A Brief Introduction to the Analysis and Design of Networked Control

Systems
Graham C. Goodwin, Eduardo I. Silva, Daniel E. Quevedo
School of Electrical Engineering and Computer Science, The University of Newcastle, NSW 2308, Australia
E-mail: [email protected]

Abstract: Networked Control has emerged in recent years as a new and exciting area in systems science. The topic has
many potential applications in diverse areas ranging from control of microrobots to biological and economic systems. The
supporting theory is very rich and combines aspects of control, signal processing, telecommunications and information
theory. In this paper, we will give a brief overview of recent developments in Networked Control with an emphasis on
our contributions. We also point to several open problems in this emerging area.
Key Words: Networked Control, Performance, Data-Rate Constraints, Delays, Data Dropouts

Introduction

Traditionally, control theory has dealt with situations where

the communication links between plant and controller can
be regarded as transparent (see, e.g., [13]). There exist,
however, cases where the links in a control system are far
from being transparent and may become bottlenecks in the
achievable performance. Control systems where this happens are collectively referred to as Networked Control Systems (NCSs).
The study of NCSs has emerged as an active research field
during the past years (see, e.g., the special issues [4, 5]).
Key questions within this framework are related to the way
in which network artifacts affect the stability and performance of control loops that employ non-transparent communication channels. Typical channel artifacts include
data-rate limits, random delays and data dropouts. A unifying framework for the treatment of general NCS analysis and design problems does not yet exist. Nevertheless,
there has been significant progress in the study of several
subproblems. For example, data-rate constraints have been
studied in, e.g., [69]. The issue of data dropouts has been
studied in, e.g., [1012] and random time delays have been
considered in, e.g., [13, 14].
A pivotal issue in any closed loop system is that of stability. When focusing on quantization issues, a key result
establishes necessary and sufficient conditions on the channel data-rate that allows one to achieve closed loop stability (in an appropriate sense; see, e.g., [7] and the many
references therein). These results are given in terms of a
lower bound on the channel data-rate (that depends on the
unstable plant poles only) over which control and coding
schemes can be constructed so as to achieve stability. These
coding schemes are quite involved and may not be attractive from a practical point of view. This fact has motivated
us to study simple coding schemes that achieve rates close
to the bounds identified above (see [15]). Recent research
has also established an important link between the average

data-rate necessary to achieve stability and signal-to-noise

ratio requirements (see [15]). This observation reinforces
the use of signal-to-noise ratio models in the context of
NCS performance analysis (see, e.g., [16]).
For SISO LTI plant models, one can design coding schemes
that, for a given initial non-networked controller design,
allow one to minimize the impact of quantization effects on
overall closed loop performance. These results emphasize
the essential role that the available degrees of freedom play
in NCSs (see [1618] and also [1923]).
The issue of performance is particularly interesting in the
context of networked control architectures for MIMO plant
models. Practical control systems for MIMO plant models
often use structurally constrained controllers such as diagonal or triangular ones (see, e.g., [1, 24]). The reasons for
this choice are manifold and include ease of design, simplified tuning, and implementation related issues such as
cabling or geographic plant distribution. It is well known
that restricting the controller architecture generally limits
the achievable performance (see, e.g., [2527]). Within
this context, networks can play significative roles. Indeed, it is easy to envisage decentralized control architectures that, when enriched with additional (non-transparent)
communication links, may provide enhanced performance.
This may (partially) overcome the limitations that arise as
a consequence of the controller structure constraint (see,
e.g., [2832]). To illustrate these ideas, in the present paper we will examine the control of MIMO LTI systems for
which a decentralized controller has been successfully designed. Unsurprisingly, for high-quality channels it turns
out that a networked MIMO architecture outperforms decentralized ones. However, an interesting finding is that,
in some situations, the networked architecture will perform
better than the decentralized one only if the channels are
extremely reliable (see [33]).
We finally address the problem of dealing with arbitrary
data dropouts and delays. We show how control laws de-

1
c 2008 IEEE
978-1-4244-1734-6/08/$25.00 

d
r

C(z)

G(z)

D
decoder

channel

encoder

An overarching issue in any closed loop system is the potential for instability. For this reason, obtaining necessary
and sufficient conditions which guarantee closed loop stability has become a central issue in control theory. NCSs
give rise to new challenges with respect to stability. This
section will give some insights into the stability question.
Consider the networked control situation shown in Figure 1. In that figure, G(z) is a SISO LTI plant, C(z) is
a LTI controller, r is a reference signal1 and d is a disturbance. Unlike standard non-networked control systems
(see, e.g., [13]) the feedback path in the control system in
Figure 1 comprises a non-transparent channel and a coding
scheme. The coding scheme is a novel aspect of networked
control that has no equivalent in traditional control theory.
The coding scheme has two parts: the encoder and the decoder. The encoder is in charge of appropriately processing the signal to be sent over the channel so as to compensate, if possible, channel characteristics and/or to translate
the measurements into symbols that the channel can understand. (This is the case of, e.g., digital channels where the
symbols are binary words.) On the other hand, the decoder
is in charge of translating back the channel symbols into
the standard signal domain.
There exist many channel models (see, e.g., [41, 42]). To
highlight the ideas, in the present section we will consider
error- and delay-free digital channels, i.e., channels that can
transmit a countable set of symbols (countable alphabet)
without errors or delays. Of course, in practice, channels
always have restricted bandwidth and the channel alphabet
is, thus, finite (not just countable; see also [7, 43, 44]).
1 All

signals in this paper are assumed wise sense stationary (wss) stationary processes.

F (z)

Figure 2: Proposed coding scheme for stabilization.

2.1

Stability

Figure 1: Basic Networked control system (including coding scheme).

channel

signed for non-networked control systems can be embellished so as to achieve good performance when used in
an NCS that employs unreliable communication channels
(see [34] and related work [3540]).
The remainder of this paper is organized as follows: Section 2 discusses the problem of stabilization of NCSs subject to data rate constraints. Section 3 discusses the design
of coding systems that optimize NCS performance. Section 4 studies some aspects in MIMO networked control.
Section 5 studies control over unreliable networks. Section
6 draws conclusions.

Key results

For the setting described above, a key result states that it

is possible to find coders, decoder and controllers such that
the resulting system is mean square stable (MSS) if and
satonly if the average data-rate in bits per sample, say R,
isfies (see [6])
>
R

np


log2 |pi | ,

(1)

i=1

where {pi }i=1, ,np denotes the set of unstable plant poles.
The above result is valid for every coder, decoder and controller in the class of (time varying and non-linear) causal
systems. Therefore, (1) establishes a fundamental separation between what is achievable in NCSs over digital channels and what is not (when the problem of interest is MSS).
We note that bounds similar to (1) arise as solutions to quite
a few different problems (e.g., observability, deterministic
stability, etc.) and under different assumptions on the channels and coding schemes (see, e.g., [6,9,41,4547]). Indeed
the quantity on the right hand side of (1) is a fundamental
measure of the difficulty of manipulating a system, as explored in [8, 48].
Proving that the rate at which a stable control system
is transmitting data must satisfy (1) (i.e., necessity) is
fairly simple and employs standard tools (see also [49
51]), whereas constructing an actual coding scheme that
achieves stability at any rate above the absolute limit (i.e.,
the proof of sufficiency) is much more involved (see [6, 9]).
These observations motivate the remainder of this section.
2.2

Some insights into the problem of stabilization

with data-rate constraints

In this section we shed some light into the limitations that

finite data-rates impose. To fix ideas, we will consider the
coding scheme depicted in Figure 2, where F (z) is a LTI
filter, and E and D are abstract systems that translate the
discrete time signal v into channel symbols s and channel
symbols into the discrete time signal w, respectively. Our
presentation will start at an heuristic level and will then
move towards a formal approach based on information theoretic considerations.
2.2.1

An additive noise model for quantization

Since the channel is digital, it becomes clear that E must

quantize v prior to transmission. We will focus on the most
simple situation where D is the identity and E is a finite

2008 Chinese Control and Decision Conference (CCDC 2008)

uniform quantizer (see, e.g., [52, 53]), i.e.,2

 

v(k)
.
s(k) = Q(v(k))  satV

(2)

In (2), V > 0 is the quantizer dynamic range, 

2V (L 1)1 is the quantization step and L is the number
of quantization levels. We say the the quantizer is overloaded if and only if |v(k)| > V for some k. (If in (2) satV
is removed and is fixed at any arbitrary positive value,
then Q becomes an infinite uniform quantizer.)
Quantization is a deterministic non-linear operation and
hence, the exact analysis of quantized systems is difficult (see, e.g., [43, 44, 54]). It has thus become standard,
particularly in the signal processing literature (see, e.g.,
[52, 55, 56]), to approximate quantization noise, i.e., the
process
q  w v,

(3)

by an additive i.i.d. noise source, uniformly distributed on

rthe interval (
2 , 2 ) and independent of the input to the
3
quantizer v. This model is valid only if is small enough,
the quantizer does not overload and v has a smooth probability density function (see, e.g., [57]). These conditions
usually do not hold in the case of quantizers that are embedded in feedback loops, as is the case of NCSs (see,
e.g., [58]). We also note that assuming that q is either independent of v or uncorrelated with v is, certainly, not a
valid assumption in networked situations. In these cases,
it makes more sense to assume that q is independent of (or
uncorrelated with) the external signals r and d.
In order to avoid quantizer overload, it is usual in practice
to choose a quantizer dynamic range such that the probability of overload is negligible (see, e.g, [52]). Indeed, if v
is wss and is any positive real value, then one can always
find a finite such that4 V = v guarantees that the probability of overload is less than ; is called the quantizer
loading factor.5 With such a choice for the loading factor,
it is immediate to see that, provided q is as in the classical
additive model described above,


3
v2
= 2 (L 1)2 ,
q2

(4)

where is the quantizer signal-to-noise ratio. Thus, bitrate limits translate into signal-to-noise ratio constraints.
It is important to note that the previous model for quantization can actually be rendered exact by a simple randomization procedure. Indeed, it suffices to consider a dithered
uniform quantizer. In this case
 

v(k) + dh (k)
,
s(k) = satV

w(k) = s(k) dh (k),

x
V if |x| > V ; x
(x)  x if |x| V and satV (x)  |x|
denotes the integer part of x.
3 Sometimes q is assumed only to be uncorrelated with v.
4 denotes the standard deviation of v.
v
5 For example, if v(k) is Gaussian, then = 4 gives an overload
probability of 6.33 105 .
2 sat

where dh is an i.i.d. random process, uniformly distributed

on (
2 , 2 ) and such that dh (k) is independent of v(k).
In this situation, and provided no overload occurs, q in (3)
becomes distributed as dh (see, e.g., [53,59,60]). In other
words, q becomes just as in the simplified model described
above. Again, in order to avoid quantizer overload, one
needs to consider an appropriate quantizer loading factor
and a signal-to-noise ratio constraint arises.
In practice, implementation of dithered quantizers is not
trivial since it requires the availability of the dither signal dh at both the sending and receiving ends. Usually
the dither is generated using pseudo-random number generators that are initialized with the same seeds. Nevertheless, even if one employs a non-dithered uniform quantizer with as small as 4 levels, it turns out that the predictions made using the simple additive quantization model
described above are surprisingly accurate (see simulation
studies in [16, 17, 55, 61]).
2.2.2

Limits imposed by signal-to-noise ratio constraints

Consider the setting in Figure 2 and assume that E and D

are such that q obeys the additive model for quantization
described in Section 2.2.1. We note that, since q depends
on the way in which E and D are designed, the variance of
q, say q2 , becomes a decision variable.6 Within this setting
a basic question arises, namely finding the conditions on
the signal-to-noise ratio that guarantee MSS.
It is possible to show via standard control theoretic arguments (see [15]) that, for any strictly proper plant model
G(z) and any wss r and d, there exists C(z), F (z) and a
finite q2 such that the resulting NCS is MSS if and only if
> inf 

 np


|pi |

1,

(5)

i=1

where {p1 , , pnp } is the set of unstable plant poles.

From this result we conclude that, as was the case for (1),
the degree of instability of the plant plays a key role in
the interplay between stability and communications constraints (as expressed by signal-to-noise ratio constraints in
our current framework). It is interesting to note that if one
adds one and takes logarithms in (5), then one recovers the
right hand side in (1). (In [15], we have shown that this
connection is not a mere coincidence.)
It is easy to see from (5) and (4) that, if the additive model
for quantization holds, then MSS is equivalent to
 

np

2
2

|pi | 1 , (6)
b > binf  log2 1 +
3
i=1
where b  log2 L is the number of bits in the quantizer.
In other words, provided no overload occurs and the noise
model for quantization holds, (6) gives a bound on the
instantaneous data-rate over the channel that guarantees
6 For

the additive model described previously q2 =

2008 Chinese Control and Decision Conference (CCDC 2008)

2
.
12

MSS.7 Clearly, this bound on b is optimistic. Indeed, it

is only exact if one employs a dithered uniform quantizer
and there is no overload. But, if r, d or the initial plant state
have unbounded support, then it is impossible to guarantee
that the input to the quantizer is deterministically bounded.
Even if r, d and the plant initial state have bounded support, and thus v is bounded, the previous analysis usually
requires a very large value for the loading factor . Equation (6) then provides only a conservative bound on binf .
In practice it is often sufficient to assume that v is such
that a sensibly small value for gives negligible overload
probability (a typical value is = 4; see [52] and also
Sections 3 and 4).
2.2.3

dh

s+

EC

ED

channel

unif. quantizer

Figure 3: Entropy coded dithered quantizer.

such that the average data-rate is upper bounded by (see

[15]).


1 ln (1 + ) + 1 log2 2e + 1.
R
(7)
2
2
12

Average data-rates

In the previous section we dealt with instantaneous datarates. There are at least two reasons that can be advanced
for abandoning that setting. First, if the external signals
have unbounded support, then it is impossible to give any
guarantee using the above reasoning, unless one employs
quantizers with infinitely many levels so as to avoid overload (thus incurring infinite instantaneous data-rates). Second, if the external signals are bounded, then the quantizer
loading factor may need to be quite large in order to accommodate v without overload. This implies that, even
in those cases, one needs to use a quantizer with either
a large number of bits (which increases the instantaneous
rate), or a large quantization step (which implies poor performance). We thus conclude that, at least from a theoretical point of view, it is interesting to study average datarates.8 (Of course, guaranteeing that average data-rates are
bounded does not ensure that instantaneous rates will be
bounded.)
A popular way to deal with average data rate constraints
is to use an entropy coded dithered quantizer (ECDQ; see,
e.g., [59]). Figure 3 shows the architecture of an ECDQ
and its relationship to E and D in Figure 2. An ECDQ uses
a dithered infinite uniform quantizer, as described earlier,
but instead of sending the quantizer output directly over
the channel, it entropy-codes it prior to transmission. The
basic idea behind the entropy-coder EC in Figure 3 is to
exploit the fact that some of the quantizer output values are
more likely than others. Thus, one can assign short channel symbols (i.e., few bits) to very frequent quantizer output values and long symbols (i.e., many bits) to infrequent
ones (see [42]). Correspondingly, the task of the entropydecoder ED is to convert the channel symbols back into the
exact quantizer output at the receiving end (see [42]).
A key property of ECDQs is that the average data-rate at
which data is transmitted can be related to the quantizer
signal-to-noise ratio . Indeed, it is possible to show that,
if the dither is as in Section 2.2 and is also independent of
r, d and the filters initial states, then the coding scheme
proposed in Figure 2, when E and D form an ECDQ, is
7 Clearly,

binf is in general greater than the right hand side in (1).

8 Alternatively, one could also examine time varying (i.e., adaptive)

quantizers that are scaled on-line according to the statistics of the input
signals. This topic is interesting, but is beyond the scope of this paper
(see references in [7]).

dh

Thus, we see that signal-to-noise ratio considerations give

an upper bound on the average data-rate. In other words,
there exists a strong link between signal-to-noise ratio constraints and average data-rate constraints. This goes beyond
the heuristic discussion in Section 2.2.1, and serves as a basis for a simple but rigorous treatment of average data-rate
constraints in networked control.
In particular, one can use (7) to show that it is possible to
achieve MSS at average data-rates that satisfy (see [15])

np


log2 |pi | +

i=1

1
log2
2

2e
12


+ 1 + F (),

where F is a positive function that goes to zero if .

Therefore, the use of an ECDQ allows one to achieve an average rate that can be made arbitrarily
to the absolute
 close

+ 1 bits per samminimum rate in (1) plus 12 log2 2e
12
ple (i.e., 1.25 bits per sample). This additional rate is composed by two terms: the first one is due to the divergence of
the distribution of quantization noise from Gaussianity, and
the second one originates in the inefficiency of the loss-less
coding scheme employed to generate the channel symbols
(i.e., the inefficiency of EC).

Performance

Next we turn to the question of performance. Motivated

by the analysis presented in the previous section, we will
utilize an independent additive noise channel model with
a signal-to-noise ratio constraint. This hypothesis covers
many situations:
Situations where the actual physical channel is additive and with a signal-to-noise ratio constraint.
Situations where the channel is digital, has a finite
alphabet and one assumes that quantization noise is
white, independent, and the quantizer is properly
scaled (recall Section 2.2.1).
Situations where the channel is digital, has a countable alphabet, and one employs an ECDQ (see Section
2.2.3).
Situations where the channel is analog, but prone to
i.i.d. data dropouts (i.e., analog erasure channel; see
also [10, 11, 62]). Indeed, it has been recently shown

2008 Chinese Control and Decision Conference (CCDC 2008)

d
r

C(z)

G(z)

F (z)

It is possible to show that, provided C(z) is an admissible

controller for G(z) and the model for quantization holds,
the model in Figure 4 is MSS if and only if F (z) is stable,
strictly proper, A(z) is stable, minimum phase (MP) and
biproper, and the signal to noise ratio satisfies
2

> ||T (z) + S(z)F (z)||2 .

A(z)1

It is illustrative to note that, if ||T (z) + S(z)F (z)||2 ,

then J(A(z), F (z)) (unless all exogenous signals
have zero spectral density, in which case J(A(z)F (z))
0). Also, if , then J(A(z), F (z)) 0 and we
recover the non-networked performance that is achieved
with C(z) and a transparent communication link.

A(z)

v
q

Figure 4: Linear model for the NCS in Section 3.

(see [63]) that, for such data dropout profiles, analog

erasure channels are equivalent, up to second order
statistics, to an additive white noise channel with a
fixed signal-to-noise ratio. (This result is actually inspired by the work [12], where the authors implicitly
use a fixed signal-to-noise ratio constraint to design
dropout compensators in a specific networked situation.)
3.1 Setup and Analysis
We will focus on the NCS depicted in Figure 4, where the
controller C(z) is a given admissible controller for G(z).9
We will show how to design the coding system filters F (z)
and A(z) so as optimize overall closed loop performance.
(We refer the reader to [64] for controller design procedures.)
It is easy to see from Figure 4 that the variance of the tracking error, i.e.,
e  r y,

(8)

is given by
2

e2 = ||Tdre (z)dr (z)||2 + J(A(z), F (z)),

(9)

where
J(A(z), F (z)) 

 2
2 
||A(z)Tdry (z)dr (z)||2 T (z)A(z)1 (1 F (z))2
2

||T (z) + S(z)F (z)||2

and where Tdre (z) and Tdry denote the transfer functions
from [d r]T to e and y, respectively, dr (z) denotes a spectral factor of [d r]T , is the channel signal-to-noise ratio
and
S(z) =

1
,
1 + G(z)C(z)

3.2

Coding system design

Based on (9) and noting that Tdrw (z)dr (z) is fixed, we

next study the problem of finding10
Jopt 

an internally stabilizing controller that defines a well possed control loop (see,e.g., [3]).

inf

A(z)U
F (z)RH2
||T (z)+S(z)F (z)||22 <

J(A(z), F (z))

and filters A(z) and F (z) that achieve Jopt (or approximate
Jopt arbitrarily well). The exact solution of this problem is
not straightforward, so we describe a simple iterative solution.
If F (z) is a given strictly proper and stable filter (as required for MSS), then the Cauchy-Schwarz inequality imF (z)
plies that the optimal filter A(z), say Aopt (z), satisfies
(see [16], [18])


 F (z) j 4
Aopt (e ) =


T (ej )(1 F (ej ))2

,
Tdry (ej )d r(ej )d r(ej )H Tdry (ej )H

(11)

where is any arbitrary positive real. We note that the

F (z)
characterization of Aopt (z) given above is explicit but is
in general not satisfied by any stable, biproper and MP rational transfer function (indeed, the 4th root of the right
hand side in (11) is usually irrational). Nevertheless, since
the condition (11) holds on the magnitude of A(z) only, it
is always possible to find a filter with the desired characteristics that achieves a performance that is as close as desired to the optimal performance. In practice, it is usually
enough to consider reasonably low order filters to approxiF (z)
mate Aopt (z) (see also [16]).
Now consider that A(z) is any given stable, MP and
biproper transfer function. For any such A(z), it is possible to show that (see [18])

T (z)  1 S(z).

A straightforward calculation shows that Tdre (z) depends

on C(z) only, which is fixed in our setting. Thus,
to optimize performance we would like to minimize
J(A(z), F (z)).
9 i.e.,

(10)

A(z)

Fopt (z) = F (z),

(12)

where F (z),  [0, 1], is defined via

F (z)  arg

inf

F (z)RH2

J1 (F (z)) + (1 )J2 (F (z)),

10 RH denotes the set of all stable and strictly proper real rational
2
transfer functions, and U denotes the set of all stable, MP and biproper
transfer functions.

2008 Chinese Control and Decision Conference (CCDC 2008)


 2
J1 (F (z))  T (z)A(z)1 (1 F (z))2 ,

(0,
)

J1 (F (z))
.
J2 (F (z))

3.3 Example
Consider a nominal loop with plant and controller given by
1
,
z 0.8

C(z) =

z 0.8
.
z1

The disturbance d is assumed zero, whilst the reference is

considered to have a power spectral density with spectral
factor
r (z) =

0.02z
.
z 0.9

We consider a finite uniform quantizer with overload factor

= 4 and b bits. We send the output of the quantizer
directly through the channel (i.e., without entropy coding).
Thus, b corresponds to the instantaneous channel data-rate
in bits per sample.
Figure 5 shows the steady state tracking error variance e2
(see (9)) as a function of the number of iterations in the proposed design algorithm for = 9.1875, which corresponds
11 Provided both situations use the same channel and the same quantizer.

x 10

(1)

Case 1
Case 2
Case 3

(13)

In (13),  is defined as follows: If there does not exist an

 (0, 1] such that J2 (F (z)) = , then  = 1. Otherwise,  =  , where  is the unique real in (0, 1] such that
J2 (F (z)) = .
We note that the problem of finding F (z) is a standard
convex problem that can be tackled using well-known tools
(see, e.g., [65, 66]). On the other hand, calculating 
amounts to solving a (scalar) line search problem. This
is also a simple problem to solve.
By utilizing the previous results, it is immediate to envisage an iterative procedure that allows one to construct approximations to the optimal A(z), F (z) pair: In a first step
one fixes A(z) (or F (z); trivial choices are F (z) = 0 and
A(z) = 1). Then, one uses (12) to calculate the optimal
F (z) for the initial choice of A(z) (or (11) to calculate the
optimal A(z) for the initial choice of F (z)). This is repeated fixing A(z) or F (z) intermittently. This algorithm
is guaranteed to converge, at least, to a local minimum.
F (z)
A(z)
In general, Aopt (z) = 1 and Fopt (z) = 0. Thus, fixing
A(z) or F (z) and optimally choosing the other filter, will
obviously provide a coding system that enhances closed
loop performance when compared with a non-coded networked situation.11 From the above, it is also clear that
the use of the proposed iterative procedure allows one to
design coding systems that will always outperform simple
coding systems that do not consider feedback in the quantizer (i.e., where F (z) = 0; see [16]). This again highlights
the role of architectures (equivalently, available degrees of
freedom) in NCSs. Further discussions regarding the interplay between coding system architecture and networked
performance can be found in [16].

G(z) =

Tracking error variance ( 2 )

and
  arg min

= 9.1875 ( 3 [bit] )

7
6.5
6
5.5
5

(3)

4.5

(2)

4
6
8
Number of iterations

10

Figure 5: Tracking error as function of the number of iterations in proposed design procedure (see Section 3.3 for
details).
1

10
Tracking error variance ( 2e )

J2 (F (z))  ||T (z) +

2
S(z)F (z)||2

Nominal performance
No coding (empirical)
Opt. coding (empirical)
Opt. coding (analytical)

10

10

3
4
5
6
Bit rate [bit/sample]

Figure 6: Tracking error as function of the channel bit rate

(see Section 3.3).

to b = 3. Cases 1 and 2 refer to iterations that start with

A(z) = 1 and F (z) = 0. In Case 1 we initially fixed A(z),
whereas in Case 2 we start fixing F (z). Case 3 refers to iterations that start with the choices suggested in [17] (where
it is assumed that is high enough).
In Figure 5 we have identified three points. The first of
these (point (1)) refers to the performance achieved without
coding (F (z) = 0 and A(z) = 1). The second (point (2))
refers to the performance achieved when employing the optimal coding system proposed in [16]. The third (point (3))
refers to the performance achieved using the approximately
optimal filters described in [17].
The results show that coding is, indeed, useful to achieve
good closed loop performance. (Compare point (1) with,
e.g., the value of e2 after 10 iterations.) It is also possible to see that use of the proposed procedure yields coding systems that perform better than the simpler proposals
in [16,17]; see also the discussion at the end of Section 3.2.
(Compare points (2) and (3) with the limiting value for e2 .)
Finally, we examine the behavior of the tracking error variance as a function of the channel bit rate b. The results
are presented in Figure 6, where Nominal performance
refers to the performance achieved by the nominal loop
(without quantization), No coding (empirical) refers to

2008 Chinese Control and Decision Conference (CCDC 2008)

simulated results12 when no coding is employed (i.e., when

A(z) = 1 and F (z) = 0), Opt. coding (empirical) refers
to simulated results obtained with the filters suggested in
Section 3.2 (after 10 iterations), and Opt. coding (analytical) refers to the corresponding predictions made using the simplified noise model for quantization. One can
see that, as expected, the effects of quantization vanish as
b . Interestingly, the predictions made by the additive
noise model turn out to be very accurate for every bit rate:
indeed, for b 3 the relative errors are of less than 1% and,
for b {1, 2}, the relative errors are around 8%. (We note
that F (z) = 0 turns out to be non-admissible for b = 1,
i.e., (10) is not satisfied. Accordingly, we have omitted the
non coded results for b = 1.)

qi
mi

Next we turn to MIMO systems. Our development here

is largely based on [33]. We focus on control of 2 2
MIMO LTI plants modes and explore the potential benefits of enhancing a traditional diagonal non-networked control architecture with additional non-transparent channels
which are subject to signal-to-noise ratio constraints. We
will show how to design LTI coding systems which optimize overall performance.
4.1 Setup
As before, G(z) denotes the plant model. We assume that
an admissible full MIMO controller, say


C11 (z) C12 (z)
C(z) =
C21 (z) C22 (z)
has already been designed for G(z). The diagonal terms
of this controller are implemented without communication
constraints, but the off-diagonal terms communicate using
non transparent communication links.
We will focus on a situation where the non transparent
communication links are as in Figure 7. In that figure,
Fi (z) is the i-th (i {1, 2}) coder transfer function, vi
is the i-th channel input and wi is the i-th channel output.
These signals are related via
wi = vi + qi ,
where qi is the i-th channel noise. Each noise sequence is
considered white, having variance 0 i2 < and power
spectral density i (ej ) = i2 , [, ]. As in previous sections, we assume that i2 is a design parameter that
is proportional to the variance of the channel input (namely,
proportional to v2i ). We define the associated i-th channel
signal-to-noise ratio as
i 

v2i
.
i2

The NCS which results from employing the links described

above to implement the off-diagonal terms of C(z) can be
visualized as in Figure 8. In that figure, u = [u1 u2 ]T is the
12 All simulations use an actual uniform quantizer with L = 2b levels.
For each b, the results correspond to the average of 200 simulations (each
one 105 samples long and using a different reference realization).

wi

Fi (z)

ni

channel

Figure 7: i-th communication link.

plant input, y = [y1 y2 ]T is the plant output, r = [r1 r2 ]T
is the reference sequence, and e = [e1 e2 ]T denotes the
tracking error, i.e.,
e  r y.
4.2

MIMO systems

Fi (z)1

vi

Analysis

It is easy to see from Figure 8 that13

2

e2 = ||S(z)r (z)||2 +

2


i2 ||S(z)G(z)i Fi (z)||2 ,

i=1

(14)
2


v2i = ||Ai (z)r (z)||2 +

j2 ||Ai (z)G(z)F (z)j ||2 ,

j=1

(15)
where r (z) is a spectral factor of r and
F (z)  diag {F1 (z), F2 (z)} ,
S(z)  (I + G(z)C(z))1 ,
A1 (z)  F1 (z)1 C12 (z)T2 S(z),
A2 (z)  F2 (z)1 C21 (z)T1 S(z).
Equation (15) allows one to establish that the NCS described above is MSS if and only if both F1 (z) and F2 (z)
are stable, MP and biproper,

 2
(16)
1 > B1  C12 (z)T2 S(z)G(z)1 2 ,

 2
T




2 > B2  C21 (z)2 S(z)G(z)2 2 ,
(17)
and
(1 B1 ) (2 B2 ) >
2

||A1 (z)G(z)F (z)2 ||2 ||A2 (z)G(z)F (z)1 ||2 .

(18)

Thus, and as expected from the SISO case examined earlier, we see that MSS imposes limits on the achievable
channel signal-to-noise ratio and, hence, on the corresponding channel data-rate.
An immediate consequence of the previous result is that,
provided (16)-(18) hold,
lim

(1 ,2 )(
1 ,
2 )

e2 = ,

where


S  (1 , 2 ) R2 : 1 and 2 achieve equality in (18) ,
13
i

is the ith element of the canonical basis in R2 .

2008 Chinese Control and Decision Conference (CCDC 2008)

r1

e1

y1

u1

C11 (z)
v2

F2 (z)1

C21 (z)

F2 (z)

Link 2

q2

Link 1

q1

G(z)

r2

e2

v1

F1 (z)1

C12 (z)

F1 (z)
u2

C22 (z)

y2

Figure 8: Partly networked MIMO control architecture (see Section 4).

for any C(z), and any choice for F1 (z) and F2 (z). As a
consequence, we see that, for any given full MIMO controller, and no matter how the coding system is chosen,
there exist sufficiently poor channels which render the performance of the resulting partially networked closed loop
arbitrary bad. In these cases, any stabilizing decentralized
controller (that makes no use of the non-transparent channels) will provide better performance than a NCS. Stated a
different way, in the situation under study, there are cases
where poor information is much worse than not having information at all. Of course, this conclusion is tied to the
fact that we are considering a pre-designed controller. If
one were to design a centralized controller considering the
communication constraints from the very beginning, then
one could outperform any decentralized design.
4.3

Coder Design

In this section we show how to design optimal coders F1 (z)

and F2 (z) under a mild simplifying assumption.
From (15) one can immediately conclude that, provided
(16)-(18) are satisfied and 1 , 2 are large enough, then
e2 J, where
J

2
2
2

||Ai (z)r (z)|| ||Sd (z)i Fi (z)||
2

i=1

i Bi

We will denote the coders that minimize J by F1o (z) and

F2o (z).
Using simple arguments (again based on the CauchySchwarz inequality), it is possible to see that
 o j 4
Fi (e ) = i

Mi (ej )Mi (ej )H

H

(Sd (ej )i ) Sd (ej )i

where
M1 (z)  C12 (z)T2 S(z)r (z),
M2 (z)  C21 (z)T1 S(z)r (z),
and i is an arbitrary positive constant. This results shows
how to synthesize coding systems that minimize the impact
of the communication links on overall closed loop performance. An interesting feature of the proposed filters is that

they do not depend on the channel signal-to-noise ratios.

This allows one to conjecture that the optimal filters will
perform well for a large class of communication channels.
4.4

Example

We end this section with an example that considers (instantaneous) bit-rate limited channels and, for simplicity, we
assume equal bit rates on each channel, i.e., b1 = b2 = b,
and take the sampling interval as 1[s]. The plant model is
given by


G(z) =

0.6
(z0.8)
1
(z0.5)

0.4
(z0.8)
1
(z0.5)

For this plant we synthesize the decentralized controller

 1.3333(z0.8)

0
(z+0.8)(z1)
,
Cd (z) =
0.8(z0.5)
0
(z+0.8)(z1)
and the full MIMO controller

C(z) =

5(z0.8)
z1
5(z0.8)
z1

2(z0.5)
z1
3(z0.5)
z1


.

We also assume that the reference description is given by

r (z) =

0.0049627(z + 0.9934)
I,
(z 2 1.97z + 0.9802)

and that the quantizers are as in Section 3.3.

Figure 9 shows the tracking error variance in the proposed
networked MIMO architecture as a function of the perchannel bit-rate b in several situations: Analytical no coding refers to the performance predicted by (14) and (15)
when no coding is employed; Empirical no coding refers
to simulated performance when no coding is considered;
Analytical opt. coding and Empirical opt. coding refer
to analytical and simulated performance when the coders
suggested by in Section 4.3 are employed. For comparison purposes, Figure 9 also shows the non networked
full MIMO performance (Ideal full MIMO) and the performance achieved when using Cd1 (z) (Decentralized).
The results allow one to conclude that, in this case, the benefits of coding are significative.

2008 Chinese Control and Decision Conference (CCDC 2008)

Tracking error variance ( e2 )

10

Decentralized
Ideal full MIMO
Analytical no coding
Empirical no coding
Analytical opt. coding
Empirical opt. coding

10

Controller

x
Plant

Figure 10: NCS over an unreliable network.

2

10

(see, e.g., [2, 3, 68, 69]). Assuming that the state x is available for measurement, in a non-networked case, the plant
input can be considered as if were given by a (possibly
time-varying and dynamic) mapping of the plant state, say

5
6
Bit rate ( b ) [bit/s]

Figure 9: Tracking error variance as a function of the perchannel bit rate (see Section 4.4).

As expected, non-networked full MIMO performance is

recovered as b , irrespective of the coders used.
Moreover, it is apparent that the non coded networked full
MIMO architecture should be preferred to the decentralized one for b > 3.36. On the other hand, the optimally
coded networked MIMO architecture provides significant
improvement in performance for b > 3.20, when compared
to the decentralized architecture. It is also interesting to
note that, if b 3.07, then the performance becomes arbitrary poor no matter what the coding is. This is consistent
with (16)-(18) as straightforward calculations reveal, and
brings back the question of how to actually design MIMO
controllers for the considered situation (not just coding systems for a given fixed controller design).

Data Loss and Delays

Up to this point in this paper, we have focused on quantization issues. However, in modern network protocols (e.g.,
Ethernet; see [67]), data is sent in large packets. Accordingly, quantization effects may become negligible in those
situations, and the fact that data packets may get corrupted,
delayed or lost becomes the dominant issue. In this section (which is based on [34]) we will present an approach
that allows one to deal with data loss and random delays.
Our proposal exploits the assumption that it is possible to
transmit relatively large packets that cover multiple datadropout and delay scenarios.
5.1

Buffer and
selection logic

10

u
Unreliable
Network

Setup

We consider a discrete-time nonlinear plant model described in state-space form via:

x(k + 1) = f (x(k), u(k)).

(19)

The plant input and state are constrained according to:

u(k) U R ,

x(k) X R ,

k N0  N{0},

where N denotes the set of positive integers, denotes the

dimension of the plant input, whilst refers to the state
dimension.
To control (19), one can employ standard discrete-time
controllers operating at the same sampling rate as the plant

u(k) = k (x(k)),

k N0 ,

(20)

where k : X U is the control policy.

We will now show how one can embed the control law (20)
in a more general control strategy where the controller
communicates with the plant input via an unreliable network (see Figure 10). This network is assumed to be able
to carry large packets of data (say, a few kilobytes). However, each packet, say U (k), may be delayed or even lost
(due to buffer overflows and transmission errors).
We assume that the delay experienced by U (k) has both
a fixed component, which can be included in the plant
model (19), and a time-varying component, say (k). The
latter depends on several factors including network load.
Since we will concentrate on a configuration where sensors, actuators and controller operate at the same sampling
rate, it suffices to consider (k) N0 . Packets which are
delayed by more than a given value, say max , will be regarded as lost. Thus, we will use the term lost to refer to
those packets which are effectively dropped by the network
as well as to those which do not arrive during a prespecified time frame. Lost packets will not be used further in the
NCS strategy described in this section.
The maximum delay max constitutes a timeout value
which can be designed based upon network delay and
dropout characteristics, see, e.g., [70]. The model adopted
above includes pure erasure channels (see, e.g., [10, 39,
40]), as a particular case, by setting max = 0.
5.2 Scenario Based Networked Control
This section briefly describes a control strategy to deal with
the situation described above. A detailed description of the
control policy can be found in [34].
At the plant side, there exists a buffer, say b(k), that contains N control input values to be sequentially passed on to
the actuator. The buffer is updated as follows: If the packet
U (k) is received (without error) at time instant k + (k)
and none of the packets
{U (k + 1), U (k + 2), . . . , U (k + (k))}

(21)

have been received by time k + (k), then U (k) is used to

replace the current content of the buffer. Otherwise, if any
of the more recent packets in (21) have been received by
time k + (k), then U (k) is discarded. This guarantee that
only fresh data is used at the plant side.
At each time instant the controller needs to send a packet
U (k). This packet is formed assuming knowledge of the

2008 Chinese Control and Decision Conference (CCDC 2008)

peq (k)  1 P { (k) = 0}

N
1 
max


j=0 i=0
(i,j)=(0,0), i+jk


P (k i j) = i (k) > 0


(k 1) > 1 (k j i + 1) > j + i 1 .
This important result, implies that, in order to analyze or
design SBNC loops, it suffices to consider a setting wherein
the network is modeled via an erasure channel with a given
dropout probability (see, e.g., [1012, 71, 72] and the many
references therein).
It should be emphasized here that if { (k)}kN0 is a sequence of independent random variables, then
peq (k) = peq ,
i.e., the equivalent probability, is a constant. However, the
sequence of dropouts in SBNC is, in general, not a sequence of independent random variables, even if the underlying network delay and dropout distributions are.
The equivalent dropout probability characterizes closed
loop control performance. It can be used as a guideline
for choosing the horizon length N and the timeout value
max . Indeed, peq (k) can be made arbitrarily small (and,
thus, the performance will become indistinguishable from
that achieved in the non-networked case). This is achieved
by choosing sufficiently large values for N and max , albeit at an increase in computational complexity (for details
see [34]). Fortunately, in practice, choosing moderate values for N and max is often sufficient to achieve good performance as illustrated below.
14 This implies that the network protocol must have acknowledgements,
as in TCP.
15 This implies that the actual delay distribution does not play any role.
16 P{} stands for probability of ().

10

Probability

current plant state x(k), the buffer state in the previous

sampling instant b(k 1),14 and the previously sent packets. With this information, the controller uses (20) to generate a set of N -length sequences of control inputs, each
one taking into account one possible value for the delay15
(k) {0, 1, . . . , max } and one possible set of packets
that may arrive before U (k). Each one of these transmission outcomes constitutes a scenario. Upon reception, and
if U (k) arrived before U (k + i), i 1, the buffer selection
logic at the plant uses time-stamping to determine the delay experienced by U (k) and, based upon the knowledge of
which packets where received in [k, k + (k) 1], selects
the corresponding control inputs sequence.
It is possible to analyze the previous strategy in a precise
fashion. Indeed, one can derive deterministic performance
guarantees (see [34]). Here, we concentrate on a key result
that states that controlling a nonlinear constrained unstable
plant model over an unreliable channel by means of SBNC
amounts (essentially) to controlling the same plant over an
erasure channel with equivalent data dropout probability
peq (k) given by16

0.2
0.15
0.1
0.05
0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Network delay ( )

Figure 11: Delay distributions of the network examined in

Section 5.3.
5.3

Example

This section presents an example to illustrate the performance of SBNC. We will assume that the network has, for
every k, a delay profile as illustrated in Figure 11. Thus, if,
for example, one sets max 4, then most packets will be
effectively lost.
Consider the following stable nonlinear plant having scalar
input


x1 (k + 1) = x1 (k) + 0.01 x2 (k) + x2 (k)3

x2 (k + 1) = x2 (k) + 0.01 2x1 (k) 3x2 (k)


+ u(k) 1 + 0.1x1 (k)2
y(k) = x1 (k) + d(k),
where x(k) = [x1 (k) x2 (k)]T is the plant state at time k,
{u(k)} is the plant input, {y(k)} is the plant output, and
{d(k)} is a piecewise constant output disturbance having
infrequent steps at random times. Both the plant state and
disturbance measurements are affected by Gaussian white
noise of variance d2 = 0.01 and x2 = 0.04I2 , respectively.
The control objective is to achieve reference tracking for
the plant output. Whilst future reference values are known
to the controller, future disturbances are not. For simplicity,
the latter are assumed to remain at their current value.
The underlying controller is given by:
k (x(k)) =

ry (k) d(k) B(x(k))

,
A(x(k))

where {ry (k)} is the reference for y(k),

A(x) = (1 + 3x22 )(1 + 0.1x21 )/9,
B(x) = (2x1 + 6x22 x1 + 9x32 4x2 4x32 9x1 )/9.
We illustrate the effect on performance of the SBNC design
parameters max and N in Figure 12,17 where typical plant
output trajectories for step-like reference and disturbances
are shown. For each pair (N, max ) we also calculated the
equivalent dropout probability peq (k) (which is constant in
this case).
As expected, closed loop performance improves when N
and max are increased. However, it is also appreciated
that there are cases where increasing N has no obvious effect on SBNC performance. This is easily explained if one
notes that, as suggested by our comments in Section 5.2,
17 As mentioned in Section 5.2, SBNC does not require knowledge of
the statistical properties of the delay profiles.

2008 Chinese Control and Decision Conference (CCDC 2008)

Output y

N=2

max = 1
3

Output y

N=6

Output y

2000

4000

peq = 0.784
0

2000

4000

1
0

peq = 0.844
0

2000

4000

1
0

peq = 0.607
0

2000

4000

peq = 0.841
0

2000

4000

1
0

peq = 0.604
0

2000

4000

2000
4000
Sample number k

eq

= 0.604

2000
4000
Sample number k

2000

2000

4000

eq

2000

1
0

4000

peq = 0.013

peq = 0.841

peq = 0.295

Output y

peq = 0.930

N = 10

max = 7

N = 15

max = 2

= 0.010
4000

peq = 0.010
0

2000
4000
Sample number k

Figure 12: SBNC closed loop performance as a function of N and max (see Section 5.3).
the equivalent dropout probability peq is the key parameter
determining SBNC performance. Indeed, cases with very
disparate values of N or max can exhibit similar performance if the corresponding equivalent dropout probabilities are similar.18
It is also informative to compare the probability of a packet
being effectively lost due to the choice of max , and the
equivalent dropout probability peq (which depends on the
interplay between N and max ). It follows from Figure 11
that if max {1, 2, 7}, then P{ (k) > max } equals
0.959, 0.877 and 0.143, respectively. The above results
show that by a proper choice of N one can decrease this
probability dramatically. It must be noted, however, that
if one chooses N too small (as compared with max ), then
peq could be larger than P{ (k) > max } (see the case
(N, max ) = (2, 7)). This is due to the fact that, in those
cases, the buffer will frequently run out of data.
It is also worth mentioning that, for this problem, using
max = 0, does not stabilize the loop for any value of N .
In addition, the simple policy that regards all delayed data
as lost and only sends the current plant input also gives
unstable behaviour. This strongly supports the use of the
SBNC scheme advocated here.
18 Consider, for example, the cases (N,
max ) = (15, 1) and
(N, max ) = (2, 2) or, more dramatically, (N, max ) = (6, 7) and
(N, max ) = (15, 7).

Conclusions

Networked Control Systems are control systems in which

the communication channels between plant and controller
are not transparent. Specifically, NCSs are control systems
which include non standard features such as quantization
(i.e. data rate limits), data dropouts and random delays.
NCS theory brings together fundamental (and contemporary) ideas from control theory, information theory, signal
processing and communications theory. It is, thus, an excellent area for the systems scientist to contribute to.
In specific areas, significant results have been found that allow one to analyze and synthesize networked control strategies. Nevertheless, we feel that the big picture is still missing. Many questions remain unanswered. For example, the
issues of robustness to plant and channel model errors and
adaptation need further work. More ambitious is the search
of explicit solutions to certain kinds of problems, which
would certainly increase the understanding of the interactions between control objectives and communications constraints.
Beyond technical details, the key open problem in NCSs is
the development of a unified framework which takes channel issues into account and gives practice-oriented design
methodologies. This makes this research area a challenging and interesting one, with a high development potential.

2008 Chinese Control and Decision Conference (CCDC 2008)

11

Acknowledgment
The authors would like to thank M.S. Derpich and J. stergaard for fundamental contributions to the work summarized here.

REFERENCES

[23]

[1] S. Skogestad and I. Postlethwaite, Multivariable Feedback

Control: Analysis and Design. New York: Wiley, 1996.
[2] G. Goodwin, S. Graebe, and M. Salgado, Control System
Design. New Jersey: Prentice Hall, 2001.
[3] K. Zhou, J. Doyle, and K. Glover, Robust and optimal control. Englewood Cliffs, N.J.: Prentice Hall, 1996.
[4] P. Antsaklis and J. Baillieul, Guest editorial special issue on networked control systems, IEEE Trans. Automat.
Contr., vol. 49, no. 9, pp. 14211423, Sept. 2004.
[5] , Special issue on technology of networked control
systems, Proc. IEEE, vol. 95, no. 1, pp. 58, January 2007.
[6] G. Nair and R. Evans, Stabilizability of stochastic linear
systems with finite feedback data rates, SIAM J. Control
and Optimization, vol. 43, no. 2, pp. 413436, July 2004.
[7] G. Nair, F. Fagnani, S. Zampieri, and R. Evans, Feedback
control under data rate constraints: An overview, Proc.
IEEE, vol. 95, no. 1, pp. 108137, January 2007.
[8] A. Savkin, Analysis and synthesis of networked control
systems: Topological entropy, observability, robustness and
optimal control, Automatica, vol. 42, pp. 5162, 2006.
[9] S. Tatikonda and S. Mitter, Control under communication
constraints, IEEE Trans. on Automatic Control, vol. 49,
no. 7, pp. 10561068, July 2004.
[10] L. Schenato, B. Sinopoli, M. Franceschetti, K. Poolla, and
S. S. Sastry, Foundations of control and estimation over
lossy networks, Proc. IEEE, vol. 95, no. 1, January 2007.
[11] P. Seiler and R. Sengupta, An H approach to networked
control, IEEE Trans. Automat. Contr., vol. 50, no. 3, 2005.
[12] Q. Ling and M. Lemmon, Power spectral analysis of networked control systems with data dropouts, IEEE Trans.
Automat. Contr., vol. 49, no. 6, pp. 955960, June 2004.
[13] J. Nilsson, Real-time control systems with delays, Ph.D.
dissertation, Lund Institute of Technology, 1998.
[14] F. Lian, J. Moyne, and D. Tilbury, Modelling and optimal
controller design of networked control systems with mutliple delays, Int. J. Contr., vol. 76, no. 6, pp. 591606, 2003.
[15] E. Silva, M. Derpich, J. stergaard, and D. Quevedo, Simple low bit-rate coding for achieving mean square stability,
in Submitted to the 46th IEEE Conference on Decision and
Control, Cancun, Mexico, 2008.
[16] G. Goodwin, D. Quevedo, and E. Silva, Architectures and
coder design for networked control systems, Automatica,
vol. 44, no. 1, pp. 248257, 2008.
[17] E. Silva, G. Goodwin, D. Quevedo, and M. Derpich, Optimal noise shaping for networked control systems, in Proc.
of the European Control Conference, Kos, Greece, 2007.
[18] E. Silva, D. Quevedo, M. Derpich, and G. Goodwin, Design of feedback quantizers for networked control systems,
Submitted to Int. J. Robust Nonlinear Control, 2008.
[19] C. Canudas de Wit, F. Rodrguez, J. Fornes, and F. GomezEstern, Differential coding in networked controlled linear
systems, in Proc. Amer. Contr. Conf., Minneapolis, MN,
2006.
[20] C. Canudas de Wit, J. Jaglin, and K. Vega, Entropy coding
for networked controlled systems, in Proc. of the European
Control Conference, Kos, Greece, 2007.
[21] I. Lopez, C. Abdallah, and C. Canudas de Wit, Gainscheduling multi-bit delta-modulator for networked con-

12

[22]

[24]

[25]

[26]
[27]

[28]
[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

trolled system, in Proc. of the European Control Conference, Kos, Greece, 2007.
Y. Minami, S. Azuma, and T. Sugie, An optimal dynamic
quantizer for feedback control with discrete-valued signal
constraints, in Proc. of the 46th IEEE Conference on Decision and Control, New Orleans, USA, 2007.
S. Azuma and T. Sugie, Optimal dynamic quantizers for
discrete-valued input control, Automatica, vol. 44, pp.
396406, 2008.
M. Salgado and A. Conley, MIMO interaction measure
and controller structure selection, Int. J. Contr., vol. 77,
no. 4, pp. 367383, March 2004.
G. Goodwin, M. Salgado, and E. Silva, Time-domain performance limitations arising from decentralized architectures and their relationship to the RGA, Int. J. Contr.,
vol. 78, no. 13, pp. 10451062, September 2005.
V. Kariwala, Fundamental limitation on achievable decentralized performance, To appear in Automatica, 2007.
E. Silva, D. Oyarzun, and M. Salgado, On structurally constrained H2 performance bounds for stable MIMO plant
models, IET Control Theory & Applications, vol. 1, no. 4,
pp. 10331045, 2007.
H. Ishii and B. Francis, Stabilization with control networks, Automatica, vol. 38, pp. 1745 1751, 2002.
J. Rawlings and B. Stewart, Coordinating multiple optimization based controllers: New opportunities and challenges, in Proc. of the 8th Int. IFAC DYCOPS, Cancun,
Mexico, June 2007.
S. Yuksel and T. Basar, Communication constraints for
decentralized stabilizability with time-invariant policies,
IEEE Trans. on Automatic Control, vol. 52, no. 6, pp. 1060
1066, 2007.
A. Matveev and A. Savkin, Decentralized stabilization
of linear systems via limited capacity communication networks, in Proc. of the 44th IEEE CDC and the ECC 2005,
Sevilla, Spain, 2005, pp. 11551161.
S. Jiang and P. Voulgaris, Cooperative control over link
limited abd packet dropping networks, in Proc. of the European Control Conf., 2007.
E. Silva, G. Goodwin, and D. Quevedo, On networked
control architectures for mimo plants, in Accepted for presentation at the 17th IFAC World Congress, Seoul, Korea,
2008.
D. Quevedo, E. Silva, and G. Goodwin, Control over unreliable networks affected by packet erasures and variable
transmission delays, To appear in IEEE Journal On Selected Areas In Communications, 2008.
A. Bemporad, Predictive control of teleoperated constrained systems with unbounded communication delays,
in Proc. IEEE Conf. Decis. Contr., 1998, tampa, Florida,
USA.
G. Liu, J. Mu, D. Rees, and S. Chai, Design and stability
analysis of networked control systems with random communication time delay using the modified MPC, Int. J.
Contr., vol. 79, pp. 288297, April 2006.
A. Casavola, E. Mosca, and M. Papini, Predictive teleoperation of constrained dynamic systems via internet-like
channels, IEEE Trans. Contr. Syst. Technol., vol. 14, pp.
681694, July 2006.
P. L. Tang and C. W. de Silva, Compensation for transmission delays in an ethernet-based control network using variable-horizon predictive control, IEEE Trans. Contr.
Syst. Technol., vol. 14, no. 4, pp. 707718, July 2006.
V. Gupta, B. Sinopoli, S. Adlakha, and A. Goldsmith, Receding horizon networked control, in Proc. Allerton Conf.

2008 Chinese Control and Decision Conference (CCDC 2008)

[40]

[41]

[42]
[43]

[44]

[45]

[46]

[47]

[48]

[49]

[50]

[51]

[52]

[53]

[54]

[55]

Communications, Control, and Computing, Monticello, IL,

USA, 2006.
D. Quevedo, E. Silva, and G. Goodwin, Packetized predictive control over erasure channels, in 26th American Control Conference, New York, USA, 2007.
S. Tatikonda and S. Mitter, Control over noisy channels,
IEEE Trans. on Automatic Control, vol. 49, no. 7, pp. 1196
1201, July 2004.
T. Cover and J. Thomas, Elements of Information Theory,
2nd ed. John Wiley and Sons, Inc., 2006.
D. Delchamps, Stabilizing a linear system with quantized
state feedback, IEEE Trans. on Automatic Control, vol. 35,
no. 8, pp. 916924, August 1990.
D. Quevedo, G. Goodwin, and J. De Dona, Finite constraint set receding horizon control, International Journal of Robust and Nonlinear control, vol. 14, pp. 355377,
March 2004.
G. Nair and R. Evans, Exponential stabilisability of finitedimensional linear systems with limited data rates, Automatica, vol. 39, no. 4, pp. 585593, April 2003.
J. Hespanha, A. Ortega, and L. Vasudevan, Towards the
control of a linear system with minimum bit-rate, in Proceeding 15th Int. Symp. Mathematical Theory Networks
System. Univ. Notre Dame, August 2002.
F. Fagnani and S. Zampieri, Stability analysis and synthesis for scalar linear systems with a quantized feedback,
IEEE Trans. Automat. Contr., vol. 48, no. 9, pp. 15691584,
Sept. 2003.
G. Nair, R. Evans, I. Mareels, and W. Moran, Topological feedback entropy and nonlinear stabilization, IEEE
Trans. on Automatic Control, vol. 49, no. 9, pp. 15851597,
September 2004.
J. Freudenberg, R. Middleton, and V. Solo, The minimal
signal-to-noise ratio requirements to stabilize over a noisy
channel, in Proc. of the 2006 American Control Conference, Minneapolis, USA, June 2006.
N. Elia, When Bode meets Shannon: Control oriented
feedback communication schemes, IEEE Trans. on Automatic Control, vol. 49, no. 9, pp. 14771488, 2004.
N. Martins and M. Dahleh, Fundamental limitations of
performance in the presence of finite capacity feedback, in
Proceedings of the American Control Conference, Portland,
USA, June 2005.
N. Jayant and P. Noll, Digital Coding of Waveforms. Principles and Approaches to Speech and Video. Prentice Hall,
1984.
R. Gray and T. S. (Jr.), Dithered quantizers, IEEE Transactions on Information Theory, vol. 39, no. 3, pp. 805812,
May 1993.
C. Gunturk, One-bit sigma-delta quantization with exponential accuracy, Commun. Pure Appl. Math., vol. 56,
no. 11, pp. 16081630, 2003.
R. Schreier and G. Temes, Understanding Delta Sigma

Data Converters. Wiley-IEEE Press, 2004.

[56] H. Bolcskei and F. Hlawatsch, Noise reduction in oversampled filter banks using predictive quantization, IEEE
Trans. Inform. Theory, vol. 47, no. 1, pp. 155172, January
2001.
[57] W. Bennet, Spectra of quantized signals, Bell Syst. Tech.
J., vol. 27, no. 4, pp. 446472, 1948.
[58] R. Gray, Quantization noise spectra, IEEE Transactions
on Information Theory, vol. 36, pp. 12201244, November
1990.
[59] R. Zamir and M. Feder, On universal quantization by randomized uniform/lattice quantizers, IEEE Transactions on
Information Theory, vol. 38, no. 2, pp. 428436, March
1992.
[60] , Information rates of pre/post-filtered dithered quantizers, IEEE Transactions on Information Theory, vol. 42,
no. 5, pp. 13401353, September 1996.
[61] M. Derpich, E. Silva, D. Quevedo, and G. Goodwin, On
optimal perfect reconstruction feedback quantizers, Undergoing second review in IEEE Transaction on Signal Processing, 2008.
[62] N. Elia, Remote stabilization over fading channels, Syst.
& Contr. Lett., pp. 237249, 2005.
[63] E. Silva, On networked control with i.i.d. data dropouts,
School of EE and CS, The University of Newcastle, Australia, Tech. Rep., 2007.
[64] E. Silva, D. Quevedo, and G. Goodwin, Optimal controller
design for networked control systems, in Accepted for presentation at the 17th IFAC World Congress, Seoul, Korea,
2008.
[65] S. Boyd and C. Barratt, Linear Controller Design-Limits of
Performance. New Jersey: Prentice Hall, 1991.
[66] S. Boyd, E. F. L. El Ghaoui, and V. Balakrishnan, Linear
Matrix Inequalities in System and Control Theory. SIAM,
1994.
[67] G. Kaplan, Ethernets winning ways, IEEE Spectrum,
vol. 38, no. 1, pp. 113115, 2001.
[68] A. Isidori, Nonlinear Control Systems, 3rd ed. Springer
Verlag, 1995.
[69] G. C. Goodwin, M. M. Seron, and J. A. De Dona, Constrained Control & Estimation An Optimization Approach. London: Springer-Verlag, 2005.
[70] C. Bovy, H. Mertodimedjo, G. Hooghiemstra, H. Uiterwaal, and P. Van Mieghem, Analysis of end-to-end delay measurements in the internet, in Proc. Passive and Active Measurement Workshop-PAM, Fort Collins, Colorado,
USA, 2002.
[71] O. Imer, S. Yuksel, and T. Basar, Optimal control of LTI
systems over unreliable communication links, Automatica,
vol. 42, pp. 1429 1439, 2006.
[72] J. Hespanha, P. Naghshtabrizi, and Y. Xu, A survey of
recent results in networked control systems, Proc. IEEE,
vol. 95, no. 1, pp. 138162, January 2007.

2008 Chinese Control and Decision Conference (CCDC 2008)

13