IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO.

2, FEBRUARY 1997 171

Adaptive Control Using Multiple Models

Kumpati S. Narendra Fellow, IEEE, and Jeyendran Balakrishnan, Member, IEEE

Abstract— Intelligent control may be viewed as the ability of stability properties are well understood. This is the case with
a controller to operate in multiple environments by recogniz- linear time-invariant (LTI) systems as well as adaptive sys-
ing which environment is currently in existence and servicing tems, the stability properties of which have been investigated
it appropriately. An important prerequisite for an intelligent
controller is the ability to adapt rapidly to any unknown but extensively. We adopt a similar philosophy for the design
constant operating environment. This paper presents a general of controllers which involve multiple models, switching, and
methodology for such adaptive control using multiple models, tuning. One of the objectives of this paper is to take a
switching, and tuning. The approach was first introduced in [1] first step in this direction by addressing the question of
and [2] for improving the transient response of adaptive systems
in a stable fashion. This paper proposes different switching
stability for a fairly broad class of switching and tuning
and tuning schemes for adaptive control which combine fixed systems.
and adaptive models in novel ways. The principal mathematical In [1] and [2] the problem of model reference adaptive
results are the proofs of stability when these different schemes control was considered, where multiple adaptive identification
are used in the context of model reference control of an unknown models were used to identify an LTI plant, and the adaptive
linear time-invariant system. A variety of simulation results are
presented to demonstrate the efficacy of the proposed methods. controller corresponding to the one yielding the minimum of
a performance index was chosen at every instant. The method
Index Terms— Adaptive control, multiple models, stability,
resulted in improved transient response in the presence of large
switching, transient response.
parametric uncertainties. In this paper, the basic idea contained
in [1] and [2] is extended into a general methodology. The
I. INTRODUCTION paper introduces different classes of switching and tuning
schemes which combine fixed and adaptive models in novel
T HE requirements of any good control system are speed,
accuracy, and stability. Achieving these in complex sys-
tems, in the presence of large uncertainty concerning the
ways. The aim is to determine flexible architectures for stable
intelligent control that can yield a fast and accurate response,
process to be controlled, is the challenge for the control yet remain computationally efficient. This constitutes the con-
theorist today. The realization that conventional controllers ceptual contribution of the paper. The theoretical contribution
do not possess all the attributes necessary to achieve such of the paper is the proof of stability of the overall system for
control has, in recent years, given rise to several definitions these different kinds of switching and tuning schemes, in the
for intelligent control. We adopt the perspective that intelligent context of model reference control of unknown LTI systems.
control is merely the ability of the control system to operate A series of simulation results is presented which demon-
successfully in a wide variety of situations by detecting the strates that the proposed methods indeed result in substantial
specific situation that exists at any instant and servicing it improvement in performance.
appropriately. External disturbances, changes in subsystem The individual concepts of multiple models, switching,
dynamics, parameter variations, etc., are examples of different or tuning are not new in control theory. Multiple Kalman
unknown environments in which the system has to operate. filter-based models were studied in the 1970’s to improve
Research has been in progress at Yale University, New Haven, the accuracy of the state estimate in estimation and con-
CT, since 1991, toward the development of a general method- trol problems, by Magill, Lainiotis, and others [4]–[6]. This
ology for the design of stable, fast, and accurate controllers was followed in later years by several practical applications
to cope with such time varying situations, using multiple [7]–[10]. In all these cases, no switching was involved, and
models, switching, and tuning. A qualitative description of only a convex combination of the control determined by
the approach may be found in [3]. The work described in this different models was used. Further, no stability results were
paper is the analytical part of the research activity, dealing reported.
with linear systems. In the context of adaptive control, switching was first intro-
It is well known that efficient design methods for various duced by Mårtennson [11]. Two kinds of switching schemes
classes of control systems can be developed only when their have been proposed in the literature. In direct switching
[11]–[15], the choice of when to switch to the next controller,
Manuscript received January 12, 1996; revised August 1996. Recommended in a predetermined sequence, is based directly on the output
by Associate Editor, B. Pasik-Duncan. This research was supported by the
National Science Foundation under Grant IRI 921-6823. of the plant. Such schemes have little practical utility. Indirect
K. S. Narendra is with the Center for Systems Science, Yale University, switching methods, in which multiple models are used to
New Haven, CT 06520-8267 USA (e-mail: [email protected]). determine both when and to which controller one should
J. Balakrishnan is with Adaptive Media Technologies, Inc., San Jose, CA
95113 USA. switch, are more attractive for applications. This approach
Publisher Item Identifier S 0018-9286(97)01328-7. was first proposed by Middleton et al. [16] and later adapted
0018–9286/97$10.00  1997 IEEE

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
172 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

in [17] and [18]. The objective in all the above efforts is be uncountable, identification of the environment takes place
to achieve stability in adaptive control with minimal prior in two stages. Assuming that the models and environments
information. In recent years, Morse [19] has been studying are parameterized suitably, the model with the smallest error,
the use of multiple fixed models and optimization for robust according to some criterion, is selected rapidly (switching)
set point control. In contrast to that, the work reported here and then its parameters are adjusted over a slower time scale
focuses on improving performance in multiple environments to improve accuracy (tuning). In switching, the problem is to
while retaining stability. determine when the current parameter value is unsatisfactory
The use of switching and tuning using multiple models (i.e., when to switch) and which one to replace it with (i.e.,
containing neural networks as components has been proposed what to switch to). In tuning, the problem is to determine the
in [20] to improve the performance of nonlinear systems. rule by which the parameter value is to be adjusted at each
However, the stability characteristics of such systems are instant.
substantially more difficult to analyze. The stability of linear
switching and tuning systems of the type analyzed in this pa- A. Architecture of the Control System
per, together with the implicit function theorem, are currently
The proposed architecture for intelligent control is shown
being used in the investigation of the stability of nonlinear
in Fig. 1. The system to be controlled has input and
switching and tuning systems. The methods presented in this
paper have found practical applications in robotic manipulator output . The objective is to make the control error
control [3] and are being investigated in the context of biologi- tend to zero, where is the desired output. The
cal gait control and control of fermentation processes. All these control system contains identification models, denoted by
different practical applications reveal that the methodology of , operating in parallel. The parameter vector of
control based on multiple models, switching, and tuning, while each may either be fixed or may be tuned from an initially
being theoretically attractive, is also practically viable. chosen value. The identification error between the output
of and that of the plant is denoted as .
Corresponding to each is a parameterized controller ,
II. GENERAL METHODOLOGY whose parameter vector is chosen such that achieves
the control objective for . The output of is denoted by
Control in Multiple Environments: Control system design . At every instant, one of the models is selected by
has traditionally been based on a single fixed or slowly a switching rule, and the corresponding control input is
adapting model of the system. This implicitly assumes that the
used to control the plant. This architecture was introduced in
operating environment is either time invariant, or varies slowly
[1], where all the models and controllers were adaptive with
with time. As control theory is extended to more complex
identical structures.
systems, many types of changes other than slow parameter
Given prior knowledge of the different possible environ-
variations are encountered, e.g., faults in the system, changes
ments, the design problem is to choose the number and
in subsystem dynamics, sensor and actuator failures, external
structure of the models and controllers as well as their parame-
disturbances, and changes in system parameters. In general,
ter vectors. The control problem is to determine suitable rules
complex systems operate in multiple environments which may
for switching and tuning these parameters to yield the best
change abruptly from one context to another.
performance for the given objective while assuring stability.
The speed and accuracy with which a controller responds to
The architecture described above is quite general and applies
sudden and large changes may be considered as a measure of
to both linear and nonlinear systems. However, in this paper it
its “intelligence.” From this point of view, “intelligent control”
is assumed that the system to be controlled is linear, primarily
is the efficient control of dynamical systems operating in
because the stability of the resulting switching and tuning
rapidly time-varying environments. While conventional robust
systems can be proved. Qualitative treatments of the nonlinear
control is restricted to sufficiently small ranges of variations,
case may be found in [3] and [20].
conventional adaptive control reacts too slowly to abrupt
changes, resulting in large transient errors before convergence.
Hence alternate control methods are needed. B. Choice of the Switching Rule
Multiple Models: When the environment of a system A natural way to decide when, and to which controller,
changes abruptly, the original model (and hence controller) one should switch, is to determine performance cost indexes
is no longer valid. If models are available for different for each controller and switch to the one with the
environments, controllers corresponding to them can be minimum index at every instant. However, since only one
designed a priori. During system operation, one has to identify control input can be used at any instant, the performance of
the existing environment to determine the correct controller. any candidate controller can be evaluated only after it used.
Such identification can again be achieved if a model for each On the other hand, the performance of all the identification
environment is known in advance. Based on these two ideas, models can be evaluated in parallel at every instant. Hence
the control strategy proposed is to determine the best model the indexes must be based on the performance of
for the current environment at every instant and activate the the models rather than the controllers, i.e., using identification
corresponding controller. errors rather than the control error . From an adaptive
Switching and Tuning: Since the number of available mod- control point of view, this rationale extends the principle of
els is finite, whereas the number of possible environments can certainty equivalence from tuning to switching.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 173

Fig. 1. A general architecture for control using N identification models and controllers.

The specific performance index proposed has the form known and constant, and is minimum phase. An LTI,
SISO, asymptotically stable, minimum phase reference model
is given with a bounded, piecewise differentiable input and
output . Its transfer function has the
(1) same relative degree as . The MRAC objective is to
where and can be chosen to yield a desired combination of determine a differentiator-free control input such that all
instantaneous and long-term accuracy measures. The forgetting signals in the overall system remain bounded, and the control
factor determines the memory of the index in rapidly error tends to zero.
switching environments and ensures boundedness of for
bounded . The reader is referred to [2] and [21] for more A. Structure of the Models and Controllers
detailed discussions on the choice of this index. Parameterization of the Plant: The plant is parameterized
The switching scheme (first proposed in [1] for the case as in conventional adaptive control. Define “sensitivity vec-
) consists of monitoring the performance indexes tors” as
at every instant. After every switch, a waiting period
of length is allowed to elapse, and then the
controller corresponding to the model with the minimum index (2)
is chosen (switched) to control the plant. The waiting period
is introduced to prevent arbitrarily fast switching. where is an asymptotically stable, controllable system
with , where is Hurwitz.
III. MODEL REFERENCE CONTROL Regression vectors are defined as
The effectiveness of the proposed methodology is illustrated
in this paper in the context of model reference adaptive
control (MRAC) of a LTI system. The structure of the models
and controllers are described in this section. The derivation (3)
of specific switching and tuning schemes and their stability
analysis are the subject of the next section. It is known [22], [23] that unique constants ,
The MRAC Problem [22]: The unknown plant to be con- , , exist such that the output
trolled is LTI, single-input/single-output (SISO) with control of the plant can be expressed as
input and output . Its transfer function is
, where and the coefficients of the
monic polynomials and constitute the unknown
plant parameter vector . It is assumed that belongs to a (4)
known compact set . Each plant in is further
assumed to satisfy the standard assumptions of MRAC [22], where represents the
i.e., the order , relative degree , and sign of are parameter vector of the plant.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
174 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

(a) (b)
Fig. 2. Structure of the identification models and controllers. The output of model Ij is y^j = p^T T
j ! , and the output of controller Cj is uj = j !. p^j and
j are constant vectors if the pair (Ij ; Cj ) is fixed and are functions of time if (Ij ; Cj ) is adaptive. (a) Identification models and (b) controllers.

Structure of the Identification Models and Controllers: The designed so that and belongs to [23]. Hence
structure of the models and controllers is shown in Fig. 2. The with no loss of generality we can assume that in
identification error of is defined as and has (7) in all the discussions that follow.
the expression Comment 1: In view of the importance of (7) in the proof
of stability, a qualitative description of its physical implication
would be helpful to the reader. denotes the actual input to
(5) the plant at instant , while denotes the control computed
by the controller . is the identification parameter error
where is the parameter error vector . The correspond- vector of , and is the corresponding control input error of
ing control input error for is defined as . Assuming that (ref. the previous paragraph), (7)
, where is the ideal control input to the plant relates at every instant , , and . If at any instant
which results in , and the ideal control parameter the control input is chosen to be , the last term in
vector is computed from as (7) is identically zero. In such a case, the equation directly
relates parametric uncertainty (i.e., ) to control input error
( ) when the th controller is being used. The last fact reflects
one’s intuitive feeling that a small identification parametric
error should result in a small control input error.
Comment 2: The problem stated above, and the results in
Section IV, can be readily extended to cases where a finite
number of distinct model structures are used, i.e., when the
sign of is unknown and only upper bounds on the order
and relative degree are known.
(6)
IV. FIXED AND ADAPTIVE MODELS
For fixed , is the ideal control parameter vector The rationale for using multiple models as proposed in
[determined using (6)] corresponding to . If is this paper is to ensure that there is at least one model with
adaptive, is tuned using a suitable tuning algorithm (e.g., parameters sufficiently close to those of the unknown plant.
gradient or least squares) and is computed at every instant The parameters of the models thus chosen may be either fixed
using (6) assuming that . Alternately, may be or adjustable. For the designer to decide which strategy to
dynamically adjusted at every instant as proposed in [23]. adopt, the rationale for using fixed models or adaptive models
The input to the plant is chosen by the switching rule to must be clearly understood.
be one of the elements of the set . It From a computational point of view, adaptive models are
can also be expressed as , where the control inefficient because of the need to update their parameter
parameter vector . vectors at every instant (see Fig. 2). A more significant
Error Equation: The following error equation is used in the drawback arises when the environment of the plant changes
stability analysis of the system: with time. If the environment is initially constant over a long
(7) interval, the adaptive parameters will converge to a
neighborhood of the parameter vector , representing that
While the relation was first given in [2], its derivation is environment. When the environment changes at this stage, i.e.,
included in Appendix A for the sake of completeness. The changes discontinuously, the change has to be detected, and
errors , which are defined in (27) in Appendix A, the parameters must be reset to their original starting
represent the deviation of the parameters and from the locations to identify the new rapidly. Fixed models do
certainty equivalence relation (6) [2], [23]. Hence not have these drawbacks, implying that the same strategy
for a fixed model and for an adaptive model with a certainty can be used in stationary and time-varying environments.
equivalence controller. For an adaptive model using dynamical However, since fixed models can represent exactly only a
adjustment of control parameters, tuning algorithms can be finite number of environments, adaptive models are needed to

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 175

improve accuracy asymptotically, as explained in Section II. a dramatic improvement in transient response when multiple
Thus, in the triad—speed, accuracy, and stability—we see adaptive models are used together with this switching scheme.
that fixed models together with switching yield the desired The number of models required to obtain a desired level of
speed, while adaptive models and tuning provide the desired transient performance depends upon the region of uncertainty
accuracy. The fundamental problem of stability consequently as well as the sensitivity of the transient response to
remains. mismatches in parameters [21].
In this section, we consider in succession four different
combinations of fixed and adaptive models. These correspond
B. All-Fixed Models
to i) all adaptive models, ii) all fixed models, iii) fixed
models and one free-running adaptive model, and iv) fixed The advantages of fixed models over adaptive models were
models with one free-running adaptive model and one adaptive discussed earlier in this section. Hence, only the stability
model with reinitializable parameters. Our objective is to aspects are discussed here.
determine the one which provides the best tradeoff between The problem differs from the all-adaptive case in three
performance and computational efficiency. As emphasized in ways. In the first place, to ensure stability with switching,
the introduction, for any one of these methods to be practically each plant in must be stabilized by at least one of the fixed
applicable, their stability properties must be well understood. controllers. This implies a sufficient density of fixed models
For the first configuration, the proof of stability has already in . Second, stability is not assured for arbitrary switching
appeared in [2]. The stability of the overall system when the schemes, since there may be controllers which destabilize the
other three configurations are used is proved in this section. given plant. Ideally, we would like to determine the class of all
The approach used here is distinctly different from that adopted stable switching schemes and then optimize from this set for
in [2]. the best performance. Since this is analytically intractable, we
seek a parameterized class of stable switching schemes which
A. All Adaptive Models provides sufficient flexibility for performance improvement.
The scheme described in Section II-B was shown in [1] and
In this configuration, which can be viewed as an extension [2] to offer such flexibility; in this subsection, we prove that the
of conventional indirect adaptive control, multiple adaptive same class results in stability with fixed models. Finally, while
identification models are used with initial parameter values a model activated at time according to this scheme has the
distributed over the set . The method was considered in minimum index at , it may deviate during the waiting period
detail in [1] and [2]. , resulting in inaccurate control inputs. Hence
Tuning Scheme: The identification parameters of and must be sufficiently small to limit such deviations. The above
the control parameters of the corresponding are tuned stability considerations are summarized in Theorem 2.
simultaneously. In particular, can be determined from Theorem 2: Consider the switching and tuning system de-
using either certainty equivalence or dynamical adjustment scribed in Section III-A, where the models are all fixed and
[23]. Any stable tuning scheme can be used, provided the the proposed switching scheme is used with ,
following conditions are satisfied [21]. and . Then, for each plant with parameter vector ,
Identification Conditions: there is a positive number and a function ,
1) such that if:
2) • the waiting time ;
• there is at least one model with parameter error
3) ;
then all the signals in the overall system, as well as the
4) (8) performance indexes , are uniformly bounded. Here
depends only upon , and also depends upon , , ,
Theorem 1 states the stability properties of the above and .
configuration. Its proof may be found in [2]. The proof is by contradiction and proceeds in three stages.
Theorem 1: Consider the switching and tuning system de- In the first stage, it is assumed that the state grows in an
fined in Section III-A, where the models are all adaptive. unbounded fashion. It is then shown that if the overall input
Assume that the tuning scheme is as given above. Let any error is sufficiently small compared to the state, the latter
arbitrary switching scheme be used such that there is some must be bounded, which is a contradiction. In the second
minimum interval between successive switches. Then stage, it is shown that if the waiting time is sufficiently
for any it follows that all signals in the overall small and the state is large enough, the above condition on
system are uniformly bounded. Further, all the identification is in turn assured by a corresponding sufficient condition
errors belong to and tend to zero asymptotically. on the identification errors. The latter condition states that
Finally, the input errors , the overall input error , if is used as the control input over an interval ,
and control error also tend to zero. the corresponding identification error must be sufficiently
The specific class of switching schemes described in small compared to the state over an interval . The
Section II-B (with ) was recommended in [2] for final stage in the proof establishes that the proposed switching
performance improvement. Simulation results demonstrated scheme, together with the fact that in Theorem 2 is chosen

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
176 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

to be sufficiently small, assures that the condition on is Since switching is forbidden over , it is not
satisfied if the state is large enough. certain that the input error will remain small over this interval.
The proof makes use of two lemmas, Lemmas 1 and 2, However, it can be imagined that if is sufficiently
which are stated and proved1 in Appendix B. small, must remain close to over the interval
Proof: From the results of [22], the overall system can . It is straightforward to show [21], using (13),
be represented as the mean value theorem, and the assumed unboundedness of
, that there exists a maximum time , proportional to
(9)
, such that if , then
where is an asymptotically stable matrix, (14)
is the input error at time , is the bounded reference input,
and . Since there is a minimum Inequalities (13) and (14) together imply (12). It thus remains
interval between switches and the individual control to establish (13). The latter is stated in terms of the input
parameters are bounded, (9) is a linear time-varying system error of the active controller. On the other hand, the switch-
with bounded, piecewise continuous coefficients and inputs. ing scheme—which determines which controller is active—is
Hence a unique solution exists on . expressed in terms of the identification error. In what follows,
Let the supremum function of the state be defined by a sufficient condition for (13) to hold is given in terms of the
identification error of the active model.
Lemma 2 in Appendix B is used for this purpose. The lemma
relates the identification error to
We assume that . With no loss of general- the corresponding signal . To apply the lemma, we use
ity we can assume that , . Then from (9), since the error (6), and the fact that for a fixed model,
can grow at most exponentially, there exist constants to equate (13) to the following:
and such that
(15)
(10)
The conditions required for Lemma 2 to be applicable are
Let . From Lemma 1 in Appendix B, it
satisfied, since the fixed models satisfy the tuning conditions
follows that if the input error satisfies the inequality
(30), there is a minimum interval of between
(11) switches, and the assumed unboundedness of assures that
it will be greater than the required lower bound in Lemma 2
for some finite time , then (9) will have bounded solu- after some time . Applying Lemma 2, it is seen that for (15)
tions, resulting in an immediate contradiction. The proof is to hold, it is sufficient if satisfies the following condition
completed in the next two stages by establishing this condition. for some :
Since belongs to the set for
every , (11) is clearly equivalent to the following:
whenever (16)
(12) where is chosen to satisfy
The above condition states that the input error of the active
controller must be small at every instant. A sufficient condition
for (12) to hold is derived in the following paragraph.
the maximum being taken over all the fixed models, and
Consider a typical switching interval , ,
is defined as
such that is activated at and deactivated at . Here
where the latter is to be chosen. For establishing
(12), is divided into three parts—the instant , the
waiting period , and the remaining interval Here the functions and , defined in the state-
(which could be empty if is deactivated ment of Lemma 2, are continuous and
at ). Let .
Equation (16) implies that a controller is activated at
instant only if the corresponding identification error is
sufficiently small compared to , over a sufficiently long
denote the set of instants where the choice of to control the
plant is made by minimizing the performance indexes. Assume interval prior to . Further, is retained over an interval
that over the set , the input error satisfies , after the end of the waiting period, only if
remains small over the interval . This
(13) completes the second stage of the proof in which (11), for
contradicting the unboundedness of , has been converted
1 J. Hespanha at Yale pointed out an error in an earlier version of the proof into (16) on the active identification error. According to the
of Lemma 2. sequence of arguments used thus far, if is small according to

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 177

(16), then is small according to (15), which in turn implies time , there exist uncertainty radii
that is small in the sense of (12), which finally implies the which assure stability for every plant in the set
boundedness of . The final stage of the proof is thus . Alternately, given ,
to establish (16) by invoking the properties of the switching , and , and a compact set to which the plant parameters
scheme. belong, the number of fixed models, to be uniformly
Let be the fixed model with parameters closest to those distributed in , can be chosen such that Theorem 2 holds.
of the plant. Let its parameter error , where Comment 3: In this section as well as in the following
is to be chosen. Its identification error satisfies sections, a nonzero value of is chosen to prevent switch-
, where is a constant such ing with infinite frequency. The proof of Theorem 2 is by
that . Letting , we obtain that contradiction and follows from the fact that the input is
small. This in turn is assured if the control error of
(17) each controller chosen, or alternately the corresponding
Choosing it is seen that (16) holds for the model identification error, is small. This is precisely what necessitates
for all . We will show that (16) holds for any arbitrary the use of a sufficiently small in the proof.
model , over the intervals in which it is actually used. The The authors believe that any finite would result
intuitive basis for this is that the switching scheme ensures that in stability, but a large value of would result in slow
switching and hence poor performance. Hence in the above
any model will be chosen in preference to at any instant
proof, our objective is only to show that an arbitrarily small
only if . The presence of the integral term
can be used. As in the case of other design
in ensures this will be true only if is smaller than
parameters in adaptive control theory (e.g., -modification,
over a sufficiently long interval prior to . By choosing
adaptive parameter for unmodeled dynamics), only the
sufficiently small, the above interval can be made larger
existence of can be assured. The choice of design
than , ensuring that (16) is satisfied.
parameters such as and is problem dependent and
From the definition of the switching scheme and the set ,
must be determined on the basis of the desired performance
it follows that
and the sensitivity of the plant response over the specified
(18) parameter set (see simulation results in Section V).
Comment 4: Fixed models cannot yield zero steady-state
To show that condition (16) is satisfied by model , we error for all reference inputs (except with exact parameter
assume the contrary, i.e., there is a and a matching). The magnitude of this error will decrease with the
such that parameter error of the closest model and increase with the
(19) input magnitude. Hence a large number of models may be
needed for a small error. The simulation results in Section V
In Appendix C, this is shown to imply that verify this. This in turn motivates the inclusion of a single
(20) free-running adaptive model in the following section.

where .
Here is a positive constant not depending on . For model C. Fixed Models and One Adaptive Model
, it follows from (1) and (17) that The previous section reveals that a large number of fixed
models may be needed to assure stability and good steady-state
(21) performance. This shortcoming may be overcome by including
a single free-running adaptive model which operates in
Choosing as parallel with the fixed models.
Switching and Tuning Scheme: The switching scheme is as
(22)
described in Section II-B. The parameters of are tuned
exactly as in the case of all adaptive models, i.e., (8) is
satisfied.
it follows from inequalities (20) and (21) that , In the case of all fixed models, the requirement that the
which contradicts (18), and thus establishes (16). Recall that parameter error of at least one model be small enough is used
the second stage in the proof demonstrated that the latter merely to ensure that its identification error is sufficiently small
assured that (11) holds, which in turn was shown in the first compared to the state (17). This condition is automatically
stage to ensure the boundedness of . Since the latter was satisfied by an adaptive model after finite time since its
assumed to be unbounded to begin with, a contradiction arises, identification error grows at a slower rate than the state of
and the theorem is proved. the system [2]. Hence there is no need to have a minimum
In light of the proof given above, Theorem 2 has the number of fixed models to ensure stability. Further, stability
following interpretation which is intuitively appealing. Given is independent of the parameters of the switching scheme.
fixed models with parameters , and a switching Finally, the magnitude of the steady-state error is independent
scheme with parameters , , and , there is a waiting time of the number of fixed models, and hence the latter need be
associated with each such that for every waiting chosen solely to improve the transient response.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
178 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

The stability properties of this configuration are summarized D. Fixed Models with One Free-Running and
in the following theorem. One Reinitialized Adaptive Model
Theorem 3: Consider the switching and tuning system de- It is commonly accepted that the convergence time of an
scribed in Section III-A with fixed models and adaptive model will be large for large initial parametric errors.
free-running adaptive models, where the latter are assumed Hence in the configuration described above, a large number of
to satisfy the identification conditions (8). Let the switching fixed models may be needed to keep the transient response
scheme described above be used with , , and . under control until the adaptive model has converged. If the
There exists a such that if , then all fixed model which is closest to the given plant is assumed to
the signals in the overall system, as well as the performance be known, faster convergence can be obtained by initiating a
indexes , are uniformly bounded. Here depends new adaptive model from the location of the former. The same
only upon the set . objective can be achieved on-line by starting adaptation from
Proof: The proof is a simple application of that given for the location of each different fixed model that is successively
Theorem 2. Let be the maximum waiting time computed chosen by the switching scheme. Either new adaptive models
in the same manner as in Theorem 2, and let . may be started from these locations, or a single adaptive model
Let be the corresponding minimum parameter may be reused by resetting its parameters to each new location
error required for stability, and as before. Since (i.e., fixed model) that is successively chosen. The second
is a free-running adaptive model, it follows from the results of method is obviously preferable since only one extra adaptive
[2] that , i.e., it grows at a slower rate than model is needed, and this is the one that we recommend.
. Hence there exists a large enough such that The free-running adaptive model is retained for the same
advantages as before.
(23)
Switching and Tuning Scheme: The switching scheme is
implying that satisfies (17). Further, from [2], the same as that described in earlier sections, with the
is also , which, from the argument used in the second reinitialized adaptive model included in the switching
stage of the proof of Theorem 2, further implies that process. The parameter vector of is determined as
follows: if a fixed model is activated by the switching
whenever scheme at any instant , then is reinitialized to the
(24) value of . Further, the initial condition of the integral in the
performance index of is also reset to the corresponding
where is the same as in (11). Hence, by replacing the role value of [ensuring that ]. Thereafter, is left
of the fixed model in the proof of Theorem 2 with , and to adapt until the next reinitialization. The tuning algorithm
resetting the time origin to , it follows that (24) also holds used to adjust between reinitializations is such that (8)
for the fixed models, and boundedness follows as in Theorem will hold if is left to run free, and further, the following
2. tuning conditions are always satisfied.
Even though a free-running adaptive model is present, Tuning Conditions:
Theorem 3 does not guarantee that the control error will tend to 1)
zero asymptotically, unlike conventional adaptive control. This
2)
is because it cannot be guaranteed that switching will settle
down at the adaptive model , which in turn is due to the
exponentially decaying integral term in the performance index. 3) (25)
Even though boundedness of all signals implies that the index
where is an equivalent state of the system [21], [22]
of will tend to zero, it is still possible that switching
comprising the output of the plant and the state of the
may alternate between two fixed models and without
controller.
encountering the adaptive model. This in turn implies that their
The inclusion of the reinitialized adaptive model does not
indexes and will be smaller than during the
alter the stability and convergence properties of the overall
switching instants; however, any increase in or during
system, as asserted in Theorem 4.
the waiting periods after switching will cause or to
Theorem 4: Let an additional reinitialized adaptive model
increase correspondingly. This in turn will cause consequent
be added to the configuration considered in Section IV-C
“bursts” in the control error . However, it is easily seen
and the switching and tuning scheme described above be used.
that the magnitude of these bursts may be made arbitrarily
Then the conclusions of Theorem 3 hold without change.
small by choosing to be sufficiently small. Based on this
The proof of the theorem is given in Appendix D. It is based
discussion, the following corollary is stated without proof. Its
on the proofs of Theorems 2 and 3, with additional arguments
proof may be obtained by adapting the proof of Theorem 3.
included to take into account the discontinuities arising from
Corollary 1: In Theorem 3, for every , there exists a
reinitialization. A qualitative discussion of the proof is given
such that if , then all signals in the
below.
overall system are bounded, and further
If there are no reinitializations after some finite time,
becomes a free-running adaptive model, and hence the proof
of Theorem 3 applies. Further, if is never chosen by the

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 179

switching scheme, stability is once again assured by Theorem contradiction [22]. It is assumed that is unbounded, and
3. The remaining situation to be considered is when hence there exist unbounded sequences and such that
is reinitialized infinitely often and becomes active between and , for . Let
reinitializations. We show that even in this case, the input and be the identification error and closed-loop estimation
error of the corresponding controller remains small error, respectively. The set is defined as that subset of
whenever is active, thus assuring stability. (i.e., with ) over which
As in the case of Theorems 2 and 3, this is shown by
demonstrating that the identification error satisfies (17).
If the input is used over , and there is no reinitial-
ization over the interval , then the argument used This in turn assures that ,
in Theorem 2 verifies (17) in this case. On the other hand, if for and some , and hence the input error satisfies
was reinitialized from a fixed model which was activated
at some instant , then the fact that the error (26)
of satisfies (16) over the interval is used
to assure the same for . It is then shown that can be chosen to be sufficiently small,
Comment 5: It can be shown in a straightforward manner and hence that decays exponentially with a minimum rate
that Corollary 1 will also carry over exactly. Further, the proof over the set . Over the set the state cannot
of the theorem only requires the presence of a model with grow faster than a maximum rate . By choosing large
sufficiently small identification error. Hence, Theorem 2 also enough, it is shown that the measure can be made
holds for the case when a reinitialized adaptive model is added. much larger than , implying that is smaller than
at the end of the interval . Since this is a contradiction,
Comment 6: The procedure described in this section boundedness of signals follows.
implements the two-stage identification process outlined in When multiple adaptive models are present, by identical
Section II for the general problem of control in time-varying arguments it follows that ,
environments. , for each . However, for the same reason as in the case
when no perturbations are present, it does not follow that each
E. Robustness input error satisfies the same condition; (26) is true only
for the actual switched input error along any permissible
It is well known that adaptive control methods which switching sequence of controllers that is used. In fact, using
assure convergence of the control error to zero in the ideal (7), it follows (see ref. Comment 1 and the derivation of (15)
case tend to become unstable when external disturbances in the proof of Theorem 2) that if is used as the input
are present. During the 1980’s numerous methods were con- over any subinterval of , then over
sequently developed to assure the robustness of adaptive that subinterval. Hence, considering the switching sequence of
controllers in the presence of bounded disturbances, time- inputs chosen by the switching scheme over , it follows that
varying parameters, and unmodeled dynamics. These methods over . Hence, boundedness of signals
involve modifications in the standard adaptive scheme such follows as in the case treated above.
as the inclusion of a dead zone in the adaptive algorithm, In the case of a single adaptive model, if unmodeled
projection of estimated parameters onto a compact convex dynamics are present and sufficiently small, it is well known
set in parameter space, and the use of -modification and - that modifying the normalization of the adaptive algorithm
modification in the adaptive algorithms. Since all the models to include a dynamic normalization term [24] will assure
used in the procedure outlined in this paper are either fixed robustness. When multiple models are present, the same pro-
or adaptive, one would expect the overall system to be cedure described above can be used to assure robustness by
robust under perturbations, if each model-controller pair is considering the actual switched input to the plant rather
individually robust. This indeed turns out to be the case. As than the individual control inputs .
in the single model case, the proofs for the multiple model In summary, the comments made in this section show that
case can be provided for each class of perturbations separately if each model-controller pair is individually robust to bounded
and do not differ significantly from the former. Due to space external disturbances or unmodeled dynamics, then the overall
limitations, we merely outline the principal steps involved in multiple adaptive model system is also robust for any arbitrary
one specific case, where all the models used are adaptive. The switching scheme with an arbitrary waiting period .
reader familiar with the robust adaptive control literature [22, Simulation results presented in [21] demonstrate that the
Appendix D] can recognize the principal arguments involved switching schemes proposed in this paper improve the transient
and can use the same procedure to conclude robustness in the response of conventional adaptive systems even when such
other cases as well. perturbations are present.
We consider the case of bounded external disturbances and
demonstrate that the parameter projection method assures ro-
bustness when the given compact set to which the unknown V. SIMULATION RESULTS
plant parameters belong is convex. In this section we show how the different strategies pro-
If is the equivalent state of the system [21], [22], posed for adaptive control using multiple models may be
the proof of boundedness for the case proceeds by used to improve the performance of conventional adaptive

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
180 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

control systems. From a mathematical perspective, there is with , where

little agreement on the term “performance,” and a variety of
indexes have proposed to measure performance. In the context
of the proposed adaptive systems using multiple models, a
rigorous demonstration that performance is improved relative
to conventional adaptive control systems involves choosing a
particular performance index and proving that for all possible
reference inputs, initial conditions for states and parameters,
etc., the value of the index for the former is less than that for
the latter. Since this is mathematically intractable at present,
we resort to computer simulations to demonstrate performance For each set , a comparison is made between conventional
improvement. adaptive control with a single adaptive model and the four mul-
The methods proposed have been simulated extensively, tiple model-based control strategies described in Section IV.
over a wide variety of scenarios, and reported in [21]. Due to The results are shown in Fig. 4, where each column shows
space limitations, only three sets of simulations are presented the responses obtained for a given level of uncertainty. In all
in this section. The first two sets pertain to transient response cases, the unknown plant has the unstable transfer function
, with parameters belonging to . The
improvement for a time-invariant plant, while continuous
initial conditions of the plant are zero.
performance improvement for a rapidly switching plant is
For the smallest region of uncertainty , a single adap-
considered in the third. tive model initiated at the center of provides satisfactory
In all simulations presented below, the plant to be controlled response [Fig. 4(a)]. Control based on a fixed model at the
has the transfer function , where same location does not even stabilize the plant. Even with
the parameters , , and are unknown. The control fixed models, there is a significant steady-state error
objective is to track the output of a reference model [Fig. 4(f)].
to a square wave reference input When the region of uncertainty is increased to , one
with unit amplitude and period 10 units of time. The least- adaptive model yields an unacceptable response [Fig. 4(b)],
squares method [25] was used to tune the parameters of the and 27 adaptive models are needed [Fig. 4(d)] to obtain a
response comparable to Fig. 4(a). With only fixed models,
adaptive models.
even with the response is unsatisfactory [Fig. 4(g)].
Example 1: This example demonstrates the importance of
With 48 fixed models and one adaptive model, the response
switching in the proposed methodology. Let be a fixed improves markedly [Fig. 4(i)] and is significantly better than
model which is close to the plant to be controlled and with one adaptive model, but still not as good as that with 27
be an adaptive model which starts far from the plant. adaptive models. By merely including a second reinitialized
The transient response using conventional adaptive control adaptive model [Fig. 4(k)], a response comparable to the latter
with alone will be unsatisfactory. If is included, the is obtained.
system will quickly switch to it because of its smaller initial The same trend was observed for the region , except that
identification error, resulting in improved transient response. a larger number of models was needed in each case to cope
will eventually take over and bring the control error to with the increased uncertainty, as seen from the third column
zero. If a second reinitialized adaptive model is allowed to of Fig. 4.
The basis for the proposed use of multiple models in
start from , the system will switch to it immediately after it
intelligent control is that improved performance in rapidly
switches to , and fast error convergence is obtained. switching environments can be achieved by ensuring fast
In the simulation, the unknown plant has the unstable and accurate transient response in each of the individual
transfer function . has the transfer environments. This is demonstrated in the next scenario, in
function , and is started from which the same approach used in the previous example is
. The response obtained using alone is applied to a rapidly switching plant.
seen from Fig. 3(a) to be unsatisfactory. When is added, Example 3: The problem is the same as in Example 2,
the system quickly switches to it at time [Fig. 3(c)] except that the parameters of the plant switch periodically
and remains there until , at which point takes over. between constant values in the set at the end of every
Fig. 3(b) shows that the resulting tracking error is small over interval of 50 units. The transfer functions of the plant during
six successive intervals are: ,
the entire interval. When a reinitialized adaptive model is
, , ,
added, the system switches to it at [Fig. 3(e)], and the , and , respectively. The response
error converges rapidly to zero [Fig. 3(d)]. obtained using a single adaptive model is compared with that
Example 2: In this example, three increasingly larger re- obtained with a combination of 320 fixed models and two
gions of uncertainty concerning the plant parameters adaptive models as used in the previous example. No attempt
are considered. These are denoted by , , and , was made to detect plant parameter changes. The substantial

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 181

Fig. 3. Illustrative scenario of transient response improvement by switching between an adaptive model (Model 1) started far from the plant, a fixed model
(Model 2) located close to the plant, and a second adaptive model (Model 3) initiated from the fixed model. The plant is unstable with transfer function
Wp (s) = 0:5=(s2 + 2s 1). The performance index used is Jj (t) = e2j (t) + 5 0 e00:001(t0 ) e2j ( ) d . (a) Adaptive model only. One fixed model,
0 t
and one free-running adaptive model: (b) plant response, and (c) switching scheme, respectively. One fixed model, one free-running adaptive model, and
one reinitialized adaptive model: (d) plant response, and (e) switching scheme, respectively.

improvement in performance using multiple models is evident It is the opinion of the authors that the above methodology
from Fig. 5. has applications in widely different contexts. The same ideas
Comment 7: The stability of the overall system, when the have been applied successfully in robotic systems and have
plant parameter vector switches at intervals between values also been used to control nonlinear systems using neural
in the finite set and control is effected using fixed networks. The theoretical basis for the design of such systems
models with the same parameter vectors , has been proved is presented in this paper in the context of linear systems.
in [21]. In particular, stability proofs were provided for different
situations, in which either switching alone, or both switching
VI. CONCLUSION
and tuning, are called for to improve performance.
As stated in the introduction, the perspective adopted by
the authors is that intelligent control is merely the ability of APPENDIX A
the control system to operate successfully in a wide variety DERIVATION OF (7)
of situations. The main idea introduced in this paper is the
The closed-loop estimation errors are defined as follows
use of multiple models, each one of which corresponds to a
[23]:
different environment in which the plant may have to operate.
The assumption made is that if the plant is sufficiently close
to a model in some sense, the latter can be used to choose
the controller. The objective, consequently, is to determine
the most appropriate model at any instant, using a suitable
performance criterion based on the identification error, and to
(27)
activate the corresponding controller.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
182 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

0
Fig. 4. Response of an unstable plant Wp (s) = 0:5=(s2 + 0:5s 2) with different control architectures and different levels of parametric uncertainty. The
performance index used is Jj (t) = e2j (t) + 3 0t e00:001(t0 ) e2j ( ) d . One adaptive model: (a)–(c). N adaptive models: (d) N = 27 and (e) N = 125,
respectively; (f) N = 64, (g) N = 216, and (h) N = 512, respectively. N fixed models and one free-running adaptive model: (i) N = 48 and (j) N = 320,
respectively. N fixed models, one free-running adaptive model, and one reinitialized adaptive model: (k) N = 48 and (l) N = 320, respectively.

From this and the definition of the parameter errors and

and (6), it follows that (27) can be rewritten as

(28)

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 183

(a)

(b)
Fig. 5. Response of a plant with parameters switching between values in S3 every 50 units of time. The performance index used in (b) is
Jj (t) = ej2 (t) + 3 0 e00:75(t0 ) e2j ( ) d . (a) One adaptive model and (b) 320 fixed models, one free-running adaptive model, and one reinitialized
t
adaptive model.

Using (28) and the definitions of , (3), , and , the Proof: Since is bounded over the interval ,
input error can be expressed as the solution exists over this interval and
is finite. Let . By the
variation of constants formula, the solution for satisfies

Since is uniformly bounded, taking supremum on

both sides, we obtain that for
(29)
which is (7).

APPENDIX B
STATEMENT AND PROOF OF LEMMAS 1 AND 2
This Appendix contains the statement and proof of Lemmas
1 and 2 which are used in the proofs of Theorems 2–4. Lemma
1 gives a sufficient condition on the inputs for the boundedness
of the state of an asymptotically stable linear system.
Lemma 1: Consider a linear system described by the equa-
tion

where is an asymptotically stable matrix and . If

showing that is uniformly bounded.
is bounded over the interval and for
, where and Lemma 2 assures that if the identification error
of a typical model is small compared to the state over
a sufficiently long interval, then the associated signal
is correspondingly small compared to the state over that
interval.
with denoting the norm of , then the state The lemma assumes that the tuning scheme for model
is uniformly bounded. satisfies (25), repeated here for convenience:

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
184 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

From the hypothesis of the lemma and the monotonicity of

1) , we then obtain that
2)

3) (30)
Choosing , it is seen that these conditions are automat- (33)
ically satisfied by fixed models.
Definition: is the class of right continuous, piece- where we have used the fact [21] that
wise continuously differentiable functions with a minimum
interval of between instants of nondifferentiability.
Lemma 2: Let denote a typical fixed or adaptive model in (34)
the MRAC system described in Section III-A. Assume that
satisfies the tuning conditions (30), and there is an interval of From the mean value theorem, there exists a
between switches. Then, there exist nonnegative contin- such that
uous functions and , with
for every fixed and , such that if:
• ,
for any , , and ; (35)
• there are no switches in the interval ; Again, by the mean value theorem, if , there
• , where is de- is a between and such that
fined in the proof of the lemma, it follows that

Here, for fixed , as (36)

Proof: Let , where
. Then from (3), From inequality (34) and the second tuning condition in (30),
, where denotes the th derivative we obtain that
of .
In the first part of the proof, we prove by induction that
there are positive constants , depending continuously
on and tending to zero as , and a continuous
function , such that
if , then
Using this, the definition of , and equations (10), (33),
(35), and (36), it follows that
(31)
In the second part of the proof, the above inequality is used
to prove the lemma.
Choosing , inequality (31) is seen to be true for
from the hypothesis of the lemma. We assume that it is true
for some . For , we choose
and let . Using
the induction hypothesis and the second condition in (30), we
obtain by partial integration that

The induction step follows since was

arbitrary. This proves inequality (31).
For the second part of the proof, let and
. Using inequality (31) and the same arguments
used to prove inequality (33), it follows that

(32)

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 185

. Consequently, there exists an interval

of the form

or
(37)

From the proof given in [2], the signal can be decomposed

as such that is continuously differentiable over . Without loss
of generality, we take . Using the mean
(38) value theorem, the right continuity of , and inequalities (10),
(39), (41), and (43), it follows that for every
Here is the first element of the control parameter vector
, and is the first element of . Hence the
first term is uniformly bounded. Further [2], the second term (44)
, with
The above, together with the continuity of over , further
implies that the sign of is constant over . Hence,
and the integral of satisfies
between discontinuities (39)

From (38) and the inequality (37), it follows that

which contradicts (40), thereby proving (42). Using (41) and

the fact , inequality (42) reduces to

(40)

if . Note from the properties of (45)

that both and are continuous functions of and
tend to zero as . The above holds for arbitrary . We where the function is defined as
now prove that if is chosen as

(41)
else.
where is the inverse of the monotonic function
, then satisfies is seen to be a continuous function of for
, . Further, for any fixed ,
as . From (38), (45), and the choice
of , it follows that for any
(42)
Note that as . To prove inequality
(42), we assume the contrary, i.e.,

(43)

for some . Since , successive (46)

discontinuities in are apart. Further, from the hypoth-
esis of the lemma, and there are no switchings in which proves the lemma.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
186 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO. 2, FEBRUARY 1997

APPENDIX C is continuous over the interval

PROOF OF INEQUALITY (20) and satisfies
The objective is to prove inequality (20), given (19). It is (49)
first shown that the latter implies that there exists a
such that then the corresponding signal satisfies
over the same interval. From the error (7) and the
(47) fact that for model , it follows that condition
(48) is satisfied if (49) is true.
From the tuning conditions (25) and (34), it can be checked For proving condition (48), we consider two possibilities:
that , for some . Using the either was not reinitialized over the interval
mean value theorem it follows from (19) that if , prior to its activation, or it was. In the former case, both
then for any and are continuous over the entire interval .
Then the argument used in the second part of the proof of
Theorem 2 applies, and it follows that (49) will be satisfied
if is chosen small enough and large enough. Hence,
by the argument in the previous paragraph, (48) is verified in
for some
this case.
Considering the second possibility, assume that
was the last instant at which was reinitialized. Let
be the fixed model from which it was reinitialized. This
implies that was activated at instant , and hence, by the
which proves (47).
proof of Theorem 2, it satisfies condition (16), i.e.,
It is given that . Using (1), (10), and (47),
it follows that (50)

Note that we may assume that and by

choosing sufficiently small in the proof of Theorem 2.
Over the remaining interval , , and are
continuous and the earlier arguments can be repeated to obtain
that
(51)

To combine the above two facts, we define a new model ,

operative over the interval , as the concatenation of
model over the interval and model over the
interval . The parameter vector of is defined as

Choosing to be the minimum of the values over all

the fixed models , , verifies (20).
and its identification error is . Hence,
APPENDIX D is continuous and satisfies over the
PROOF OF THEOREM 4 interval which contains the interval .
Moreover, satisfies (25) over this interval. Hence, applying
Proof: The proof consists of demonstrating that the input Lemma 2 (Appendix B) to and and using the
error of the reinitialized adaptive model satisfies definition of , it follows that ,
. From this, as in the proof of Theorem 2, we
whenever
conclude that (48) holds for this case also.
(48)
REFERENCES
for some large enough, where is as defined in the proof
of Theorem 2. Since the proofs of Theorems 2 and 3 ensure [1] K. S. Narendra and J. Balakrishnan, “Performance improvement in
adaptive control systems using multiple models and switching,” in
that the same is true for the rest of the fixed and free-running Proc. Seventh Yale Wrkshp. Adaptive Learning Syst., Center for Systems
adaptive models, boundedness follows as in the previous cases. Science, Yale University, New Haven, CT, May 1992, pp. 27–33.
As in the case of Theorem 2, this is shown via (16). Let [2] , “Improving transient response of adaptive control systems using
multiple models and switching,” IEEE Trans. Automat. Contr., vol. 39,
be any arbitrary switching interval over which pp. 1861–1866, Sept. 1994; also in Proc. IEEE 32nd Conf. Decision
is active. From Lemma 2 (Appendix B) applied to , it Contr., San Antonio, TX, Dec. 1993.
[3] K. S. Narendra, J. Balakrishnan, and M. K. Ciliz, “Adaptation and
follows that for the given in (48), there is an learning using multiple models, switching, and tuning,” IEEE Contr.
and an interval such that if Syst. Mag., June 1995.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.
NARENDRA AND BALAKRISHNAN: ADAPTIVE CONTROL USING MULTIPLE MODELS 187

[4] D. T. Magill, “Optimal adaptive estimation of sampled stochastic pro- [21] J. Balakrishnan, “Control system design using multiple models, switch-
cesses,” IEEE Trans. Automat. Contr., vol. AC-10, pp. 434–439, 1965. ing, and tuning,” Ph.D. dissertation, Yale Univ., New Haven, CT, 1995.
[5] D. G. Lainiotis, “Partitioning: A unifying framework for adaptive sys- [22] K. S. Narendra and A. M. Annaswamy, Stable Adaptive Systems.
tems—I: Estimation; II: Control,” Proc. IEEE, vol. 64, pp. 1126–1143 Englewood Cliffs, NJ: Prentice-Hall, 1989.
and 1182–1197, Aug. 1976. [23] M. A. Duarte and K. S. Narendra, “A new approach to model reference
[6] M. Athans et al., “The stochastic control of the F-8C aircraft using a adaptive control,” Int. J. Adaptive Contr. Sig. Proc., vol. 3, pp. 53–73,
multiple model adaptive control (MMAC) method—Part I: Equilibrium 1989.
flight,” IEEE Trans. Automat. Contr., vol. AC-22, pp. 768–780, 1977. [24] L. Praly, “Global stability of a direct adaptive control scheme which
[7] D. W. Lane and P. S. Maybeck, “Multiple model adaptive estimation is robust with respect to a graph topology,” in Adaptive and Learning
applied to the Lambda URV for failure detection and identification,” Systems, K. S. Narendra, Ed. New York: Plenum, 1986.
in Proc. IEEE 33rd Conf. Decision Contr., Lake Buena Vista, FL, Dec. [25] G. C. Goodwin and D. Q. Mayne, “A parameter estimation perspective
1994, pp. 678–683. of continuous time model reference adaptive control,” Automatica, vol.
[8] C. Yu, R. J. Roy, H. Kaufman, and B. W. Bequette, “Multiple-model 23, no. 1, pp. 57–70, 1987.
adaptive predictive control of mean arterial pressure and cardiac output,”
IEEE Trans. Biomed. Eng., vol. 39, pp. 765–778, Aug. 1992.
[9] R. L. Moose, H. F. Van Landingham, and D. H. McCabe, “Modeling and
estimation for tracking maneuvering targets,” IEEE Trans. Aerospace
Elec. Syst., vol. AES-15, pp. 448–456, May 1979. Kumpati S. Narendra (S’55–M’60–SM’63–F’79)
[10] X. R. Li and Y. Bar-Shalom, “Design of an interacting multiple model is currently a Professor of Electrical Engineering
algorithm for air traffic control tracking,” IEEE Trans. Contr. Syst. Tech., and Director of the Center for Systems Science at
vol. 1, pp. 186–194, Sept. 1993. Yale University. His research interests have included
[11] B. Mårtensson, “Adaptive stabilization,” Ph.D. dissertation, Lund Insti- stability theory, adaptive control, learning automata,
tute of Technology, Lund, Sweden, 1986. and more recently, intelligent control of complex
[12] M. Fu and B. R. Barmish, “Adaptive stabilization of linear systems systems using neural networks.
via switching control,” IEEE Trans. Automat. Contr., vol. 31, pp. Dr. Narendra is a fellow of the IEE and AAAS.
1097–1103, Dec. 1986. He serves on several national and international tech-
[13] K. Poolla and S. J. Cusumano, “A new approach to adaptive robust nical committees and editorial boards of technical
control—Parts I and II,” Coordinated Science Laboratory, Univ. Illinois, journals and has received numerous awards from
Urbana, Tech. Rep., Aug. 1988. different societies of the IEEE. In 1995, he received an honorary D.Sc.
[14] D. E. Miller and E. J. Davison, “An adaptive controller which provides degree from Anna University in Madras, India, was elected a member of the
Lyapunov stability,” IEEE Trans. Automat. Contr., vol. 34, pp. 599–609, Connecticut Academy of Science and Engineering, and delivered the Bode
June 1989. Lecture at the 34th IEEE Conference on Decision and Control.
[15] D. E. Miller, “Adaptive stabilization using a nonlinear time-varying
controller,” IEEE Trans. Automat. Contr., vol. 39, pp. 1347–1359, July
1994.
[16] R. H. Middleton, G. C. Goodwin, D. J. Hill, and D. Q. Mayne, “Design
issues in adaptive control,” IEEE Trans. Automat. Contr., vol. AC-33, Jeyendran Balakrishnan (S’94–M’95) received
pp. 50–58, Jan. 1988. the B.E. degree in electronics and communication at
[17] A. S. Morse, D. Q. Mayne, and G. C. Goodwin, “Applications of hys- the College of Engineering, Madras, India, in 1987,
teresis switching in parameter adaptive control,” IEEE Trans. Automat. the M.E. degree in electrical communication at the
Contr., vol. 37, pp. 1343–1354, Sept. 1992. Indian Institute of Science, Bangalore, India, in
[18] S. R. Weller and G. C. Goodwin, “Hysteresis switching adaptive control 1989, and the Ph.D. degree in electrical engineering
of linear multivariable systems,” IEEE Trans. Automat. Contr., vol. 39, at Yale University, New Haven, CT, in 1995.
pp. 1360–1375, July 1994. Subsequently, he worked for SGS Thomson
[19] A. S. Morse, “Supervisory control of families of linear set point Microelectronics, Inc., San Jose, CA, on architecting
controllers,” in Proc. IEEE 32nd Conf. Decision Contr., San Antonio, an MPEG-2 video decompression chip. Currently,
TX, Dec. 1993. he is with Adaptive Media Technologies, Inc.,
[20] K. S. Narendra and S. Mukhopadhyay, “Intelligent control using neural San Jose, where he is working on multiresolution algorithms for video
networks,” in Intelligent Control, M. M. Gupta and N. K. Sinha, Eds. compression. His research interests include adaptive and switching algorithms
New York: IEEE, 1994. for control, signal processing, and multimedia compression.

Authorized licensed use limited to: University of Waterloo. Downloaded on November 24,2023 at 01:09:17 UTC from IEEE Xplore. Restrictions apply.