0
The goal of this example is to investigate the applicability of
–1 PCAC with VRF on time-varying plants. Consider a discrete-
time asymptotically stable SISO plant (46) that abruptly changes
–2 at step k c to a discrete-time unstable NMP SISO plant (53). Note
–10 –5 0 5
Re (s) that the orders of both plants are n = 2 with relative degree one.
Let u min = -50, u max = 50, Tu min = -25, Tu max = 25, , = 20,
CZ From u1 to y~1 CZ From u2 to y~1 Qr = 2I , - 1, Pr = 5, R = I ,, nt = n = 2, and P0 = 10 3 I 2nt . No output
CZ From u to y~
1 2 CZ From u to y~
2 2
constraint is considered in this example. The plant is initialized
CZ From u1 to z~ CZ From u2 to z~ with y -1 = -0.2, y -2 = 0.4, and u -1 = u -2 = 0. PCAC uses the
strictly proper FIR initial model i 0 = 60 1 # 3 1@ T and uses (37)–
(b)
(40) with the measurement y k . Let v k = 0.
Figure 20 shows the response of PCAC for rk / 1 using CRF
FIGURE 16 Example 6: Transmission zeros (TZs) and channel with m = 1 and VRF (21) with h = 0.9, x n = 5, and x d = 10 for
zeros (CZs) of the continuous-time asymptotically stable multiple- k c = 30, k c = 200, k c = 400, and k c = 600. Figure 20(a) shows
input, multiple-output plant (71)–(73). (a) The transfer function
from u to [yu T zu T] T has two nonminimum-phase (NMP) TZs at that, in the absence of forgetting, that is, m = 1, reidentification
approximately z t = 0.94. (b) Each transfer function has one NMP of the plant after the abrupt change is poor, and the output does
CZ at approximately 0.94. not approach the step command. In contrast, Figure 20(b)

88 IEEE CONTROL SYSTEMS » DECEMBER 2021

 10

z (t )
0

~
Open Loop
–10

"
n=3

"
n=4
20

"
n=6
y (t )

0
~

–20

0
uk

–2

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
t (s)
2
log (PSD(z ))
~

0
–2
–4

0 π /4 π /2 3π /4 π
θ d (rad/Sample)

40
Gain (dB)
Gain (dB)

40
u2 to y~1

20 Plant
u1 to y~1

Plant
0
20
–20

360 180
Phase (°)
Phase (°)

u2 to y~1
u1 to y~1

180
0
0
–180 –180
20
Gain (dB)

20
Gain (dB)

0
u2 to y~2
u1 to y~2

–20 0
–40 –20
–60 –40
–80
0 720
Phase (°)
Phase (°)

u2 to y~2
u1 to y~2

360
–360
0
–720 –360
Gain (dB)

20
Gain (dB)

20
u2 to z~
u1 to z~

0
10 –20
–40
0 –60
0 360
Phase (°)
Phase (°)

u2 to z~

180
u1 to z~

–180 0
–180
–360 –360
0 π /4 π /2 3π /4 π 0 π /4 π /2 3π /4 π
θ d (rad/Sample) θ d (rad/Sample)

FIGURE 17 Example 6: Unmatched broadband disturbance rejection for the continuous-time asymptotically stable multiple-input, multiple-
output plant (71)–(73) with a nonminimum-phase (NMP) transmission zero and an NMP channel zero for nt = 3, nt = 4, and nt = 6 using the
discrete-time broadband disturbance w k = [w 1, k w 2, k] T ! R 2, where w 1,k and w 2,k are zero-mean, Gaussian white noise sequences with 1.7
and 2.5, respectively. Note that, for all nt , the magnitude of the open-loop power spectral density (PSD) of zu in the range r/16 # i d # r/4
is reduced, which indicates suppression of the broadband disturbance w k . For all six channels, namely, from u 1 and u 2 to yu 1, yu 2, and zu ,
the 12 bottom-most plots compare the frequency response of the identified model at t = 5000 s to the frequency response of the plant. As
in Example 4, this example shows that a larger model order nt yields more accurate estimates of the plant dynamics.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 89

 shows that, with VRF, reidentification of the plant after the
0.5 abrupt change is better compared to Figure 20(a), and, thus, the
output approaches the step commands. G
TZ From u to y~
Im (s)

0 TZ From u to z~
Example 9: Abrupt Change From an Asymptotically
Stable Plant to an Unstable Nonminimum-Phase
–0.5 Plant With Sensor Noise
–1 0 1 2 3
Re (s) The goal of this example is to investigate the effect of zero-
(a) mean, Gaussian white sensor noise on VRF. This example uses
the same setup as Example 8 except that PCAC now uses noisy
2
measurements with the command profile
1 Z
] 20, 0 # k 1 100,
] -20, 100 # k 1 200,
Im (s)

0 rk = [ (76)
]] 3, 200 # k 1 400,
–1 -50, k $ 400.
\
–2 Figure 21 shows the response of PCAC for two different levels
–10 –5 0 5
of sensor noise using k c = 300 and VRF (21) with h = 0.5,
Re (s)
x d = 5x n, and various values of x n . As shown in Figure 21(a)
and (b), as x n increases, the duration during which forgetting is
CZ From u1 to y~1 CZ From u2 to y~1
active increases. Hence, as x n increases, reidentification be-
CZ From u to y~
1 2 CZ From u to y~
2 2
comes slower and, for large values of x n, the output does not
CZ From u1 to z~1 CZ From u1 to z~2
approach the command. Note that, as x n increases, the VRF m k
CZ From u2 to z~1 CZ From u2 to z~2
becomes less sensitive to sensor noise.
(b) Note that the speed of convergence of the coefficients after
the abrupt change determines whether or not the output ap-
FIGURE 18 Example 7: The transmission zero (TZ) and channel proaches the command. In particular, with a large choice of x n,
zero (CZ) of the continuous-time unstable multiple-input, multiple- the reconvergence of the model coefficients is slow, which
output plant (74). (a) The transfer function from u to [yu zu T] T has yields a divergence of the output. Furthermore, in the presence
four nonminimum-phase (NMP) TZs at approximately 0.00, 0.94,
and 3.33. (b) The eight transfer functions have 21 NMP CZs at of zero-mean, Gaussian white sensor noise, where x d = 5x n,
approximately 0.00, 0.13 ! 1.42., 0.14 ! 1.18., 0.15 ! 0.99., 0.19, increasing the value of x n reduces the activation of forgetting
0.95, 2.10, and 3.33. due to sensor noise. G

Guidelines for Selecting Hyperparameters

T his is a list of the hyperparameters that need to be select-
ed for predictive cost adaptive control as well as heuristic
changes but is more sensitive to sensor noise. On the
­o ther hand, larger x n yields identification that reacts
guidelines on how to select them. slowly to changes but is less sensitive to sensor noise.
• Identification model order nt : Choose nt to be greater than • VRF parameter h: 0.001 # h # 1. Larger values of h yield
or equal to the dimension of the unstable subspace of the more aggressive forgetting in response to changes.
plant. Add +2 for each harmonic disturbance. • Prediction horizon ,: Select the prediction horizon , so that
• Initial identification covariance P0: Typically, choose ,Ts is greater than the settling time of the plant. For unstable
-4
P0 = aI np
t (m + p) + mp, where 10 # a # 10 4. Larger values of plants, excessively large values of , may lead to divergence.
a yield more aggressive identification and may lead to • Cost-to-go output weight Q r : aI (, - 1)p t, a 2 0.
more severe transients in the response. • Terminal output weight Pr : aI p t, a 2 0.
• Initial identification coefficient vector i 0: Use the best • Control move-size weight R: aI ,m, a $ 0.
prior estimate of the plant coefficients when available; • Constraint relaxation weight S: aI ,n c, a $ 0. Choose non-
otherwise, set i 0 = 0.01 $ 1 [npt (m + p) + mp] # 1. zero a when infeasible command and constraints are
• Constant forgetting factor m: Typically, choose 0.95 # m # 1. expected.
Choose m = 1 when persistency is lacking. Not relevant • Control magnitude constraints u min, u max: Select these
when using variable-rate forgetting (VRF). values to reflect actuator magnitude constraints.
• VRF parameters x n, x d: x d = 5x n, where 1 # x n # 100. • Control move-size constraints Tu min, Tu max: Select these
Smaller x n yields identification that reacts quickly to values to reflect actuator move-size constraints.

90 IEEE CONTROL SYSTEMS » DECEMBER 2021

 60
40 rk
z1(t ) 20
0
~

–20
–40 S = 0.1 I2
40 S = 10 I2
20 S = 1000 I2
z2(t )

0
~

–20

100

0
y (t )
~

–100

10
0
uk

–10

6000 40
20
εk∞

4000
0
2000 10 15 20 25
0
0 10 20 30 40 50 60
t (s)

0.5 0.5
u1 to y~1

u2 to y~1
Im (z)

Im (z)

0 0
–0.5 –0.5

0.5 0.5
u1 to y~2

u2 to y~2
Im (z)

Im( z)

0 0
–0.5 –0.5

0.5 0.5
u1 to z~1

u2 to z~1
Im (z)

Im (z)

0 0
–0.5 –0.5

0.5 0.5
u1 to z~2

u2 to z~2
Im (z)

Im (z)

0 0
–0.5 –0.5

–2 –1 0 1 2 –2 –1 0 1 2
Re (z) Re (z)

Plant Pole Model Pole (S = 10 I2 )

Plant Zero Model Zero (S = 10 I2 )
Model Pole (S = 0.1 I2 ) Model Pole (S = 1000 I2 )
Model Zero (S = 0.1 I2 ) Model Zero (S = 1000 I2 )

FIGURE 19 Example 7: Command following for the continuous-time unstable multiple-input, multiple-output plant (74) with the nonmini-
mum-phase (NMP) transmission zero and NMP channel zero for nt = 2 and various values of S using the three-step command (75). For
each S, the tracking output approaches the three-step command. Note that, as S increases, the output-constraint violation becomes
less severe. Due to the poor initial model, the output constraint is relaxed by the slack f k until t = 20 s. For all t $ 20 s, note that, for
each S, as the model becomes more accurate and zu 1 (t) approaches zero, the output constraint is enforced. The eight bottom-most plots
compare the poles and zeros of the identified model at t = 60 s to the poles and zeros of the plant.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 91

Example 10: Abrupt Change From an Asymptotically
xo = ; E x + ; E u, (78)
-0.1 -0.6 4
Stable Plant to an Unstable Plant With Sensor Noise 0.5 0 0
and Output Constraints y = 60 2@ x, (79)
The goal of this example is to demonstrate PCAC with VRF
with output constraints. Consider a continuous-time asymp- and a realization of (63) is given by (64) and (65).
totically stable SISO plant The data are sampled with sample period Ts = 1s. Let
4 u min = -2, u max = 2, Tu min = -1, Tu max = 1, , = 20, Q r = 10I , - 1,
G (s) = 2 (77)
s + 0.1s + 0.3 rP = 20, R = I ,, P0 = 10 3 I 2nt , and i 0 = 60 1 # (2nt - 1) 1@ T. The initial
that abruptly changes at time 300 s to a continuous-time un- plant (77) is initialized with x(0) = 62.5 - 1.4@ T, and the modi-
stable SISO plant (63) with g = 0.01. Note that the order of (63) fied plant (63) is initialized with x(0) = 6-0.039 1.508 0 0@ T.
is n = 4, and the order of (77) is n = 2. A realization of (77) is The output constraint is given by (5), where C = [1 -1]T , and
given by D = 6-20 -20@ T. For this example, PCAC uses (41)–(45) with

50 rk
0
50

yk
rk –50 kc = 30
0 –100 kc = 200
kc = 400
yk

–50 50
–100 kc = 30 kc = 600
kc = 200 uk 0
50 kc = 400
kc = 600 –50
0
log ∆IR log yk – rk
uk

2
0
–50 –2
4 –4
log ∆IR log yk – rk

2 –6
0 10
–2
–4 5
–6 0
10
5 1
Jk (Θk+1)

0 0.5
0
20 1
Jk (Θk+1)

10
λk

0.8
0
0 100 200 300 400 500 600 0 100 200 300 400 500 600 700
k (Step) k (Step)
k = kc – 1 k = 650 k = kc – 1 k = 700
0.5 0.5
0.5 0.5
Im (z )

Im (z )

Im (z )

Im (z )

0 0 0 0
–0.5 –0.5 –0.5 –0.5
–0.5 0 0.5 0.2 0.4 0.6 0.8 1 1.2 1.4 –0.5 0 0.5 0.2 0.4 0.6 0.8 1 1.2 1.4
Re (z) Re (z ) Re (z ) Re (z )

Plant Pole Plant Zero Model Pole Model Zero Unit Circle
(a) (b)

FIGURE 20 Example 8: Command following with dynamics that abruptly change at step k c represented by the dashed vertical line for
rk / 1 and various values of k c. At step k c, the discrete-time asymptotically stable minimum-phase single-input, single-output plant (46)
changes to the discrete-time unstable nonminimum-phase plant (53). (a) Constant-rate forgetting with m = 1. Note that, as for larger
values of k c, the command-following error after the change is larger. In particular, for k c = 600, the slow reidentification after the abrupt
change causes the control input to saturate and the output error to diverge. Note that the maximum value of the recursive least squares
(RLS) cost J k (i k + 1) (12) increases as k c increases. This is due to the fact that m = 1, and, thus, RLS does not discount data from before
the abrupt change. The two bottom-most plots compare the poles and zeros of the identified model with the poles and zeros of the plant
at k = k c - 1 and k = 650. (b) Variable-rate forgetting (VRF) (21) with h = 0.9, x n = 5, and x d = 10. In contrast with (a), for each k c, y k
approaches the command after the abrupt change. The sixth plot from the top shows that the VRF factor m k is activated near k c. TIR:
the L 2-norm difference between the 30-step impulse response of the plant and model.

92 IEEE CONTROL SYSTEMS » DECEMBER 2021

S = 10 6 I 2 . A strictly proper model is used for identification in tion of the model facilitates identification and, thus, con-
PCAC. Figure 22 shows the response of PCAC for two different straint enforcement. G
levels of sensor noise and various values of nt using the com-
mand profile (76) and VRF (21) with h = 0.5, x n = 25, and CONCLUSIONS
x d = 125. This article presented a numerical investigation of PCAC,
Figure 22 shows that, for a plant that abruptly transi- which uses closed-loop identification to identify a model for
tions from (77) to (63) with g = 0.01, output transients and constrained receding-horizon control optimization. The iden-
constraint violations occur after the abrupt transition. tified model uses an input–output structure with BOCF realiza-
However, as the identification accuracy improves, the out- tion, which enables output-feedback control without the need
put transients dissipate, and the constraints are satisfied for an observer. PCAC was applied to 10 numerical examples
asymptotically. Additionally, note that overparameteriza- that demonstrate command following and disturbance

50 50
rk rk
0 0
–50 –50
yk

yk
tn = 1 tn = 1
–100 tn = 5 –100 tn = 5
tn = 20 tn = 15
50 tn = 25 50 tn = 20
0 0
uk

uk
–50 –50
log ∆IR log yk – rk

Jk (Θk+1) log ∆IR log yk – rk

6 6
4 4
2 2
0 0
–2 –2
–4 –4

10 10
5 5
0 0
15 15
Jk (Θk+1)

10 10
5 5
0 0
1 1
0.8 0.8
λk

λk

0.6 0.6
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
k (Step) k (Step)
k = kc – 1 k = 700 k = kc – 1 k = 700
0.5 0.5 0.5 0.5
Im (z )

Im (z )

Im (z )

Im (z )

0 0 0 0
–0.5 –0.5 –0.5 –0.5
–0.5 0 0.5 0.2 0.4 0.6 0.8 1 1.2 1.4 –0.5 0 0.5 0.2 0.4 0.6 0.8 1 1.2 1.4
Re (z ) Re (z ) Re (z ) Re (z )

Plant Pole Plant Zero Model Pole Model Zero Unit Circle

(a) (b)

FIGURE 21 Example 9: Command following with dynamics that abruptly change from (46) to (53) at step k c = 300 with sensor noise.
Predictive cost adaptive control uses the command profile (76) and variable-rate forgetting (VRF) (21) with h = 0.5, x d = 5x n, and various
values of x n. (a) The sensor noise v k is a zero-mean, Gaussian white noise sequence with a standard deviation of 0.1. Note that, before
the abrupt change and for all values of x n, y k approaches each step command. After the abrupt change, however, y k approaches the
command for x n = 1, x n = 5, and x n = 20, and it diverges for x n = 25. Furthermore, for x n = 25, the impulse-response error and recur-
sive least squares (RLS) cost diverge approximately at k = 430. Note that, as x n increases, the VRF factor m k becomes less sensitive
to sensor noise and is activated only at the abrupt change and at each step-command change. (b) The sensor noise v k is a zero-mean,
Gaussian white noise sequence with a standard deviation of 0.3. Note that, compared to (a), since the sensor-noise level is larger, the
value of x n for which the output, impulse-response error, and RLS cost diverge is lower; namely, x n = 20. TIR: the L 2-norm difference
between the 30-step impulse response of the plant and model.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 93

rejection subject to state and control constraints. Based on totically unbiased; and exigency prioritizes the identification
these examples, “Guidelines for Selecting Hyperparame- of model features that are essential for achieving the control
ters” presents guidelines for selecting hyperparameters. objective. When the persistency of the commands and dis-
The performance of PCAC was investigated numerically turbances is not sufficient for identification, PCAC was
to highlight the interplay between closed-loop identifica- shown to provide additional self-generated persistency
tion and control. Three issues were considered, namely, per- through the control input.
sistency, consistency, and exigency. Persistency is needed Since the control input produced by PCAC is correlated with
for the unambiguous estimation of model parameters; con- the disturbance and sensor noise, the parameter estimates ob-
sistency ensures that the parameter estimates are asymp- tained from closed-loop identification based on RLS are not

400
200 rk 1000 rk
0
y(t)

y(t)
–200 0
"

n=2

"
n=2
" "

n=4

" "
n=4
–400 –1000
"

n=3

"
n=5 n=3 n=5
2 2
0 0
uk

uk
–2 –2
Jk (Θk+1) log ∆IR log yk – rk

Jk (Θk+1) log ∆IR log yk – rk

2 2
0 0
–2 –2
–4 –4
8 8
6 6
4 4
2 2
0 0
–2 –2
100 10
50 5
0 0
1 1
0.8 0.8
λk

λk

0.6 0.6
0 100 200 300 400 500 600 0 100 200 300 400 500 600
t (s) t (s)
t = (tc – 1) s t = 600 s t = (tc – 1) s t = 600 s
1 1 1 1
Im (z)

Im (z)

Im (z)

Im (z)

0 0 0 0

–1 –1 –1 –1
–1 0 1 –2 –1 0 1 –1 0 1 –2 –1 0 1
Re (z) Re (z) Re (z) Re (z)

Plant Pole Plant Zero Model Pole Model Zero Unit Circle

(a) (b)

FIGURE 22 Example 10: Command following with dynamics that abruptly change from (77) to (63) at 300 s with sensor noise and an
output constraint for various values of nt . Predictive cost adaptive control uses the command profile (76) and variable-rate forgetting (21)
with h = 0.5, x n = 25, and x d = 125. (a) The sensor noise v k is a zero-mean, Gaussian white noise sequence with a standard deviation
of 0.01. Note that, before the abrupt change occurs and for each value of nt , y (t) approaches the step commands, and the output con-
straint is satisfied at all times. In the underparameterized cases nt 1 n = 4, after the abrupt change occurs, the double integrator and
lightly damped poles are poorly identified by recursive least squares (RLS), causing y(t) to oscillate and violate the constraint until the
end of the simulation. In contrast, in the exact and overparameterized cases nt $ n = 4 , after the abrupt change occurs, y(t) only tempo-
rarily violates the constraint during reidentification Additionally, after t = 338 s, y(t) is more damped than the underparameterized cases,
and the output constraint is enforced at all times. Note that the last segment of y(t) is infeasible. (b) The sensor noise v k is a zero-mean,
Gaussian white noise sequence with a standard deviation of 0.1. Note that, due to the higher sensor-noise level, the oscillation amplitude
and output-constraint violation for nt 1 n = 4 are more pronounced in (b) than in (a). Furthermore, note that, in both (a) and (b), for nt = 2,
RLS is able to capture the two integrators of the plant, whereas, for nt = 3, the three poles of the model are unable to capture the two
integrators and two undamped poles of the plant, causing the identification and transient response to be poorer. TIR: the L 2-norm dif-
ference between the 30-step impulse response of the plant and model.

94 IEEE CONTROL SYSTEMS » DECEMBER 2021

consistent and, thus, are asymptotically biased. Although the Michigan, Ann Arbor, Michigan, USA. His research interests
bias was shown to degrade the accuracy of PCAC, this effect include control theory for systems with state and control con-
was mitigated to some extent by using RLS with VRF. PEM and straints as well as control applications for aerospace and auto-
IV were considered as alternatives to RLS/VRF. Although PEM motive systems. He received the Ph.D. degree from the Univer-
and IV are more accurate than RLS/VRF for both open- and sity of Michigan in 1995. He is a Fellow of IEEE and is named
closed-loop identification, these methods were found to be less as an inventor on more than 100 U.S. patents.
compatible with PCAC than RLS/VRF.
Finally, the interplay between control and identification
REFERENCES
was further examined in terms of exigency, which is the abil- [1] W. Kwon and S. Han, Receding Horizon Control: Model Predictive Control for State
ity of the closed-loop identification to prioritize the model Models. Berlin: Springer-Verlag, 2006.
[2] S. Di Cairano and I. V. Kolmanovsky, “Real-time optimization and model pre-
features needed to achieve the control objective. Numerical
dictive control for aerospace and automotive applications,” in Proc. Amer. Contr.
examples, such as the estimation of the plant dc gain for set- Conf., 2018, pp. 2392–2409. doi: 10.23919/ACC.2018.8431585.
point command following, showed that exigency is manifest- [3] U. Eren, A. Prach, B. B. Koçer, S. V. Rakovic´, E. Kayacan, and B. Aç ikmes¸e,
“Model predictive control in aerospace systems: Current state and opportuni-
ed in the continuing refinement of the identified model.
ties,” J. Guidance, Control, Dynamics, vol. 40, no. 7, pp. 1541–1566, 2017. doi:
The numerical examples in this article were designed to ex- 10.2514/1.G002507.
pose features of PCAC that motivate the investigation of stabil- [4] T. Samad et al., “Industry engagement with control research: Perspective and mes-
sages,” Ann. Rev. Contr., vol. 49, pp. 1–14, 2020. doi: 10.1016/j.arcontrol.2020.03.002.
ity and performance guarantees. A major next step is to extend
[5] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Cam-
the method to various classes of nonlinear systems [74]. Finally, bridge Univ. Press, 2004. doi: 10.1017/CBO9780511804441.
the performance of PCAC on linear and nonlinear numerical [6] S. S. Keerthi and E. G. Gilbert, “Optimal infinite-horizon feedback laws for a
general class of constrained discrete-time systems: Stability and moving-horizon
examples provides further motivation to apply this technique
approximations,” J. Optim. Theory Appl., vol. 57, no. 2, pp. 265–293, 1988. doi:
to physical examples. 10.1007/BF00938540.
[7] E. F. Camacho and C. Bordons, Model Predictive Control, 2nd ed. Berlin: Spring-
er-Verlag, 2007.
ACKNOWLEDGMENTS [8] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive Control: Theory,
We thank Nima Mohseni, Adam Bruce, and the reviewers for Computation, and Design, 2nd ed. Nob Hill Publishing, LLC, 2017.
numerous helpful comments and suggestions. This research [9] S. V. Rakovic´ and W. S. Levine, Handbook of Model Predictive Control. Birkhäus-
er, 2019.
was supported by the U.S. Office of Naval Research under BRC
[10] K. Fruzzetti, A. Palazog˘lu, and K. McDonald, “Nonlinear model predictive
grant N00014-18-1-2211. control using Hammerstein models,” J. Process Control, vol. 7, no. 1, pp. 31–41,
1997. doi: 10.1016/S0959-1524(97)80001-B.
[11] H. H. J. Bloemen, T. J. J. Van Den Boom, and H. B. Verbruggen, “Model-based
AUTHOR INFORMATION predictive control for Hammerstein–Wiener systems,” Int. J. Control, vol. 74, no. 5,
Tam W. Nguyen received the M.Sc. degree in mechatronics en- pp. 482–495, 2001. doi: 10.1080/00207170010014061.
gineering and the Ph.D. degree in control systems from the Free [12] J. Zhang, K.-S. Chin, and M. Ławryn´czuk, “Nonlinear model predictive con-
trol based on piecewise linear Hammerstein models,” Nonlinear Dynam., vol. 92,
University of Brussels, Belgium. He then served as a postdoc-
no. 3, pp. 1001–1021, 2018. doi: 10.1007/s11071-018-4105-5.
toral researcher in the Aerospace Engineering Department at [13] T. Hu and Z. Lin, Control Systems With Actuator Saturation: Analysis and De-
the University of Michigan, Ann Arbor. He is currently an as- sign. Birkhäuser, 2001. doi: 10.1007/978-1-4612-0205-9.
[14] S. Tarbouriech, G. Garcia, J. M. G. da Silva, and I. Queinnec, Stability and Sta-
sistant professor in the Department of Electrical and Electron-
bilization of Linear Systems With Saturating Actuators. Berlin: Springer-Verlag, 2011.
ics Engineering at the University of Toyama, Toyama, Japan. doi: 10.1007/978-0-85729-941-3.
His research interests include constrained control for aerial- [15] D. S. Bernstein and A. N. Michel, “A chronological bibliography on satu-
rating actuators,” Int. J. Robust Nonlinear Control, vol. 5, no. 5, pp. 375–380, doi:
robotics applications.
10.1002/rnc.4590050502.
Syed Aseem Ul Islam received the B.Sc. degree in aerospace [16] D. Limón, I. Alvarado, T. Alamo, and E. F. Camacho, “MPC for tracking
engineering from the Institute of Space Technology, Islamabad, piecewise constant references for constrained linear systems,” Automatica, vol. 44,
no. 9, pp. 2382–2387, 2008. doi: 10.1016/j.automatica.2008.01.023.
and the Ph.D. degree in flight dynamics and control from the
[17] G. C. Goodwin, D. S. Carrasco, D. Q. Mayne, M. E. Salgado, and M. M.
University of Michigan, Ann Arbor, Michigan, USA. He is cur- Serón, “Preview and feedforward in model predictive control: Conceptual
rently a research fellow at the University of Michigan, Ann Ar- and design issues,” IFAC Proc. Volumes, vol. 44, no. 1, pp. 5555–5560, 2011. doi:
10.3182/20110828-6-IT-1002.02298.
bor, Michigan, USA. His research interests include data-driven
[18] S. M. Kuo and D. Morgan, Active Noise Control Systems: Algorithms and DSP
adaptive control for aerospace applications. Implementations. Hoboken, NJ: Wiley, 1995.
Dennis S. Bernstein received the Sc.B. degree from Brown [19] L. Grüne and J. Pannek, Nonlinear Model Predictive Control. Berlin: Springer-
Verlag, 2017, pp. 45–99.
University and the Ph.D. degree from the University of Michi-
[20] N. L. Ricker, “Model predictive control with state estimation,” Ind. engineering
gan in Ann Arbor, Michigan. He is currently a professor in the chemistry Res., vol. 29, no. 3, pp. 374–382, 1990. doi: 10.1021/ie00099a013.
Aerospace Engineering Department, University of Michigan, [21] A. Wills, D. Bates, A. Fleming, B. Ninness, and R. Moheimani, “Application of
MPC to an active structure using sampling rates up to 25kHz,” in Proc. Conf. Dec.
Ann Arbor, Michigan, USA. His research interests include iden-
Contr., 2005, pp. 3176–3181. doi: 10.1109/CDC.2005.1582650.
tification, estimation, and control for aerospace applications. He [22] I. V. Kolmanovsky and V. Winstead, “A receding horizon optimal control ap-
is the author of Scalar, Vector, and Matrix Mathematics, published proach to active state and parameter estimation in automotive systems,” in Proc.
IEEE Int. Conf. Control Appl., 2006, pp. 2796–2801.
by Princeton University Press.
[23] K. Zhang and Y. Shi, “Adaptive model predictive control for a class of con-
Ilya V. Kolmanovsky ([email protected]) is a professor in strained linear systems with parametric uncertainties,” Automatica, vol. 117, p.
the Department of Aerospace Engineering at the University of 108974, 2020. doi: 10.1016/j.automatica.2020.10897.

DECEMBER 2021 « IEEE CONTROL SYSTEMS 95

[24] V. Adetola, D. DeHaan, and M. Guay, “Adaptive model predictive control [50] K. F. Aljanaideh and D. S. Bernstein, “Closed-loop identification of unstable
for constrained nonlinear systems,” Syst. Control Lett., vol. 58, no. 5, pp. 320–326, systems using noncausal FIR models,” Int. J. Contr, vol. 90, no. 2, pp. 168–185,
2009. doi: 10.1016/j.sysconle.2008.12.002. 2017. doi: 10.1080/00207179.2016.1172733.
[25] D. W. Clarke, C. Mohtadi, and P. S. Tuffs, “Generalized predictive control [51] F. Sobolic, K. F. Aljanaideh, and D. S. Bernstein, “A numerical investi-
— Part I. The basic algorithm,” Automatica, vol. 23, no. 2, pp. 137–148, 1987. doi: gation of direct and indirect closed-loop architectures for estimating non-
10.1016/0005-1098(87)90087-2. minimum-phase zeros,” Int. J. Contr, vol. 93, no. 6, pp. 1251–1265, 2020. doi:
[26] Y. I. Lee and B. Kouvaritakis, “Receding horizon output feedback control for 10.1080/00207179.2018.1501609.
linear systems with input saturation,” in Proc. Conf. Dec. Contr., 2000, pp. 656–661. [52] L. Ljung, “Prediction error estimation methods,” Circ., Sys. Sig. Proc, vol. 21,
doi: 10.1109/CDC.2000.912841. no. 1, pp. 11–21, 2002. doi: 10.1007/BF01211648.
[27] N. Saraf and A. Bemporad, “Fast model predictive control based on linear [53] M. Gilson and P. V. den Hof, “Instrumental variable methods for closed-loop
input/output models and bounded-variable least squares,” in Proc. Conf. Dec. system identification,” Automatica, vol. 41, no. 2, pp. 241–249, 2005. doi: 10.1016/j.
Contr., 2017, pp. 1919–1924. doi: 10.1109/CDC.2017.8263930. automatica.2004.09.016.
[28] J. Barlow and M. Q. Phan, “Direct data-based model predictive controller [54] J. I. Yuz and G. C. Goodwin, Sampled-Data Models for Linear and Nonlinear
design,” in Proc. AIAA Guidance, Navigation, and Control Conf. and Exhibit, 2006, p. Systems. Berlin: Springer-Verlag, 2014. doi: 10.1007/978-1-4471-5562-1.
6421. doi: 10.2514/6.2006-6421. [55] K. J. Åström and B. Wittenmark, Computer-Controlled Systems, 3rd ed. Engle-
[29] P. J. Goulart and E. C. Kerrigan, “Output feedback receding horizon con- wood Cliffs, NJ: Prentice-Hall, 1997.
trol of constrained systems,” Int. J. Control, vol. 80, no. 1, pp. 8–20, 2007. doi: [56] “Rigid tube model predictive control for linear sampled–data systems,” IFAC-
10.1080/00207170600892949. PapersOnLine, vol. 50, no. 1, pp. 9840–9845, 2017. doi: 10.1016/j.ifacol.2017.08.903.
[30] C. E. Garcia and A. Morshedi, “Quadratic programming solution of dynamic [57] F. Fontes, “Discontinuous feedbacks, discontinuous optimal controls, and
matrix control (QDMC), Chem. Eng. Commun., vol. 46, nos. 1–3, pp. 73–87, 1986. continuous-time model predictive control,” Int. J. Robust Nonlinear Control, vol.
doi: 10.1080/00986448608911397. 13, nos. 3–4, pp. 191–209. doi: 10.1002/rnc.813.
[31] S. Rakovic´, “Robust model-predictive control,” in Encyclopedia of Systems and [58] R. Findeisen, L. Imsland, F. Allgöwer, and B. Foss, “Towards a sampled-data
Control, J. Baillieul and T. Samad, Eds. Berlin: Springer-Verlag, 2013, pp. 1–11. doi: theory for nonlinear model predictive control,” in New Trends in Nonlinear Dy-
10.1007/978-1-4471-5102-9 2-1. namics and Control and their Applications, W. Kang, C. Borges, and M. Xiao, Eds.
[32] M. Cannon and B. Kouvaritakis, Model Predictive Control. Berlin: Springer- Berlin, Heidelberg: Springer-Verlag, 2003, pp. 295–311. doi: 10.1007/978-3-540-
Verlag, 2016. doi: 10.1007/978-3-319-24853-0. 45056-6_19.
[33] J. Lofberg, “Approximations of closed-loop minimax MPC,” in Proc. Conf. [59] J. W. Polderman, “A state space approach to the problem of adaptive pole
Dec. Contr., 2003, pp. 1438–1442. doi: 10.1109/CDC.2003.1272813. assignment,” Mathematics Control, Signals Syst., vol. 2, no. 1, pp. 71–94, 1989. doi:
[34] W. Langson, I. Chryssochoos, S. Rakovic´, and D. Mayne, “Robust model 10.1007/BF02551362.
predictive control using tubes,” Automatica, vol. 40, no. 1, pp. 125– 133, 2004. doi: [60] K. J. Åström and B. Wittenmark, Adaptive Control, 2nd ed. Reading, MA:
10.1016/j.automatica. 2003.08.009. Addison-Wesley, 1994.
[35] F. A. Fontes, S. Rakovic´, and I. V. Kolmanovsky, “Rigid tube model predic- [61] S. A. U. Islam and D. S. Bernstein, “Recursive least squares for real-time
tive control for linear sampled–data systems,” IFAC-PapersOnLine, vol. 50, no. 1, implementation,” IEEE Control Syst. Mag., vol. 39, no. 3, pp. 82–85, 2019. doi:
pp. 9840–9845, 2017. doi: 10.1016/j.ifacol.2017.08.903. 10.1109/MCS.2019.2900788.doi:
[36] M. Shouche, H. Genceli, P. Vuthandam, and M. Nikolaou, “Simultaneous [62] A. L. Bruce, A. Goel, and D. S. Bernstein, “Convergence and consistency of
constrained model predictive control and identification of DARX processes,” Au- recursive least squares with variable-rate forgetting,” Automatica, vol. 119, pp.
tomatica, vol. 34, no. 12, pp. 1521–1530, 1998. doi: 10.1016/S0005-1098(98)80005-8. 109–152, 2020. doi: 10.1016/j.automatica.2020.109052.
[37] M. Lorenzen, M. Cannon, and F. Allgöwer, “Robust MPC with recursive [63] A. Goel, A. Bruce, and D. S. Bernstein, “Recursive least squares with variable-
model update,” Automatica, vol. 103, pp. 461–471, 2019. doi: 10.1016/j.automat- direction forgetting: Compensating for the loss of persistency,” IEEE Contr. Sys,
ica.2019.02.023. vol. 40, pp. 80–102, Aug. 2020.
[38] A. Weiss and S. Di Cairano, “Robust dual control MPC with guaranteed con- [64] P. M. J. V. den Hof and R. J. P. Schrama, “Identification and control — Closed-
straint satisfaction,” in Proc. Conf. Dec. Contr., 2014, pp. 6713–6718. doi: 10.1109/ loop issues,” Automatica, vol. 31, no. 12, pp. 1751–1770, 1995. doi: 10.1016/0005-
CDC.2014.7040443. 1098(95)00094-X.
[39] R. Soloperto, J. Koehler, M. Müller, and F. Allgöwer, “Dual adaptive MPC for [65] M. Gevers, “Towards a joint design of identification and control?” Progr. Sys.
output tracking of linear systems,” in Proc. Conf. Dec. Contr., 2019, pp. 1377–1382. Contr. Theory, vol. 14, pp. 111–151, 1993. doi: 10.1007/978-1-4612-0313-1_5.
doi: 10.1109/CDC40024.2019.9029693. [66] H. Hjalmarsson, “From experiment design to closed-loop control,”
[40] P. Ioannou and B. Fidan, Adaptive Control Tutorial. SIAM, 2006. doi: Automatica, vol. 41, no. 3, pp. 393–438, 2005. doi: 10.1016/j.automatica
10.1137/1.9780898718652. .2004.11.021.
[41] G. Chowdhary, M. Muhlegg, and E. N. Johnson, “Exponential parameter and [67] A. A. Feldbaum, “Dual control theory,” Avtomatika i Telemekhanika, vol. 22, no.
tracking error convergence guarantees for adaptive controllers without persis- 1, pp. 3–16, 1961. [Online]. Available: http://mi.mathnet.ru/at12149
tency of excitation,” Int. J. Contr., vol. 87, no. 8, pp. 1583–1603, May 2014. doi: [68] B. Wittenmark, “Adaptive dual control methods: An overview,” Adap-
10.1080/00207179.2014.880128. tive Syst. Control Signal Process., vol. 28, no. 13, pp. 67–72, 1995. doi: 10.1016/
[42] M. Campi, A. Lecchini, and S. Savaresi, “Virtual reference feedback tuning: S1474-6670(17)45327-4.
direct method for the design of feedback controllers,” Automat.a, vol. 38, no. 8, pp. [69] N. M. Filatov and H. Unbehauen, Adaptive Dual Control. Berlin: Spring-
1337–1346, 2002. doi: 10.1016/S0005-1098(02)00032-8. er-Verlag, 2004.
[43] D. Xu, B. Jiang, and P. Shi, “Adaptive observer-based data-driven control for [70] K. B. Ariyur and M. Krstic, Real-Time Optimization by Extremum-Seeking
nonlinear discrete-time processes,” IEEE Trans. Auto. Sci. Eng, vol. 11, no. 4, pp. Control. Hoboken, NJ: Wiley, 2003.
1037–1045, 2014. doi: 10.1109/TASE.2013.2284062. [71] T. W. Nguyen, S. A. U. Islam, A. L. Bruce, A. Goel, D. S. Bernstein,
[44] T. Maupong and P. Rapisarda, “Data-driven control: A behavioral approach,” and I. V. Kolmanovsky, “Output-feedback RLS-based model predictive
Sys. Contr. Lett., vol. 101, pp. 37–43, 2017. 10.1016/j.sysconle.2016.04.006. control,” in Proc. Amer. Contr. Conf., 2020, pp. 2395–2400. doi: 10.23919/
[45] M. Tanaskovic, L. Fagiano, C. Novara, and M. Morari, “Data-driven control ACC45564.2020.9148011.
of nonlinear systems: An on-line direct approach,” Automatica, vol. 75, pp. 1–10, [72] N. Mohseni, T. W. Nguyen, S. A. U. Islam, I. V. Kolmanovsky, and D. S.
2017. 10.1016/j.automatica.2016.09.032. Bernstein, “Active noise control for harmonic and broadband disturbances
[46] D. Piga, S. Formentin, and A. Bemporad, “Direct data-driven control of con- using RLS-based model predictive control,” in Proc. Amer. Contr. Conf., 2020,
strained systems,” IEEE Trans. Contr. Sys. Tech., vol. 26, no. 4, pp. 1422–1429, 2018. pp. 1393–1398. doi: 10.23919/ACC45564.2020.9147440.
10.1109/TCST.2017.2702118. [73] M. Vidyasagar, Control System Synthesis: A Factorization Approach. MIT,
[47] C. D. Persis and P. Tesi, “On persistency of excitation and formulas for data- 1985. doi: 10.2200/S00351ED1V01Y201105CRM002.
driven control,” in Proc. Conf. Dec. Contr., 2019, pp. 873–878. [74] T. W. Nguyen, I. V. Kolmanovsky, and D. S. Bernstein, “Sampled-data
[48] C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization, op- output-feedback model predictive control of nonlinear plants using online
timality, and robustness,” IEEE Trans. Auto. Contr, vol. 65, no. 3, pp. 909–924, 2020. linear system identification,” in Proc. Amer. Contr. Conf., New Orleans, LA,
doi: 10.1109/TAC.2019.2959924. 2021, pp. 4682–4687.
[49] U. Forssell and L. Ljung, “Closed-loop identification revisited,” Automat.a, 
vol. 35, no. 7, pp. 1215–1241, 1999. doi: 10.1016/S0005-1098(99)00022-9.

96 IEEE CONTROL SYSTEMS » DECEMBER 2021