Hydrological Sciences Journal

ISSN: 0020-6024 (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/thsj18

ON THE SYSTEMS APPROACH IN HYDROLOGY

VENKATESWARARAO VEMURI & NARASIMHAMURTY VEMURI

To cite this article: VENKATESWARARAO VEMURI & NARASIMHAMURTY VEMURI (1970) ON

THE SYSTEMS APPROACH IN HYDROLOGY, Hydrological Sciences Journal, 15:2, 17-38, DOI:
10.1080/02626667009493951

To link to this article: https://doi.org/10.1080/02626667009493951

Published online: 30 Dec 2009.

Submit your article to this journal

Article views: 5859

View related articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
 Bulletin of the International Association of Scien:$fic t-lydrology, X V , 2 611970

ON THE SYSTEMS APPROACH IN HYDROLOGY

VENKATESWARARAO VEMURI
Department of A A & ES
Purdue University
Lafayette, Ind. 47907
and
NARASIMHAMURTY VEMURI
Vemuri oari street
Tuni P . O . , “ A . P . . , India

ABSTRACT
There is a long felt need for a general theoretic frame work for hydrologic systems research which
imbeds dynamical, structural, spatial and behavioral aspects of modeling with predictive powers. Many
of these problems can be effectively studied by systems analysis. The process of system modeling for
large-scale, nonlinear, time-lag systems can be rationalized by suitably identifying and modeling
subsystems. When computers are used as modeling tools methods of formulating the problem, choice
of the computer used, and choice of performance criteria greatly influence the results. This in turn
imposes certain limits on the validity of computer simulated models. These aspects are discussed by
studying the nature of hydrologic systems and the nature of the associated inverse problems and
requirements for validating the model$.

INTRODUCTION

During the past decade study of the hydrologic cycle and its various components has
undergone a substantial change. At the turn of this century most of the research was confined
to a delineation of the components of the hydrologic cycle and a descriptive discussion of the
intervening phenomena. Since the 1930’s, however, some significant advances were made
in the quantification of hydrologic information and a great deal of work was done in this
area by many eminent hydrologists.
During the current decade, there has been a considerable upsurge of activity in a systematic
study of the theoretical and computational aspects of hydrological problems. Incidentally,
this period coincided with the International Hydrological Decade. This world-wide activity is
partly responsible for a movement to upgrade the status of hydrology and relieve the empiricism
involved in it. It is not an exaggeration to say that this goal has been achieved, at least in part.
The fact that many mathematicians and engineers from other disciplines are currently engaged
in research activities that traditionally belonged to the hydrologist speaks for itself the changing
nature of problems that hydrology can offer. Due to this widened interest in hydrology, from
without, the subject has been flooded by numerous papers. In particular, the concept of a
system, borrowed from electrical engineering, found widespread use by the hydrologist. The
deluge of material written on system theory (mostly by electrical engineers and mathematicians)
concerns general theorems but very little of usable techniques for obtaining practical results.
Early attempts to bring this new tool to the aid of the hydrologist also caused some problems
of communication which were accentuated by a proliferation of alien terminology. This resulted
in a gap-a gap due to a lack of understanding of the meaning of a system as it was conceived
and developed by its originators. This gap is widening as new generations of hydrologists
bring in new tools of research which are regarded so far as lying beyond the frontiers of interest
of the traditional hydrologist.

17
 So a couple of attempts have been made (Amorocho and Hart, 1964; Dooge, 1968)
along the lines similar to the one taken in this paper. This paper complements the above two
works. The presentation starts with some notions about a system and then attempts to classify
the nature and role of various classes of problems that a systems engineer faces. Even though
.many illustrative examples from hydrology are cited, the expose is valid for any engineering
system.

NATURE OF AN ENGINEERING SYSTEM

The essence of all engineering endeavors lies in the study of physical systems. Let us, for
moment, leave the concept of a system undefined and let us tacitly assume that the general
nature of a system can be understood intuitively. A basic problem in natural and behavioral
sciences is first to describe the behavior of a system in some convenient fashion, then use the
description to predict the future behavior, and finally apply this clairvoyance in some useful way.
In mathematical parlance the goal is to determine the state of a system at some future time
from a knowledge of its present or initial state. For example, we may wish to determine
tomorrow's weather from today's weather conditions. To determine the behavior of the weather
at a particular time in the future, generally we must be able to determine the weather at any
time in the future. This is a formidable computational job and the problem of avoiding the
generation of this proliferation of data is of fundamental importance.
The mathen1atician attempts to solve this problem by using very simple device--by using
differential equations. If x(t) represents the state of a system at any time t, one can use a
differential equation of the form

g == 0)

to obtain the value of x(t) at all future times and therefore at any particular time. In the majority
of cases, the basic equation is a nonlinear partial differential equation, or a functional equation
more complicated in form than the ordinary differential equation shown above.
Hydrology is the science of studying the properties of water and its movement in various
parts of the water cycle and the science of engineering and managing this water as a natural
resource. If this definition is accepted, problelTIS in hydrology can be divided into two classes:
The scientific or technological aspects of the problems and (2) the management or behavioral
aspects of the problem. Even though science and technology are vital to the continued
development of natural resources to serve mankind's needs, the social structure within which
technology is applied is equally vital. Hydrological systems are some of those that have to
operate within a constraining behavioral frame of reference and study of one aspect (say,
the technological) without due regard to the other (behavioral) aspect will be meaningless.

'TECHNOLOGICAL SUBSYSTEM

Problen1s typical of environmental sciences are characterized by their large size and are
responsible for a new genre of mathematical sciences-the study and control of large scale
systen1s. A significant practical aspect of the problem here is not even a question of control;
that is, far too ambitious. It is a question of learning enough about a system to permit the
development of a meaningful policy for operation. The general problem of operating a large
system with limited amount of time available for observation, data processing and implemen-
tation of control generates new kinds of mathematical questions that have not yet been precisely
formulated and certainly not resolved.
Uncertainty in the measurement process is another factor of fundamental importance in
the study of large scale systems. The problem of decision making in complex situations is

18
meaningful only if we kno\V' how to deal with this element of uncertainty during the model..
building phase.
A third problem of importance in techno-behavioral (technavioral ?) sciences is the element
of time scale. In many physical sciences, time periods for which mathematical functions apply
are necessarily short whereas months, years and even decades are typical time units in a majority
of problems in hydrosciences. Either the systems are heavily overdamped (such as rainfaII-
runoff systems) or they have long time delays (delay between investment and benefit).Very
often these time delays are functions of time making the governing equations very complex.
A fourth factor of mathematical significance is that the input (i.e., the exciting signals)
is not generally under the control of the observer and sometimes may not even be directly
observable.

BEHAVIORAL SUBSYSTEM

One of the foremost areas of research of fundamental importance, and ironically the least
explored, is the effect of behavioral structure within which a technological system operates.
The reason for this situation may be a lack of understanding about the nature of problems in
this area that are significant and susceptible for research. This observation is particularly true

la

---------)tt.

Fig. 1

1--~
1 ~·L-=~

Fig. Ie

Fig. 1 Configurations of a systenl.

19
 hydrology. The devious natural habits of water often ignore the jurisdictional boundaries
created by man and pose the need for collective action of both public and private organizations.
fact, scientific and technological capability to handle water~-water management needs- are
almost powerless unless translated by effective and adequate institutional arrangements into
significant social values.
Intuitively, a system is nothing but a collection of interacting components subject to various
and producing various outputs. It is often convenient to conceive of a system in terms
a block diagram-the" black-box" representation as shown in figure 1 a. Often the situation
is more realistic if the black-box has a variety of inputs and outputs as shown in figure 1 b.
In the study of atmospheric systems, geological, hydrological and other environmental systems,
it is useful to think in terms of diagrams such as the one in figure 1 c which brings clearly the
existence of internal subsystems whose behavior can neither be examined nor influenced directly.
The behavior of systems is not always exemplary: economic systems are subject to inflation
and depression; hydrological systems are subject to, say, floods; ecological systems are subject
to pests and drought. One of the n1ain tasks of an engineer is to control the behavior of systems
and to derive benefits from his ability to control.

TACIT ASSUMPTIONS UNCERTAINTY

USEFUL DEFINITIONS

The following definitions serve to introduce some concepts that one commonly uses in
system theory and are useful in the sequel.

Systeln
A system S is a set of ordered pairs of signals.

s= i = 1, 2, ... etc.

where the braced parenthesis indicate the set of ordered pairs. (Xi, Yi), for i I , 2, ... etc. with
an underlying functional relation characterizing the mapping of Xi (t) into Yi (t).
Essentially, a system S consists of three parts: the input signal x(t), the output signal
and the plant P.

Input
The members of the domain of S, i.e., the set {Xi (t)} are said to be the inputs to the plant P.

20
Output
The members of the range of S, the set constitute the output of the P.

Plant
The plant P of a system S contains a nlathematical characterization of the process that
relates the inputs and outputs.
This distinction between a system S and its plantP is very inlportant and many
erroneously regard the terms system" and "plant as synonymous. A together
its inputs and outputs constitute a system. To make this distinction more meaningful, any
physical entity whose behavior is required to be controlled may be regarded as a plant. A
combination of plant(s) or components that act together and perforn1 certain function may
then be regarded as a system. However free use of these two terms interchangeably is
common in the literature with no ambiguity to the discerning eye.

Charaterization
A plant P of a system S is said to be characterized by a differential operator (or an
operator, an integro-differential operator or a functional to 111ention
an auxiliary condition (such as an initial condition) if and pair
provides a solution to the associated equation and the and
pair is a member of S.
That is, a system (or characterization describes relation.

Function
A. function is

A signal as defined above nlay be regarded either as a function of tinle or as a set of ordered
pairs.

Functional
A functional F[.] is a single valued set of ordered pairs of scalar valued functions of one
real variable, . ) and scalars

F[.] ==
Whereas a function defines a mapping fron1 one point to another point, a functional defines
a mapping from a function to a point. A simple example of a functional is the pair

where a mapping from the function to the point whose value is defined the definite:
integral g(~)d~ is implied.

By letting

r g d(

21
 functional can be written in an alternate form

.F

Heuristic
\lc11""\,.11'U'''~'
is a rule of thunlb, strategy, simplification, or any other kind
drastically limits the effort in search for solutions of problems. All that can
useful heuristic is that it offers solutions which are good enough most of the time.

"Algorithnl
An algorithm (or algorithmic methods) is a decision procedure which guarantees the
solution sought given enough time.
A variety of physical problems can be studied effectively by adopting the systems point
of view. A basic advantage of such an approach is in the possibility for standardization of
analytical and computational techniques. Once a handful of techniques are mastered, it becomes
only necessary to extend the arguments to be applicable to any physical system. Furthermore,
the representation helps to identify analogous situations \vhich may otherwise escape
The extent to which systems approach can be profitably employed becomes clear
considering as an example the task of developing a generalized planning system to predict
study the future water requirements of a conlmunity.Various facets of such an effort
are diagrammatically shown in figure 2. Fronl this figure it is clear that the social structure
within which science and technology are applied is as vital as the technology itself. Modernization
water law, institutions related to water resources development and managenlent
and water resource are but a fe\v areas where meaningful
contributions can using the systems 't·'I,rf....
<:It... <:l''·h

ENGU\lEl:':RING PROBLEMS

Problenls of interest to a systems engineer can be classified as direct and inverse

Direct problems are characterized by a complete specification of the contents of
box in 1a and one is required to study or predict the response of the box to any
specified input. This problem is also cailed the analysis problem. From a mathematical point
of view an analysis problem, in general, constitutes the solution of a differential equation,
integral, integro-differential or some such functional equation.
In engineering analysis, the system and the excitation are specified and the response is to
be found. The precise nature of the excitation and response depends, of course, on the physical
area to which the systenl belongs. In the case of a distributed system, a field, the specification
must include both the distributed characterics of all points within the field and the geometric
location of the field boundaries.
The inverse problenl is much more complex. Here the response to a particular input or
inputs is known, but either the equations describing the process are unknown, or the inputs
themselves are unknown. Depending upon the specific circumstance, an inverse problem may
present itself as (a) design or synthesis problem, (b) identification or modeling problenl, and
(c) instnnnentation or control problem. While the direct problem generally has a unique solution,
if it has any solution at all, the same is not true for an inverse problenl. An inverse problem
often leads to a multitude of mathematically acceptable solutions out of which one physically
acceptable solution, if any such solution exists, has to be selected. Criteria for this selection
depend upon the nature of the inverse problem.
In a design or synthesis problem, the nature of the expected excitation and the nature of the
required response are specified and a system having this excitation-response relationship has
to be physically realized or designed as it is often called. For example, it is possible to have
many possible alternatives in the design of a penstock or surge tank that satisfies some stipulated

22
fall within the
performance requiren1ents. To choose one out of many possibilities it is necessary to establish
design criteria or design specifications. Perhaps it is necessary to nlinimize the cost, perhaps
is necessary to meet the peak load demands, perhaps the drop in hydraulic head should not
fall belo\v a certain limit. In each case a different design may represent an optimum solution.
Some times more than one design may satisfy a given criterion or no design may exist that
satisfies an the specifications. While powerful synthesis techniques exist in a few areas (for
example, electrical filter design in circuit theory), by and large the predominant approach is
still "cut and try" based on empirical formulas. Even this heuristic method may demand the
of computers if there exist several allowable alternatives in the design of a subsystem and
the optimal subsystem fails to be optimal when it takes its place in a larger system.
A typical example falling in this category is the development of a computer aided
methodology for the automatic design of waste collection and treatment Given
information about topography, flows, costs of pipe and excavation, the problem determining
minimun1 cost trunk sewer is a typical design problem. This problem can be posed and
solved by a mathematical programming approach.
The identification or modeling problem is somewhat different from the design problem and
perhaps most difficult of all inverse problems. An identification problem generally arises while
attempting to control the behavior of a physically existing system. Therefore the criterion for
selecting one out of many possible solutions is different here. The criterion to be adopted here
not physical realizability but physical plausibility. In general, an identification problem is
characterized by a specification of a finite set of observations on the input and output signals
of a system and the goal is to find a mathematical characterization of the While the
of particular design can always be checked, either by studying model or the
it is almost practically impossible to the validity of an identified model. The
reason this can be attributed to many possible causes. In an identification problem the
is not generally under the control of the observer. Furthermore,. the period of observation
limited and therefore the validity of the model is correspondingly limited to the
"\A~. ~u.~J'-'~.J of input signals which appeared within the observation interval. In other
a model cannot contain more information than is available in the data from which it derived
(Brillouin, 1962).
The task of building a rainfall-runoff model to a watershed is a typical example of the
identification problem.
The third type of inverse problem, called the instrulnentation problem, involves the determi-
nation of an excitation signal, given the systen1 specification and response. Such an instrumen-
tation problem may arise in the literal sense of the word as in the determination of the true
voltage existing at the input terminals of a voltmeter from a knowledge of the voltage registered
on the dial. An instrun1entation problem may also arise in the general context of an identification
problem. For instance, certain physical situations permit the application of an (additional)
input signal, called the probe signal, specifically for the purpose of parameter estimation. In
these cases, there are often constraints on the length of the observation time and the magnitude
of the disturbance to the system produced by the input signal. Design of a proper input signal
for these applications can significantly improve and influence the accuracy of the estimated
parameters particularly when observations are corrupted by noise.
For example, one proposed method of identifying the parameters of a ground water basin
envisages the use of explosives (detonations) at certain places and to study the disturbances
caused by these detonations on the transient elevation of the water table elsewhere in the
basin. The idea is to study the system's sensitivity (Vemuri et al., 1969) to an external disturbance
and to estimate the unknown parameters using sensitivity analysis. The immediate instrumen-
tation problen1 here is to find the optimum nature of the externally introduced disturbance.
Looking from another angle, an instrumentation problem is often posed as a control problem.
In a control problem one is interested in the determination of the characteristics of an external
signal that drives a system from a given initial state to a desired tern1inal state. This control
problem formulation has received a great deal of attention and all other aspects of an inverse
problem have been studied in the general context of a control problem.

24
 Essential elements of a control problenl can be listed as follo\\1s:
A mathematical model of the system to be controlled;
A description of the desired output of the systenl'
A set of admissible inputs or controls;
A performance or cost functional which measures the effectiveness of a given control
criterion.
A typical example of a control problem in hydrology is the task of determining a pumpingl
recharge to\vards the optimum management of a ground water basin,

MORE ABOUT THE IDENTIFICATION PROBLEM

Characterization of a system and identification of a system are t'NO fundanlentaI problems

of system theory, Broadly speaking, characterization is concerned with the setting up of various
classes of mathematical models of physical systems. There are t\VO principal lines of attack
which are in vogue to characterize a system: the dynanlical (or differential) equation approach
and the kernel (or the weighting) function approach.
In the dynamical equation approach the unknown plant is assumed to be characterized
a known set of dynamic (very often, differential) equations of a given order but with unknown
very constant) coefficients. This assumption regarding knowledge of the form. of
the differential equations is, of course, restrictive. Nevertheless, in practice, this information
can often be gained from physical considerations. State variable and transfer function methods
belong to this class.
The kernel function approach starts without making any assumptions of linearity or of
rhln<;ln-'\I~<;ll '-'4
1..1UIIVli'" of knovvn form. The problem is thus nonparanletric.A characterization
sought in the time domain. In this approach, the unknown system (or plant) is
characterized by an analytic function over a function space and represented by the associated
power series (also known as Volterra) expansion (Balakrishnan, 1963). Identification of impulse
response (unit hydrograph) essentially belongs to this class.
Whatever may be the method adopted to characterize a system, the computational process
involved in solving the identification problem can be represented pictorially as shown in
figure 3. It is evident from this figure that the problenl itself divides into three parts.
Determination of the form of the model (i.e., selection of the model dynamic equations)
and isolation of the unknown parameters.
(b) Selection of a criterion function by means of \vhich the "goodness of fit" of the model
responses to the actual system. responses can be evaluated.
(c) Selection of an algorithm for adjustment of the parameters in such a way that the difference
between model and system responses, as measured by the criterion of (b) above, are
minimized.

SELECTION OF THE MODEL

The particular model (or class of models) that is chosen in a given situation depends upon
a variety of considerations. Some of them are briefly summarized below.

a) Input
The input signal nlay be deterministic or random. Some times, the input signal may consist
of a naturally occurring process added with a probe signal-the latter being intentionally
introduced by the designer. Proper design of this probe signal belongs to the area of instrumen-
tation referred to in the previous section. Some models can be theoretically identified only

25
 A

26
when the input is white gaussian noise. If the actual input process does not meet this requirelTIent,
such a model cannot be used for identification purposes. However, on some occasions, a white
gaussian noise which is uncorrelated with the actual input may be used as a probe signal.

In service/out of service
The kinds of inputs the system will be subjected to and the models that can be used, depend
to a great extent on whether the system is in service or not. If the plant can be put out of service,
at least temporarily, almost any input can be used and hence any theoretically acceptable
model would be adequate. In some naturally occurring processes, it is aln10st impossible to
put the plant out of service and the identification has to be carried out with the actual input with-
out interrupting the process. Identification and control of such a process has to be done" on-
line".
Black box versus gray box
A truly black box situation is one in which nothing is known about the plant or the nature
of the noise corrupting the output signal. In general, the engineer has some a priori knowledge
the system and so the black box is really a gray box. It is evident that both black-box and
lTIodeling problems can be divided into three parts as listed earlier.

Stable and unstable

Certain processes are unstable and many identification
sta ble plant. It is not true that unstable are useless systen1s.The IJ~\.-"V'.LV"-'.L r'\r"hIIA-n·H.~
involved in the identification of an are more complex.
be as a kind of instability in a system.
In to the considerations, the nl0dels chosen to characterize a must
lTIodifiable and theoretically identifiable.

SELECTION OF PERFORMANCE CRITERION

Formulation of a problem and selection of a particular method to characterize a system

constitute only a first phase in the total problem solving effort. The task of approximation and
conlputation is an equally important phase. Whatever the category of the particular problem
rnay be, the first step in the general formulation of a computational strategy is the selection
of an objective. That is, we set some goal to be achieved by our process or system through the
application of a properly selected influence policy. Usually an objective is specified as the
acquisition of some desired state for the process. A constraint is a limit or a required charac-
teristic deliberately imposed on the system for any reason. One question that naturally arises
in this connection is whether or not means for influencing the process under the imposed
constraints are sufficiently strong to allow the achievement of the specified objective
(controllability). If such means exist, then we have a properly formulated problem.
In general, there exist a number of ways in which the objective may be achieved. Within
the set of possibilities, taking into account all the imposed constraints, one may wish to choose
systematically the best approach with respect to some perfonnance criterion, cost function or
function, as it is variously called. The nature of the criterion chosen, that is the method
measuring the" goodness" of a policy, obviously depends upon the particular system under
study. Some of the general considerations useful in measuring the goodness of a policy are
Transfers of state must take place within a reasonable period of time.
The energy necessary to effect the changes in state must be constrained.
The expenditure of power required to effect changes must be constrained.

27
 Consider, for example, the problem of selecting two parameters K and T in order to optimize
the performance of a string of pumping wells in a ground water basin. The paran1eter K could
characterize the pumping" thrust" and T some other parameter. In this context, it is conceivable
to define optimum performance as that one which will minimize the total energy consumption
and the time T required to achieve a desired state. This perforn1ance criterion can be written as
·"T

=k1T+kzl dt
." 0

where T,K) represents the deviation of the computed water level h(K, T, t) from a desired
level and kl and k2 are some constants. For each set of parameter values K and
J assumes a new value. The goal of the optimization procedures is to seek computational
methods for automatically adjusting K and T in order to minimize J. In certain cases it is
desirable to find a time function, such as a pumping modulation program. K(t), rather than a
fixed value of K. In ground water studies, this latter situation usually arises when one is
interested in determining the" best" pumping program. In such cases, the problem becomes
one of fllnctional optinlization.
There is a wide choice of criterion functions available. A particular function will be chosen,
over the others, on the basis of mathen1atical and engineering considerations. Quadratic
functions, such as the integral square error criterion

h) = J: -h dt

are usually chosen because of certain mathematical advantages. When using

criterion function, it is important to understand its properties. For example, Jise in equation
a in the output space since it depends on the desired water table elevation
and the computed value h(K, T, t).However, it is an ordinary /lulction of the parameters.
insofar the parameter optimization (or parameter identification, as it is often
called) is concerned, use of the above criterion allows solution to proceed as in the minimization
(or maxirnization) problem of ordinary calculus which is concerned with functions, rather
than as in the problem of the calculus of variations which is concerned with functionals. This
is an extremely important point to remember.
It is not always necessary, nor is it desirable, to choose the integral square error criterion.
This criterion· gives a measure of the deviation of the actual water table elevation from the
desired elevation integrated over the entire duration of observation. At the end of a particular
experiment, this criterion gives a number which will be large if errors persist over an interval
of time. Indeed, any function which satisfies the requirelnents of a distance in the appropriate
space can be used as a performance criterion. For example,

"T IhD(t) - h (K, t, r)1 dt

Jo
sup jhDCt) - h(K, t, (1
K

are all equally valid performance criteria.

The mean square error criterion, J ms , is relatively easy to handle mathematically. It applies
and is particularly useful when input(s) are statistical in nature. On the debit side, J ms is useful

28
only when the systen1 is stable with bounded inputs. System identification based on the J ms
criterion often leads to lightly damped higher order systems. The J ms criterion penalizes large
errors much more severely than small ones and is relatively insensitive to parameter changes.
Owing to this fact, the Jn~s criterion is not highly recommended for use in parameter identifi-
cation procedures. However, while solving a design problem, relative to a particular performance
criterion, the system should be designed so as to minimize degradation of performance
(criterion) due to parameter variations and under these conditions J ms is highly useful.
The integral square error criterion, Jise, has a close resemblance to J ms and essentially
has the same advantages and disadvantages as J ms .
A performance criterion which provides increased sensitivity is the integral oj' the absolute
value of the error Jiae. This criterion weighs large errors less heavily and small errors more
heavily than Jise. However, this is computationally more difficult to implement.
There exist a variety of tailor-made criteria to suit a given situation. In the case of static
optimization problems, it is possible to define an instantaneous criterion function, such as

J(h D, h(t)) == [hD(t)-h(K, t, r)]2

rather than one which depends on the integration over a fixed interval. This kind of optimization
can be carried out continuously since the effect of a change in the parameter is reflected
immediately in a change in the criterion function. Static optimization problerTIs of this kind
are excellently suited for analog computer solution.
In dynamic optimization problems, evaluation of the performance criterion, in
constitutes a major cornputational process. If the minimization of J is based on the n1ethod
of gradients, then evaluation of J not only requires a knowledge of the solution of the dynamic
equation of the system but also that of an additional one called the derived equation. For
example, in the sensitivity analysis method (Vemuri et al., 1969), this additional equation
called the sensitivity equation which is linear and posed as an initial-value problem.
If dynamic programming method is used, this derived equation is the adjoint vvhich
is posed as a final-value problem (Vemuri and Karplus, 1969).

SELECTION OF AN ALGORITHM

Techniques useful in the minimization of J (i.e., fitting of nlodel to a given set of data) fall
in seve:ral classes.

. Analytical Methods
In some cases, analytical expressions for the unknown parameters can be derived. This is
true, for example, for least squares estimates with certain assumptions on the nature of
disturbance or noise entering figure 1.

2. Search Techniques (Wilde, 1964)

Are useful when the number of unknown parameters is small. In such cases, successive
values of parameters are selected, either at random or in accordance with some preselected
grid pattern in the parameter space and the corresponding values of J become J(1) J(2)
A simple comparison test is used to determine which value of J is a minimum.
Search techniques can be further subdivided into the following categories.
a) "Brute Force " Methods
The most obvious method of finding the optimal set of parameters is to discretize all the
parameters and compute J for all combinations of these parameters. This method is also known
as "exhaustive enumeration". The minimum value of J (and the corresponding parameter

29
values) are then selected from the output listing. It is evident that such methods are usefu
only when the number of possible parameter combinations is low.
b) EliTnination Methods
Elimination methods attempt to eliminate exhaustive search by confining attention only
to a preselected part of the rriterion surface. If the minimization procedure is visualized as a
mountain climbing operation, where the mountain is assumed to have a single peak, then
one can eliminate sections of the surface from further consideration by learning from the results
of previous search. A particularly interesting elimination technique based on the principle of
dynamic programming, is known as Fibonacci search (Wilde and Beightler, 1967).
c) Rando 111 Search
Another alternative to exhaustive enumeration is the use of random search in parameter
space (Brooks, 1958). In this case, successive trial parameter values are selected at random.
This method works even if the criterion surface has multiple minima. However, the number of
explorations needed to insure a sufficiently high probability of obtaining a value near the true:
mininlum may be extremely large. This method was used successfully (Hufschmidt, 1962) in
the design of a water resource system.
Since search techniques require repeated solution of the model equations, they are well
suited for hybrid computers (Bekey and Karplus, 1968), where rapid solution of differential
equations is performed on the analog portion 'vvhile programmed control and supervision of
the computations remain in the digital part. RandolTI search techniques are particularly useful
as a method of optimizing nonlinear system parameters.

Hill-Climbing Methods
Rather than searching over the whole range of parameters, the various climbing or descending
methods are based on finding the local properties of the criterion surface (Wilde, 1964,
Dawdy and O'Donnell, 1965).
a) Relaxation Methods
This method is based on searching for the mInImum along one parameter, finding the
local minimum, setting the parameter at this value, continuing the search along the second
parameter and so forth.
b) Gradient Methods
This is a direct approach to satisfy the system differential equations and constraints while
iterating on the control signals until each new iterate drives the value of the cost function to
a new minimum. If there are several local minima, gradient methods may have to be used
repeatedly with different initial parameter values until a global minimunl is found (Bryson
and Denham, 1962; Vemuri and Karplus, 1968).
The main advantage of the method of gradients is the relative independence of convergence
on the initial estimates of the unknown quantities. As the minimum is approached, the
magnitude of the gradients decrease and correspondingly, the convergence becomes slower.
Such techniques as quasilinearization (Bellman and Kalaba, 1965) and stochastic approximation
and Wolfowitz, 1952) may be viewed as extensions of the gradient methods.

SPECIAL METHODS FOR LINEAR SYSTEMS

In those special cases when the system can be assumed to be linear, a number of specialized
techniques are available for solving the identification problem. Consider, for example, a single
input, single output constant coefficient linear system whose state at any time t is represented

30
by the solution of the vector differential equation

.x=Ax+Bu = 0) = (1
Solution of equation (13) can be written as

(t) = cP(t- + ," ep(t-r)' u dt

",' to
where lJ>(t to) is the system transition matrix, defined by

(j> (t -
The identification problem of linear systems consists of estin1ating the elements of the
matrix A or equivalently the transition matrix lJ> from noisy measurements on the system
response Y given by
y = ex Du
We shall examine some of these techniques.

Weighting jllnction identification

For simplicity, consider a scalar case of equation (13) in which the response can be described
the convolution integral

h u

where the systen1 welgllltlnlg function. If the input is a unit impulse, u(t) then

h (t - r)' dr = h
o
Hence, the weighting function can be obtained, at least in principle, by measuring the impulse
response of system. In practice it is not possible to create an ideal impulse, so, a unit step
response
,rtt

= Jo h(r)dr
is first obtained and differentiation on both sides yields h (t), the impulse response.
The unit hydrograph (Sherman, 1932; Nash, 1957, 1959, 1960; and Dooge, 1959) is
essentially a weighting function of a watershed. Even though this method proved to be. an
excellent tool, assumptions regarding linearity of the watershed and the impulse nature of the
function on which the theory is based are almost always violated in practice.
An alternate approach to the "continuous" unit hydrograph method is the discrete unit
hydrograph method. Because it is more practicable to record rainfall-runoff data at discrete
intervals, equation (18) can be rewritten in discrete form

h u

which
h
h(1"')u +h u
= h(2 u +h u + h(O) u
31
from which the following recursive relations can be derived

h -u h
u

-u h -u(2

This procedure is called "numerical deconvolution and yields satisfactory results in many cases.

b) Correlation Methods

If enough data is available permitting statistical evaluation, the weighting function (or unit
hydrograph) may also be obtained from correlation measurements by using the Weiner-Lee
relation

= J h
represents the autocorrelation function of the input and represents the
tr JrH 11-( )lJ 111111cross correlation. In cases vvhere anlounts of data are available, this
"I.cU,I.c101.1,",,'A.l nlethod is better than the unit method.

Irtllo/.?Of1l11 Decomposition
In addition to the differential equation and function rerJre:seJotattlC)nS
linear system can also be modeled as collection filters whose are
orthogonaL (Mishkin and Braun, 1961 ; Dooge, 1968). That is

where (Pi form an orthonormal set. The advantage of such a choice is that the adjustment of
each parameter C'i is now independent of all others. If h (t) is nlodel response and the
observed response of the system, then the error is

If the mean square error is minimized, it is possible to compute the optimum values of Ci in
a single computer run.

:TransferFunctionMethods
A third method is based on generating the impulse response of a linear system by way of
functional transformations such as Fourier and Laplace transforms. Methods based on the
use of moments (Nash, 1959), harmonic coefficients (O'Donnell, 1960) and Laguerre coefficients
(Dooge, 1964) belong to this class.

32
METHODS FOR NONLINEAR SYSTEMS

Experience indicates that most of the component processes in hydrology are nonlinear
in nature.
In general, any method available for the representation of a nonlinear functional can be
chosen as a mathematical model for the identification of a nonlinear system. Some of the well
known models available to characterize a nonlinear system are:
a) State vector representation;
b) Wiener model using orthogonal expansions;
c) The functional power series.

When a nonlinear system is represented by a state vector model, such as

x= (1., t)

where both the order of the system as well as the form of I(x) are unknown, the state variable
representation has no advantage over the others. Ho\vever, in a more tractable problem in
which the form of the dynamic equations are known (as in equation 8) the value(s) of the
unknown parameters can be estimated by using techniques based on quasilinearization, sensitivity
analysis (Tomovic, 1962), numerical inversion of Laplace transforms (Bellman and
Kalaba, 1966), in conjunction with such computational algorithms as the gradient descent
nlethod or mathematical programming.
Using an orthogonal development of nonlinear functionals in series of Fourier-Hermite
functionals (Cameron and Martin, 1947), Wiener developed a model for the identification of
nonlinear dynamical systems. In order for the approach to \-\lork, however, the input must be
a white gaussian noise. In hydrology, most of the input processes are naturally occurring
processes and therefore utility of this method appears to be limited.
In 1887 Volterra introduced the functional power series

p(t p(t

for the expansion of the functional r(.) in terms of various homogeneous functionals. It has
been shown (Amorocho, 1963) that the linearity assumption is tantamount to truncating the
above series after the second term. When Volterra series is used, as a mathematical model,
the identification of a given nonlinear system reduces to the problem of determining the
Volterra Kernels (i.e., the generalized impulse responses) ho, hi, ... , etc. Since hi is a function
of i variables, the detennination of these functions for i ~ 3 becomes a formidable task. Some
initial efforts to apply this technique were nlade by some hydrologists (Amorocho and
Orlob, 1961).
A variation of the above method is to decompose the nonlinear time-lag system into two
subsystems: one a linear time-lag subsystem and the other a nonlinear no-time-Iag subsystem
(Jacoby, 1966).
A basic disadvantage with this power series method lies in the difficulty to interpret the
meaning of the functions hi in terms of the physical parameters of a system but the method is
general enough to be applicable both to linear and nonlinear systems. The transfer function
approach has several advantages in linear cases and attempts to generalize this method to
nonlinear systems (George, 1959) are not yet popular.

3J
 SIMULATION OF SYSTEMS ON COMPUTERS

MODELS VERSUS REALITY

While solving scientific and engineering problems it is important to keep in mind distinction
between theories and reality . Theoretical results are derived from certain axioms by using
principles of deductive logic. In physical sciences, the theories are so formulated that they
correspond, in some useful sense, to the real world. The separation between the conceptual
\\lorld (model) and the physical world (reality) should be clearly imprinted in the analysts mind
at all times.
It is important to remember that the word "model" is meant to imply a manifestation of
the interpretation that the scientist gives to the observed facts. Facts remain unchanged, but
models change. Thus the term" model" covers a vast variety of configurations. It can be a
laboratory mechanism, such as a Hele-Shaw model or an electrolytic tank analog, it can be
merely a system of mathematical equations for which a simple and straightforward laboratory
tool cannot be easily devised (such as Maxwell's equations of electromagnetic field theory)
and in the extreme case of abstraction, a model may exist only in the mind of a scientist (such
as Bohr's model of an atom). Similarly, a model may be very modest in its aims like the
A>Vl"'Ar·ln",.t3n·h~· it is expected to sum up.

SOLUTIONS SOUGHT

At this point it is instructive to make a distinction bet\veen scientific and

analysis. A mathematician, when confronted with a problem, worries about existence
U111quleness of solutions, validity of certain commutative multiplication or some such delicate
without concerning himself with the physical meaning of his
\,.{'-"'''''U\.Jt'-JJJ. Mathematics
an edifice built on a set of axioms and as long he conducts without the
axioms, lemmas and theorems a mathematician would be satisfied. The interests
of course, lie in the physical world; ho\vever, his attitude is more or less philosophical. The
of a scientist is to construct theories to explain the observed behavior of the physical
world around him and for this purpose he has to take into account all possible alternatives
to certain phenomena. The goal of engineering analysis is to obtain specific answers
to questions with a specified accuracy at a minimum cost in time, labor, and equipment.
Moreover, while an engineer would certainly like to obtain general solutions, he is generally
'willing to settle for nunlerical solutions.

ROLE OF SIMULATION

One of the earliest methods known to man to study the nature of physical systems is by
simulation. When confronted with the problem of studying the behavior of physical phenomena,
the engineer or scientist constructs a scale model following Newton 's Law of Similitude which
states that models must be dynamically and geometrically sitnilar to the structures that they
represent. In its primitive sense, a model is a small scale representation of the prototype in all
features which are pertinent to the problem under investigation. Often it is impossible to satisfy
all the requirements of similitude simultaneously. In this event, skill is required to design the
model in order to reduce the error by ignoring some of these requirements to a negligible
value.
A major drawback of the method based on similitude is a lack of flexibility. Electrical
analogs (Karplus, 1958) introduced an element of flexibility and played a very prominent role
in the study of hydrologic systems. With improved sophistication in the available hardware,
the term "simulation" also emerged with new meaning and depth. The word simulation now'
stands to mean trial and error experimentation in which the validity of a model is verified:

34
sensitivity to environment is explored; and vanatlon of performance to parameter
evaluated. With this broader meaning simulation has become a means of solving various
facets of the inverse problems described earlier. Where optimal programming methods are
not suitable because of the complexity of the model, simulation offers the possibility of
an experimental approach to the decision problem. When the models are very complicated
complex quantitative relationships that make explicit mathematical solution impossible or
difficult, the use of modern computers in conjunction with numerical approximation methods
offers a feasible alternative.

V ALIDATION OF COMPUTER SIMULATED MODELS

Experimental verification of physical models is done by subjecting them to the characteristic

signals that the physical system is likely to encounter and examining the response
Construction of scale models with precision and observing the results with precision is possible
in case of systems occurring in the physical sciences. In the case of behavioral systems it is not
easy to construct true models. If, for instance, the Inanagement of a ground water basin
influenced by factors such as socio-economic or political considerations, it is almost inIpossible
to simulate the total environmental systenl by using scale models. Mathematical nIodels
a greater degree of abstraction are used invariably to study the behavior of such systems. It
obvious therefore that the Law of Similitude in this respect would mean not the "'~~","'I",~T1,",··I'
.. q 11T'(7

of parts but similarity of Jogical process in the behavior of the systenI and its
In problems of inference from model to the real system, the question arises whether the
model under exanIination is a valid one, or to what extent can it be considered a valid model?
In the field of most of the efforts so far seem to have been directed in
complex models on COlTIputers and the of validation does not seem to have ~ar·a1'(,,:::>.ri
much attention. Basically, the validation of sinIulatedmodel does not pose a problenl different
in principle from the validation of any other scientific hypothesis, but the complexity that is
typically built into such models is so great that the process of validation is different.
Important features of a scientific method may be briefly summarized as follows. Careful
and accurate classification of facts and observation of their and correlation
(2) Discovery of scientific laws by the aid of creative imagination' Equal validity for
normally constituted minds' (4) Self criticism.
In order to test the validity of a theory or a nlodel we may use these criteria by putting
the following questions and trying to elicit positive answers. Does the theory permit careful
and accurate classification of facts and observance of their sequence and correlation? (2) Does
the theory provide scope for discovery of scientific laws by creative imagination? (3) Is the
theory equally valid for all normally constituted minds? (i.e., Are we sure that we will not
encounter a Maxwell's demon" as we extend the theory and try to generalize it?) (4) Is the
theory or the model capable of withstanding criticism? If we get positive ans\vers to these
questions, it means that our theory has stood the four-fold test and is scientifically valid.
Insofar as computer silTIualtion is concerned, we may say that it satisfies this four-fold test for:
(1) Preparing the model for sinIulation means writing a suitable program which in turn involves
processing the data classification and correlating the data. (2) Interpreting the conlputer
results and predicting the behavior of the real system involves logical inference and creative
imagination. (3) The logical mechanisnl of the conlputer ensures uniform results whosoever
feeds the same data thus obviating the differences between nornlal and abnornlal nIinds.
(4) The final and the most important touchstone of validity of a scientific hypothesis is it's
ability to sustain criticism. A scientific hypothesis must be capable of being disproved. Theories
or models should be subjected to tests capable of showing them to be false. As Karl Popper (193]
said, "So long as a theory withstands detailed and severe tests and is not superseded by another
theory... we may say that it has proved it's mettle or that it is corroborated;
Having come to the conclusion that a scientific hypothesis or model should be subjected
to the most severe tests aimed at disproving it, we should discover the relevant· tests that are

35
most appropriate to the problem at hand. A physical model is sought to be validated by ensuring
similitude in construction and verifying whether the changes in state produced in it are truly
representative of the theoretically expected values of the real system. Validation of behavioral
models is however not that easy because of the unpredictable behavior of the variables involved.
The need to construct behavioral models in hydrology occurs because of the inherent interaction
of scientific and technical variables with such factors as social, political, legal, and ethical
considerations. When such factors enter into the equations of optimization, then the validity
of the overall models can be judged from two points of view, namely, determining the truth
value of the results obtained and from a utilitarian point of view.
Determination of the truth value involves verification at three stages to be sure; namely
verification of the validity of the concept, validity of inference and verification of empirical
concordance. Several philosophical theories have been in existence which discuss this problem
of conceptual validity. Important among them are rationalism, empiricism, and positivism.
Rationalism holds that a model or theory is simply a system of logical deductions from a series
of synthetic premises of unquestionable truth values; not themselves open to verification. The
phrase '" synthetic a priori" has been coined by Kant (Russell, 1946) to describe the premises
of this type. Thus, for the rationalists, the problem of verification of a theory reduces to the
problem of searching for a set of basic assumptions underlying the behavior of the system
under study. At the other end of the spectrum is empiricism. Empiricism refuses to admit any
postulates or assumptions that cannot be independently verified. Empiricists ask that we begin
with facts and not with assumptions.

Inference is scientifically valid if the inferred could be drawn by every logically trained
mind and if the inference is drawn from known things to unknown things. Establishing the
validity of inference in simulation problems is of paramount importance.

Experbnental Verification

The final stage of verification is the experimental verification. In social sciences, one cannot
always conduct experiments under controlled conditions and at a particular instant of time.
Experimental verification at this stage means testing how far the results predicted by
simulation agree with actually observed values. Here one can use the non-parametric methods
of statistical inference to test the correspondence between fact and theory and establish
confidence intervals on the performance of computer simulated models. For instance, one
can use the simple"' chi square test" and decide whether there is any correspondence between
observations and predictions at a particular probability level.
Perhaps the best method of verification may be subjecting the model to multistage verifi-
cation: first, searching for Kant's synthetic a priori, secondly, attempting to verify the postulates
subject to the limitations of statistical tests and finally, testing the truth value and ability of
the model to forecast the behavior of the system under study.

CONCLUSIONS

Hydrologic systems, in a broad sense, consist of two (not necessarily distinct) subsystems
interacting with each other. One of these subsystems deals \vith processes which are
essentially hydrologic in nature and the other deals with behavioral aspects which enter
the scene as a result of human intervention.

36
2. Hydrologic systems are, in general, large-scale, time-varying, nonlinear, time-lag systems.
Variation with time may be due to human intervention (as in the effect of urbanization of
a watershed) or may be due to such natural factors as seasonal variations.
The fact that inputs are not generally under the control of the observer and also the fact:
that precise measurements are not easy to come by introduces an element of uncertainty
in the study and modeling of hydrologic systems.. By cleverly establishing measures of
uncertinty, it is possible to construct meaningful models and establish confidence intervals
on the validity of the models obtained.
4. Where some physical insight into the operating process of a system is easy to obtain, the
dynamic equation approach to model building is more useful' it gives an opportunity to
relate macro- and micro-properties of the system. Where almost nothing is known about
the physical process and nothing about the structure and order of the dynamic equations,
non-parametrization is more meaningfuL
5. Finally, if computers are used as model building tools, the algorithms and programs should
be conceived and written with the following points in mind. (a) The programs should
permit heuristic modification of an algorithm. (b) During simulation of a model, the
program should permit for "on-line" ADD and DELETE instructions that permits the
operator to substantially change the sequence of operations during the solution process.
(c) The program should contain built-in check routines to check the validation of the
computer output.

ACKNOWLEDGMENTS

The authors wish to express their gratitude to Dr. Walter B. Langbein of the U.S. Geological
and to the referees for their suggestions that substantially contributed to an improvement
in the quality of the paper.

REFERENCES

[1] AMOROCHO, J. and ORLOB, G. T. (1961), "Nonlinear analysis of hydrologic systems'" Water
Resources Center Contribution No. 40, University of California, Berkeley, California, November 1961.
[2] AMOROCHO, J. (1963), Measures of the linearity of the hydrologic systems" J. Geophys. Res.,
h

vol. 68, no. 8, pp. 2237-2249.

[3] AMOROCHO, J. and HART, W. E. (1964), A critique of current methods in hydrologic systems
H

investigation" Trans. Amer. Geophys. Union, vol. 45, no. 2, pp. 307-321, June 1964.
[4] BALAKRISHNAN, A. V. (1963), Determination of nonlinear systems from input-output data ",
H

Proc. 54th Meeting of the Princeton University Conference on Identification Problems in Communi-
cation and Control Systems, Princeton, N.J.
[5] BEKEY, G.A. and KARPLUS, W.J. (1968), Hybrid Computation, John Wiley, New York.
[6] BELLMAN, R. E. and KALABA, R. E. (I965), Quasilinearization and Nonlinear Boundary Value
Problenls, American Elsevier, New York.
[7] BELLMAN, R. E., KALABA, R. E. and LOCKETT, J. (1966), Numerical Inversion of Laplace Transforms,
American Elsevier, New York.
[8] BRILLOUIN, L. (1962), Scientific Uncertainty and Information, Academic Press, New York. 2nd ed.
(9] BRILLOUIN, L. (I 962a), Science and Information Theory, Academic Press, New York, 2nd ed.
[10] BROOKS, S.H. (1958), "A discussion of random methods for seeking maxima" Operations Research,
vol. 6, pp. 244-251, March 1958.
[11] BRYSON,A. E. and DENHAM, W. F. (1962), "A steepest descent method of solving optimum
programming problems" J. Appl. Meeh., vol. 29, no. 2, pp. 247-257.
[12] CAMERON, R.H. and MARTIN, W.J. (1947), "The orthogonal development of nonlinear functions
in Fourier-Hermite functionals" Annal. Math., vol. 48, pp. 385-392, April 1947.
[13] DAWDY, D. R. and 0 'DONNELL, (1965), Mathematical models of catchment behavior" Proc.
Am. Soc. Civil Engineers, J. of Hydraulics Division, voL 91, no. 4, pp. 123-137.
[l4]DoOGE, J. C. I. (1959), "A general theory of unit hydrograph " J. Geophys. Res., vol. 64, no. 2,
pp. 241-256.
[t 5] DOOGE, J. C. I. (1964), "Analysis of linear systems by means of Laguerre functions" J. SIAM
Control, vol. 2, no. 3, pp. 396-408.

37
I161 DOOGE,J . C . I . (1968), “The hydrologic cycle as a closed system”, Bull. Intern. Assoc. Sci. Hydrol.,
vol. 13, no. I , pp. 58-68.
[17] GEORGE, D.A. (1959). Continuous Nonlinear Systems, Res. Lab. of Electronics, M.I.T., Cambridge,
Mass., Tech. Report 355, Julyc?4.
[ 181 HUFSCHMIDT, M. M. (l962), Analysis by simulation: examination of response surface”, Design
of Water Resource SystEms, Harvard University Press, pp. 391-442.
[I91 JACOBY,S.L.S. (1966), A mathematical model for a nonlinear hydrologic system”, J. Geophys.
Res., vol. 71, no. 20, pp. 4811-4824, October 15.
[20] KARPLUS, W. J. (1958), Analog S i m u l a t i ~ ~Solution
: of Field Problems, McGraw-Hill, New York.
1211 KIEFER,J. and WOLFOWITZ, J. (1952), Stochastic estimation of the maximum of a regression
function”, Annal. Math. Stat., vol. 23, no. 3, pp. 462-466.
[22] MISHKIN,E. and BRAUN,L. (ed.), (1961), Aduptioe Control Systems, McGraw-Hill, New York.
[23] NASH, J. E. (l957), “The form of instantaneous hydrograph”, Publication No. 45, International
Assoc. Sci. Hydrolog;y, vol. 3 , pp. 114-121.
[24] NASH,J.E. (1959), Systematic determination of unit hydrograph”, J . Geophys. Res., vol. 64,
no. 1, pp. 111-114.
[25] NASH, J.E. (1960), “ A unit hydrograph experiment”, Rept. Hydraul. Res. Sta. and Res.,
Warlingford, Berks, Eng!md.
[26] O’DONNELL, T. (1960), Instantaneous hydrograph derivation by harmonic analysis”, Commun.
Surface Waters, Intern. Assoc. Sci. Hydrology, vol. 51, pp. 546-557.
[27] POPPER,K . R . (1961), The Logic of Scientific Discovery, Scientific Editions Inc., New York.
[28] RUSSELL,B. (1946), History of Western Philosophy (A chapter on Kant), Simon and Schuster,
New York.
[29] SHERMAN, L.K. (1932), “Stream flow from rainfall by the unit hydrograph method”, Engineering
Neicis Record, vol. 108, pp. 501-505.
[30] TOMOVIC, R. (1962), Sensitivity Anulvsis 0” Dynamic Systems, McGraw-Hill, New York.
1311 VEMURI,V. and KARPLUS,W.?; (1969), Identification of nonlinear parameters in distributed
systems via hybrid computation , Wuter Resources Resenrcb, vol. 5, no. I , pp. 172-185, February
1969.
[32] VEMURI, V., et cil. (1969), “Sensitivity analysis method of system identification and its potential in
hydrologic research”, U‘uter Resources Research, vol. 5, no. 2, April 1969.
1331 WIENER.N . (1949). Seminar i n Nonlinear N e t u w k s . Res. Lab. of Electronics. M.I.T.. Cambridge,

indafions of Oprin?izat6m, Prentice-Hall, Englewood

38