PROCEEDINGS OF THE

IEEE, VOL. 55, NO. 11, NOVEMBER

1885 1967

CONCLUSIONS based on the usual assumptions. This would be an auto-

An optimization technique has been presented for the matic means with an on-line control computer to revise the
economic allocation of realand reactive power applying the loss formula for currentsystem changes.
method of Lagrangian multipliers and a gradient method. The real and reactive power allocation option of the com-
The optimization procedure uses solutions of network puter program provides a method of planning economic
equations to account for the effects of transmission losses. reactive capability. It also provides the potential of on-line
A precalculated transmission loss formula is not required. control of reactive power scheduling.
The computer program provides the option to obtain REFERENCES
only a real power schedule or both real and reactive power [‘I M. J. Steinberg and T. H.Smith, Economy Loading of Power Plants
allocation. The real powerdispatch obtained by the method and Electric Systems.New York : Wiley, 1943.
[21 L. K. Kirchmayer, Economic Operation of Power Systems. New
presented provides the following benefits.
York: Wiley, 1958.
A solution can be obtained which reflectscurrent sys- H. H. Happ, J.F. Hohenstein, L. K. Kirchmayer, and G. W. Stagg,
“Directcalculation of transmission loss formula-11,’’ ZEEE Trans.
tem operating conditions as well as anychanges in the Power Apparatus and Systems, vol. 83, pp. 702-707, July 1964.
network due toline additions or outages. [41 H. E. Brown, C. E. Person, L. K. Kirchmayer, and G. W. Stagg,

The results include, in additionto the economic load- “Digitalcalculation of three-phaseshortcircuits by matrixmethod,”
Trans. AZEE (Power Apparatus andSystems), vol. 74, pp. 1394-1 397,1955.
ing of each generating unit, complete voltage and 15] A. F. Glimn and G. W. Stagg, “Automatic calculation of load
power flow information associated with the genera- flows,” Trans. AZEE (Power Apparatusand Systems), vol. 76, pp. 817-823,
tion schedule. 1957.
16’ R. Courant, Differential and Integral Calculus, vol. 2. New York:
The method presented can be used in planning studies to Interscience, 1949.
[’I H. A. Spang, 111, “A review of minimization techniques for non-
determine operating costs for both present and proposed linear functions,” SZAM Rev., vol. 4, pp. 34S363,October 1962.
generation and transmission facilities. The method can be [*I L. P. Smith, Mathematical Methodr for Scientists and Engineers.
adapted also to on-line control of generation.[”] This would NewYork: Dover, 1953.
[’I G. W.Stagg and E. L. Wizemann, “Computer program for load
provide the opportunity to check voltage and line loadings flow study handles ten-system interconnection,”Electrical World, August
before initiating generation changes. Security analyses can 1, 1960.
be performed also by simulating generator and line outages. [‘‘I G. W. Stagg, J. F. Dopazo, M. Watson, J. M. Crawley, G. R.
Bailey and E. F. Alderink, “A time-sharing on-line control system for
In addition, the elements of the coefficient matrices a economic operationof a power system,”Proc. Znstrum. SOC. Am., October
and /3 can be used to determine a transmission loss formula 1966.

Optimization in Engineering Design

ALLAN D. WAREN, MEMBER, IEEE, LEON S. LASDON, MEMBER, IEEE, AND
DANIEL F. SUCHMAN, MEMBER, m

spch problem are brie6y described. Two reasonably complexexamples,

lrPingthesemethods,Prepresented:tbetirstd&ibthed&ofm&-
water~rsyst~Pndtheseeoeddesc~tbe~of.rridebrndcrpt41
mer.

I. INTRODUCTION

Manuscript received July 5, 1967; revised August 30, 1967.

E NGINEERING designisviewed here as a three-
stage iterative process: 1) the selection or modifica-
tion of a structure for the system and the identifica-
tion of the design variables in this structure, 2) the assign-
A. D. Waren is with the Departmentof Electrical Engineering,Cleve- ment of numerical values to thesedesign variables, 3)
land State University, Cleveland, Ohio. At the time this paper was sub- evaluation of the resulting design and the decision as to
mitted he was with Bell Telephone Laboratories, Inc., Murray Hill, N.J.,
for the summer. which of steps 1) or 2), if either, must be repeated.
L. S. Lasdon is with the Operations Research Department and Systems Computer-aided design generally refers to the useof
Research Center, Case Institute of Technology (nowcalled Case-Westem- digital computer programs in the analysis stage and has
Reserve University), Cleveland, Ohio.
D. F. Suchman is with the Ordnance Division, Clevite Corporation, greatly reduced the time required to evaluate proposed
Cleveland, Ohio. design^.['^-^^^ Parameter valueselection,however, is an
1886 PROCEEDINGS
NOVEMBEROF THE IEEE, 1967

area in which computers have not yet found wide usage. vector of design parameters andz an independentvariable.
This paper deals with the use ofa computer systematically
to For example, in filter design, F is the insertion loss, e the
modify the parameters of a givendesign, generating a element values,and z the frequency ;in array design, F is the
sequence of new,improved designs at each step. directivity, e includes the transducer module driving volt-
If the computer is to aid in the assignment of design ages and parametersspecifying module location, and z is an
variables then it must be provided with somemathematical angular coordinate. In control system design, F could be
criterion, a measure of “goodness,” by which to evaluate the step response, e the variable gains, time constants, etc.,
alternate designs. Usually this measure is a scalar valued and z a time coordinate. It is assumed that f i e , z ) is con-
function of the design variables and one realistic design tinuous for all (e, z ) of interest.
goal for the computer is to maximize, or minimize, this For all of these systems,the design speafications require
measure subject to whatever constraints the variables must the response function to be above, or below, speciiied levels
satisfy. In many practical problems these constraints are for various ranges of z. For example, in the bandpass filter
simply upper and lower bounds, imposed by economic or design case, the insertion loss must lie above certain stop-
realizability criteria, on the parameter values. The com- band specifications and below other passband levels.
puter must also be provided with a method of generating Fortunately, it is not necessary to consider all values of z
new and superior designs. Mathematicalprogramming in theseranges. Since f i e , z ) is continuous in z, it is sufficient
algorithms can be readily used to accomplish this, given to consider a discrete set of values for z, {zi}, such that
the design criterion and the constraints. For example, it was satisfaction of the speafications at these points implies
shown by Lasdon and Warent41 that filters with bounded, satisfaction everywhere.
lossy elements could be designed by solving a nonlinear, Define
inequality constrained, optimization problem.
The purposes of this paper are twofold. The first is to F i e ) = F(e, zi)
extend the nonlinear programming formulation givenby
and define indexsets I, and I,, not necessarily disjoint, such
Lasdon and Wared4] toa wider class of designproblems. that I, contains those values of i for which Fi(c) must be
In particular, the designof planar electromagnetic and above specdied levelsai, and I, those for whichFi(e)is to lie
acoustic transducer arrays to meet, or exceed,speclfica- below other levels pi. Then the design problem becomes
tions on directivity is treated in detail. The second isto show that of finding a vector e satisfying
that the iterative use of optimization techniques can sigmfi-
cantly speed the design process. Most previous work on Fi(e) 2 ai i E I, (1)
optimization in engineering design has viewed its applica- Fi(e) I pi i E I, (2)
tion as a single-stage process culminating in an “optimal”
design. The viewpoint taken here is that optimization theory and
and practice is mostvaluable to design when used iteratively l j I ej Iuj j = l,...,n (3)
in much the same way as analysis programs are presently
used. Applied in this way a number of optimization prob- where the l j and uj are the lower and upper bounds on the
lems are solved, each of which may have different objective jth parameter. More complex forms of constraints can also
functions, constraints, or designvariables.Such anap- be handled. Again, for filter design, I, could be the index set
proach is necessaryfor a variety of reasons. It is usuallynot for stopband frequencies, ai the stopband specifications,
obvious how to formulate an optimization problem whose and I , and pi the corresponding passband quantities. In
solution will satisfy all the design goals. Exact designspeci- this case, the lower and upper element value bounds are
fications are seldom given; instead, there are conflicting chosen to guarantee physical realizability.
requirements and tradeoffs are usually necessary.Moreover, Of course, selecting values forthe ai and pi and the sets
the structure selected for the system may be inadequate or I, and I, may be difficult sincethe capabilities of the design
overlycomplex and costly; in either case, the structure structure maybe unknown and the design specifications
must be modified, new designvariables selected, and anew will probably not be phrased precisely as above. One way to
optimization problem solved. overcome these difficulties is as follows. An initial set of
This paper is divided into three main sections. Section points ( z i } is selected and specifications assigned at these
I1 formulates a class of engineering design problems as points. This determines an initial set of values for {ai, pi, I,,
inequality constrained optimization problems. Section 111 I B } .The correspondingoptimization problem is then solved.
discusses solution techniques for suchproblems. In Section The performanceof this optimal design is compared to the
IV two examples are presented. The first example gives a actual goals, and then a new set (ai, pi, I,, I,} is determined
detailed discussion of the design of a planar hydroacoustic so that the solution of the modified optimization problem
transducer array for a target-seeking system. The second will come closer to meeting the systemobjectives. This
deals with lossy, bounded element value filter design. approach leads again to an iterative procedure with the
designer closingthe loop, i.e., providing from his experience
11. PROBLEM FORMULATION
and intuition the new {ai,pi, I,, I,}.
A . Inequalities In some cases, more than one response function may be
Often, the adequacy of a design is determined by the required in order to adequately describe the system. An
behavior of a “response function,” f i e , z ) where e is the example of such a system is presented in Section IV-A.
WAREN ET A L . ; OPTIMIZATION IN DESIGN 1887

B. OptimalDesign
There may be many vectors e satisfying inequalities
(1H3) (i.e., many satisfactory designs), or there may be
none at all. In either case a “best” design can be meaning-
fully definedas one which maximizesthe minimum amount
by which the specifications are exceeded or, if it failsto meet
the specifications,minimizes the maximum amount by ,
whichtheyfail to be met. The solution to the following
nonlinear programming problem represents such an opti-
mal design.
Determine (e, [) to
minimize [ (4)

subject to Fig. 1. ( <0 and no constraints binding.

Fi(e) + 5 2 ai i EI,
F,(e) - [ Ipi iEZB (5)
l j I ej I u j j = l,...,n (6)

where [ is an additionalindependent variable.

For any vector (e, [) satisfying inequalities ( 5 ) and (6)
with [ < 0, the specifications are exceeded at all specified
values of zi by at least )[I (see Fig. 1). If [<O and at least one
of the constraints ( 5 ) is binding (satisfied as an equality), then w

151 is the minimum amount by which the specifications are

exceeded. If [> 0 and at least one of the constraints is
binding, then [ is the maximum amount by whichthe s p e d -
cations fail to be satisfied(see Fig. 2).
At a minimum of the above problem, at least one of the
L

Fig. 2.
INMPENDENT
.-
VARIABLE -2
i>0 and one constraint binding.
-
constraints ( 5 ) must be binding ; otherwise, [ could be re-
duced without violating any of the constraints. Thus, if
C. i7me Domain Example
min 5 <0, the solution to (4H6) defies an optimal design
which maximizesthe minimum amount by whichthe s p e d - Many design problems require that, for some s p e d e d
cations are exceeded. Similarly, if min [> 0, the specifica- input, the output of a system approximate a desired output
tions cannot be met at all points and the resulting design over some time interval. For linear systems the impulse
minimizes the maximum amount by whichthe specifications or step response is often specified. Such problems can be
are not met. Of couse, in either case, the parameter value readily converted to the form (4)-(6) if the approximation
constraints aresatisfied. is carried out in the discrete Chebyshev sense.[’]
More flexibility can be obtained by replacing the con- Let the impulse response of a system, with parameters e,
straints ( 5 ) by at time ti be Fi(c) and the specifiedresponse at ti be fi,
i= 1, . . ,m. If, in ( 5 )
+
Fi(e) wi[ 2 ai iEI,
Fi(e) - w i t I pi i EIB

where the { w , , ai} are non-negative weighting factors.

Using these,various portions of the response may be empha- then the solution of the optimization problem
sized more heavily than others, or the response can be
maintained at present levels insome intervals and improved minimize [
(e, i)
in others. For example, givenan initial filter designin which
the stopband rejection is satisfactory but the passband ripple subject to
too great, solving the minimization problem with
Fi(e) 2 fi - wi[
wi = 0 iEIA i = 1,2,*..,m
mi = 1 iEIB Fi(e) I fi + oil
will generate a design which minimizesthe ripple and main- yields the desired approximation. Note the inequalities ( 5 )
tains the present stopband levels. Values for the weights have been written above in a slightly different formto pro-
other than 0 or 1 may also be usedto advantage. vide a simple pictorial interpretation. Fig. 3 illustrates these
 1888 PROCEEDINGS OF THE IEEE, NOVEMBER 1967

m i

where r >0. Let

R = { x ) g i ( x )2 0}, i = 1, ,m
be the constraint set and
RO = {x(gi(x)> O}, i = 1, * ,m
its interior.’ Choose r1 > 0, x . E Ro, so that all constraints
are positive for xo, and consider the problem of minimiling
P ( x , rl) starting from x . subject to no constraints. Intuitive-
ly, one would expectthat a minimum of P ( x , rl) exists in Ro
since on the boundary of this set, i.e., for some gi(x)=O,
+
P ( x , rl)+ 00. Thus, the trajectory of steepest descent
leading from xo, a path on which P ( x , rl) is strictly decreas-
ing, cannot penetrate the boundary of the constraint set.
The minimizing point depends, of course, on the choice of
TIME, t ( K C ) rl, and is denoted by x ( r l ) . By the above arguments,
Fig. 3. Time domain example. x (. r__, )E Ro.
Now repeat this minimizing process for a sequence of r
values rl > r2 > r,>O. Each minimizing point x(rk)will
inequalities and shows that decreasing forces the system
response to come closer and closer to the desired one.Such be in Ro. Further, by reducing r the influence of the term
an approach has been used successfullyby Athanasso- m

~ ]the design of low-pass filters to achieve a speci-

p o u l o ~ [in r 1
i= 1
i/gi(Xh
fied impulse response.
which “penalizes” closeness to the constraint boundaries,
111. SOLUTION TECHNIQUES is reduced and in minimizing P , more effort isconcentrated
A . Penalty Function Approach upon reducing f ( x ) . Thus, the sequence of points x@’),
The chief diiliculty in solving the optimization probkm x(rz), . .,x ( r 3 can come closer and closer to the boundary
(4)-(6) is the nonlinear constraints (5). According to of the set R if it isprofitable, in terms of reducing f , to doso.
Zoutendijkl6I the best available algorithm for solving such One would expect that as r approaches zero, the minimizing
problems seems to be the s u m technique of Fiam and point x(r) approaches the solution of the original problem
M~Cormick.[’~,[*~ This method uses a “penalty function” (7), (8). Conditions on f ( x ) and the gi(x)which guarantee
to convert an inequality constrained problem to a sequence convergence are given by Fiacco and McCormick.[’] Such
of unconstrained minimization problems, for which efficient a method is particularly attractive when dealing with prob-
solution techniques exist. lemswith markedly nonlinear constraints sinceit ap-
The method is most easily explained if the original prob- proaches the solution value from inside the constraint set.
lem isreformulated as Motion along the boundaries of this set, which can be very
cumbersome when the boundaries have large curvatures,
minimize f ( x ) (7) is completely avoided.
s
For design problems, this approach offers a number of
subject to additional advantages. Suppose analytic techniques are
used to obtain an initial design. If the design satisfies the
g,(x) 2 0, i = 1,. * .,m. (8)
parameter constraints (6), then each subsequent design
The problem (4H6)can be easily converted to this form, satisfiesthese also. Moreover, since is reduced at each step,
with x =(e, e), as shown below : each P-minimization leads to animproved design satisfying
all the constraints on the design parameters. Hence, the
minimize5 method can be terminated whenever the current design is
(e, n
satisfactory. This may result in a significant time saving,
subject to since in many cases the true optimum need not be closely
+
F,(e) wi[ - ai 2 0 iEI,
(9)
approximated.
Often, for example in filterdesign,itisverydifficult to
+
-Fi(e) wic + pi 2 0 iEZB
obtain a “reasonably good” initial designsatisfying the
ej - l j 2 0 parameter bounds. If the initial designisvery poor, the
j = I,... , n (10)
uj - e j 2 0. For the problem (9), Ro is nonempty since for any e satisfying the
bounds, imay be chosen large enoughto make the other constraints
posi-
D e h e the function, tive.
WAREN ET AL.: OPTIMIZATION IN DESIGN 1889

minimization algorithm may converge slowly, if at all. Stage 2: Evaluate g(a) at the points a=O, 1, 2, 4, . . . ,
However, if it is relatively easy to generate a “good” but a, b, c, where c is the first of these values at which g has in-
infeasibledesign, then the following approach may be creased over its previous value. Then,
adopted. Replace the constraints (lo), due tothe bounds, by
a < a* < c.
e, - Sic I ui
If g ( 1)>g ( 0 ) divide the components of s by some factor, e.g.,
+
e, Sic 2 I , 2 or 3, and repeat.
where 6 , are non-negative constants whichscale the [ Stage 3: A quadratic polynomial is now fittedto the three
variable for each element.As soon as any of these constraints values g(a),g(b),and g(c),and its minimum a, is
is satisfied,the corresponding 6, is setto zero.
If an initial feasible design is available, the parameter
bounds (6) need not necessarily be incorporated in the P-
function. Since these are simple linear constraints itis
possible to minimize the P-function subject to them, often If
with substantial savings in time.[61The recent development
of an efficient method for linearly constrained minimiza-
tion[’] makes this approach an attractive possibility.
then a, is accepted as the estimate of a*. Otherwise b is
B. Unconstrained Minimization taken as a*.
In order for this penalty approach to beeffective, the In the particular application discussedhere, the finite
unconstrained minimization must be done efficiently. step size maycause a constraint to be violatedat some stage.
Recommended methods come from a class of recently de- The linear minimizer must incorporate logic to cut back the
veloped quadratically convergent procedures. The tech- step size in this case. Since the P-function is so constructed
nique used depends on whether first derivatives of F,(e) that a minimum, within the constraint set, always exists,
exist and can be easilycomputed from analytic expressions. then cutting the step back will eventually lead to a success-
If so, then the method of Fletcher and Powell seems to be ful one-dimensional minimization.
the best not, If then either Powell’s
IV. EXAMPLES
method[”] or Stewart’s[’31 extension of the Fletcher-Powell
technique, using finite differences, is preferred.[l41 A . Design of Planar Transducer Arrays
Acoustic and electromagnetic target seekingsystems
C . One-Dimensional Minimization
generally have two conflicting requirements placed upon
Each of the above minimizers reduces an unconstrained their directional characteristics. Let the system be a planar
minimization problem to a sequence of onedimensional transducer array which must fit within a specified area;
minimizations which proceed in the directions chosen by for example, the hull of a torpedo. First, in order that the
the method. More than90 percent of the computing time is array look only at the target, it is desired that its three-
spent solving these single-variable problems, so the effec- dimensional beam pattern have the narrowest possible
tiveness of the technique used is critical. The authors have main lobe. Second, all points off this main lobe must lie
used twointerpolative techniques, one requiring derivatives, below a specdied threshold level, defined by the type of
the other not. The former, using cubic interpolation, is signalprocessingused,which willdifferforeachcross-
described elsewhere[41J91 and the latteris described below. sectional plane. Unfortunately for the designer, these two
Quadratic Interpolation :This technique solves the prob- primary requirements are at odds with one another. It is
lem of finding a= a* to minimize g(a)=f(e+as), where e a relatively simple matter to decrease the level of minor
is the current parameter vector and s a direction provided lobes. One simply radiates a greater portion of the total
by the unconstrained minimization algorithm. Initially, input power from elements located at the arraycenter. This
the variables e,, * . . ,e, are scaled so that a unit change in technique, called amplitude shading, effectively shortens the
any variable is a significant but not too large percentage array, moving the first moment (or dipole equivalent) in
change in that variable. For example, if a capacitor is ex- toward the center. However, this also produces a broader
pected to have a value around 500 pF then a 1-pF change beam, thereby violating the first requirement.
wouldbe considered significantwhereas a 10-pF change Until recently, work in this field has been concerned only
would be too large. The method proceeds in three stages. with the special case where the element spacings in both
The first stage normalizes the s-vector so that a step size planes are linear multiples of one another. This is primarily
a= 1 is acceptable ; the second stage establishes bounds because the simplitications which result in the equations
on a* ; and the third stage interpolates its value. describing the array serve to ease the calculations consider-
Stage I :Calculate ably. Indeed, until the recent advent of high-speed digital
computers, no other approach was practical. Some work
A = max I s j /
j has been done on arrayswith a variable interelement spac-
ing,[’51*[’61 but the results obtained are more applicable to
and divide each component of s by A. This insures that s is electromagnetic arrays, where grating lobes are significant,
a reasonable change in e. than to acoustic ones.
 1890 PROCEEDINGS OF THE IEEE, NOVEMBER 1967

The classic and widely referencedsolution to the general ARRAYCO-ORDINATESYSTEM

linear problem (i.e., designthe linear n-element array fitting
within a specified aperture which has simultaneously mini-
mum beamwidth and lowest minor-lobe level)wassug-
gested by Dolph.["] The authors have recently shown["]
that signdicant improvement over Dolph's results is possi-
bleby introducing the additional degrees of freedom of
variable interelement spacing and asymmetric excitation.
In the following, this work is extended to the planar array
case.
Acoustic beam patterns are diffraction phenomena, so
the equations defining the directional characteristics of
electromagnetic radiation are directly applicable. Consider
the ith element of a transducer array, whose location is
described by the two position parameters ri and 6, as
shown in Fig. 4. Define a plane at an arbitrary angle 4 XIX3-PLANE THROUGHTRANSDUCER
(VERTICALPLANE)
with respect to the horizontal. A plane acoustic wave inci- XI X2- HORIZONTAL PLANE
dent froman angle JI in the 4 plane will excite this element 0 P - IN PLANE (SHADED)

before reaching the phase center of the array, whichis Fig. 4. Planar array variables.
defined at the origin. Therefore, the phase of this element is
To construct the response functions, the directivity E ,
2xri
A,($, 4) = - -cos
1
(ei - 4) sin $. (12)
is firstnormalized by dividing by
N

If the arrayelements are isotropic andall excited withunit

amplitude, then the contributionto the total patternby the
ith element would be and expressing the result in decibels. Hence the response
function for the kth pattern is
Ci($, 4) = expj(ot - A i ( $ , 4))- (13)
However, practical array elements are generally noniso-
tropic since finite element size, mutual impedances, and
Fk(e, 4, $1 = 2o loglo
( 4, $1
/i':1
1 IvkiAi(o, 411)
reflections all cause their response to vary withtarget angle. The normalization causes F k ( e , 4, o) to be zero for the sum
In calculating the pattern of the entire array, these factors patterns. Ususally it is sufficient to sepcify Fk in the three
are accounted for by introducing the measured element planes 41= 0" (horizontal plane), 43= 90" (vertical plane),
directivities Ai($,4) as multiplicative factors in (13). In and 42=45" (skew plane). Thus, for each pattern in each
addition, if the ith module output is weighted by an ampli- plane there is a response function of the form
tude factor vi, this also enters (13) multiplicatively.
Given a s@c geometry, {ri, e,}, a number of different Fk,j(e, $1 Fk(e, 4 j , $).
types of directivity patterns can be generated by summing Each of these response functions is now considered only at a
the element contributions with different weightingor shad- set of discrete angles { $ p } where it must lie below speclfied
ing factors. The shading factor for the ith element's con- levels. Thus, for each response function, the set ZA is empty,
tribution to the kth pattern is denoted t.'ki. These patterns I , contains the indices of the selected angles $, and the
generally fall into two classes : sum and difference patterns inequalities (1H3) become
(see Fig. 5). The kth array pattern, &(e, 4, $), is
F k , j ( $e p, ) I bk,j($p) = ''. 9
N
j = 1,2,3
4, $1 = 1
i= 1
qiAi($, 4)
p = l;..,P
O I r i I R
ri cos (6, - 4) sin
0 < 13~I 271 i = l;.-,N
-V vk,i 5 V k = l;..,K.
where e corresponds to the variables { u k i , T i , e,}. If K pat-
terns are considered then e has N(K + 2) components. For This system of inequalities represents specifications for an
the problem actually considered, the modules were located N transducer array to generate K patterns. These patterns
(and driven) with quadrantal symmetry, thus reducing the are specified at P points in each of three planes. The overall
number of variables to N(K +2)/4. Specifically, for sum system has a maximum radius of R and each element shad-
patterns all four quadrants areidentical, while for difference ing is bounded by V .
patterns the top and bottom (left and right) halves have Introducing the variable and the non-negative weights
shadings differing only by a minus sign. { w , , ~ , ~ } the
, optimal array design problem is
WAREN ET A L . ; OPTIMIZATION IN DESIGN 1891

as expected, they increased in regions where no pointswere

assigned. When this became unacceptable, the minimiza-
subject to tions were stopped, the (0,) reassigned based on the present
minor-lobe structure, the ( B j } modified, and the minimiza-
tions begun on this new problem. For this particular exam-
ple, this sequence was repeated twice.
Table I shows the values of the initial and final parameter
matrices. Note thatthe starting arraywas symmetricwithin
a quadrant, while the optimal array was not. This follows
from the asymmetry of the specifications.

TABLE I

Shading
Shading
Hori- Shading
r e X Y zontal
Sum Vertical
Diff. Diff.
K 3 P
1 Parameters of Initial Array i.= 4 inches
k= 1 j = 1 p = 1 Fk,j,p(e) + @k,j,pi - bk,j,p 4.37 22" 4.052 1.637 0.980 1.660 0.950

;[' 4.21 67" 1.645 3.875 0.940 0.700 1.450

+ i=l ri
+-R -1 ri +-+-
1
Qi
1
271 - Oi
1.50 45; 1.068
Parameters of Final Array
1.061 1.570 1.140 1.150

4.194 21.30" 3.907 1.523 1.287 3.052 0.782

3.710 66.67' 1.470 3.406 1.228 0.843 1.690
1.484 32.51" 1.251 0.798 1.528 1.037 1.266

Problem: Design a 15-kHz,12-element planararray

capable of fitting in an 8Sinch diameter circle, composed Figs. 5 and 6 show the calculated patterns in the hori-
of isotropic radiators, having quadrantal symmetry, which zontal plane, from the original and optimal parameter
will produce simultaneously three patterns. The amplitude matrices. Although there is improvement in both patterns,
response of each is specifiedas to beamwidth and maximum it is not startling. Figs. 7 and 8 present similar information
minor-lobe level in three planes, horizontal, vertical, and in the vertical plane. The improvement in the difference pat-
a plane 45" intermediate to the two. Furthermore, let two tern is very great. Note that in this plane it was possible to
of these patterns be differencepatterns, onewith a vertically get both patterns to meet the specifications.Note also that
oriented notch, and onewith a horizontally oriented notch. the sum pattern was deteriorated, not improved, in the
In addition, make the beamwidth of the vertical-difference optimum case.The pattern,however, still exceedsthe speci-
patternbroaderthanthat of the horizontal-difference fications established initially. Figs. 9 and 10 present similar
pattern by a specified amount. data in the intermediate plane. In this plane, itproved
This problemwasformulated as described earlier. In possible to complywith the minor-lobe specilication on
orderto adequately characterize each response function two of the three patterns and miss it by only 2 dB in the
over the 90" range of interest for II/,approximately twenty third. The starting point, incidentally, represents the best
to twenty-five discrete angles would be needed. Since the available computer-aided design effort: 3 hours of work,
computing time is proportional to the number of angles at using the machineonly as apattern dalculation device
which the response must be calculated, it is desirable to use (about 12 minutes of computer time).
as few points as possible. By using an iterative design
method, it was possible to obtain a solution using less than B. F i l t e r D e s i g n
ten points for each response function. The problem here is to design a quartz resonator band-
An initial design was obtained and analyzed. Then, for pass filter to meet or exceed the following specifications:
each response function, the associated II/,
were distributed
center frequency 68.0 MHz
in those ranges of II/ where the minor-lobe levels were least 6 dB bandwidth 100.0+2.0 kHz
satisfactory and specifications/?k,](II/,) assigned. In addition. 60 dB bandwidth < 180.0 kHz
one of the II/,
was assigned to determine the 10-dB major- maximum
passband
ripple
attenuation
ultimate
1.O dB
> 80.0 dB
lobe beamwidth. If the beamwidth was already satisfactory
(i.e., narrower than required) the associated weight was set In addition, the resulting element values must be easy
to zero, thus maintaining this situation. Otherwise the cor- to realize. This requirement includes the incorporation of
responding weight was setto unity, maintained at this value realistic losses and bounds on all element values.
until the desired beamwidth was attained, and then set to For quartz resonator filters, 100kHz is a ratherwideband
zero. requirement. Because of this, and thehigh ultimate attenua-
As the P-function minimizations progressed, the minor- tion, a cascade half-lattice structure was selected. Initially,
lobe levels at the assigned points generally decreased, while, it was felt that two four-resonator sections could meet the
1892 PROCEEDINGS OF THE IEEE,NOVEMBER 1967
WAREN ET AL.: OPTIMIZATION INDESIGN 1893

- F; fe.90.6.I STARTING POINT - I

0 15 30 45
MECHANICAL ANGLE (DEGREES)
60
-9
Fig. 7. Initial response function in vertical plane.
- 75 90

0
FINAL RESULT -
VERTICAL PLANE

DIFFERENCE

-20 - VALUE SPEC

-25 -
-30 -
-35 .

MECHANICAL ANGLE (DEGREES) - 9d

Fig. 8. Final response function in vertical plane.
1894 PROCEEDINGS OF THE IEEE,NOVEMBER 1967

MECHANICAL ANGLE (DEGREES) - $J d

Fig. 9. Initial response function in skew plane.

MECHANICAL ANGLE (DEGREE@ -$+

Fig. 10. Final response function in skew plane.
WAREN ET AL.: OPTIMIZATION IN DESIGN 1895

c6 dB
o dB
0 I I M / 4- REFEREWE OL I I I I I
-200 -100 0 100 200 -200 -100 0 100 200
DEVIATIONFROM CENTER FREQUENCY (kHz) DEVIATIONFROMCENTERFREQUENCY (kHz)
Fig. 1 1 . Response of lossless single-sectionfilter. Fig. 13. Response of filter after one optimization cycle.

BO
I I I I I -- I I I I I
CASCADE OF 2 LOSSY
6" ORDER SECTIONS
70 - 70 ~

60 - 60-

-
m
50- c
m
50-
E U
v

40-
\ In
4z 40-
0 0
I-
0: K
W W
VI VI
z 30- 30-

20 - 20 -

M I N 6 dB
dB

w-
e6
10- 4 10-

c 0 dB REFEREM o dB REFERENC

0 I I I I I 0, I 1 I 1 I
-200 -100 0 too 200 -200 -100 0 100 200
DEVIATION FROM CENTER FREQUENCY (kHz) DEVIATION FROM CENTERFREQUENCY (kHz1
Fig. 12. Response of cascade of two lossy filter sections. Fig. 14. Finalfilterresponse.
1 8% PROCEEDINGS OF THE IEEE,NOVEMBER 1967

(a) FILTER CONFIGURATION

Cs = 1.0PF
Q=60K
0

(b) SECTIONS I AND 3 (C) SECTION 2

Fig. 15. F i i filter design.

specifications. Since an adequate synthesis procedure for V. CONCLUSIONS

cascade lattices does not exist, a singlesection was designed The inequality formulation presented here seems to be
using the O’Meara-Sch~effler[’~I*[~~] synthesistechnique. applicable to a wide class of designproblems. However, only
Fig. 11 shows the ideal lossless insertion loss for the sixth- recently have efficientsolution techniques become available
order l d B Chebyshev filter selected. Introduction of real- for solving the resulting optimization problems. It is ap-
istic losses (Qof 50 for inductors and 60 OOO for resonators) parent that the optimization process fits very naturally into
distorts the filter characteristic, but impedance scaling of the usual computer-aided design procedure and leads to
the terminating tuned circuits helps compensate for this significantly improved designs with considerably less effort
effect. on the part of the designer. It is hoped that this paper will
Two of these sections were cascaded and, as shown in arouse interest in the use of these techniques and motivate
Fig. 12,this design fallsfar short of the specifications. It was others to develop new mathematical programming formu-
felt that this structure could achieve a much better response. lations for design problems.
Therefore, itwas usedas the initial point for an optimization
REFERENCES
stage, as described previously (see also Lasdon and “Computer-aided design,” pts. 1-10, Electronics, September 1966-
WarenL4]),and the resulting design comes much closer to July 1967.
the specifications (Fig. 13). IZ1 F. F. Kuo, “Network analysis by digital computer,” Proc. IEEE,
At this point, some motional capacitance valueswere vol. 54, pp. 820-829, June 1966.
H.C. So, “OLCA: An on-line circuit analysis system,” this issue,
modified, for easeof realizability, with negligiblechange in p. 1954.
insertion loss. L. S. Lasdon and A. D. Waren,“Optimaldesign of filterswith
Since the specifications are still not met, the structure bounded, lossy elements,” IEEE Trans. Circuit Theory, vol. CT-13, pp.
175-187, June 1966.
wasmodified by introducing a third half-lattice section. J. A. Athanassopoulos, J. D. Schoeffler, and A. D. Waren, “Time-
This wasa fourth-order, half-dBChebyshevdesign, in- domain synthesis by nonlinear programming,” Proc. 4thAnn.Allerton
serted between the two sixth-order sections. The insertion Con on Circuit andsystem Theory, pp. 7 6 7 7 5 , 1966.
‘klG. Zoutendijk, “Nonlinear programming-Anumerical survey,”
loss for this filter comes quite close to the requirements. SIAM J . on Control, vol. 4, no. 1, pp. 194-210, 1966.
One more short optimization cycle was used to obtainthe [‘I A. V.Fiacco and G. P. McCormick, “The sequential unconstrained
final design (Figs. 14 and 15) which exceeds all the s m - mhimization technique for nonlinear programming, a primaldual
rnetld,” Mumgement Science, vol. 10, no. 2, pp. 3&366,1964.
cations. [*I A. V. Fiacco and G. P. McCormick, “Computational algorithm
Note thatin this example an initial structure was selected for the sequential unconstrained minimization technique for nonlinear
and subsequently modified, a computer analysis program progamming,” Munagement Science, vol. 10, no. 4, pp. 601417, 1964.
D. Goldfarband L. Lapidus, “A conjugate gradient methodfor non-
was used frequently forthe evaluation of proposed designs, linear programming,” Dept. of Chem.Engrg.,PrincetonUniversity,
and theoptimization cycle was used twice. Princeton, N.J, November 1967.
PROCEEDINGS OF THE IEEE. VOL. 5 5 , NO. 11, NOVEMBER
1897 1967

R. Fletcher and M. J. D. Powell, “A rapidly convergent descent [I6] Y. T. Lo and S. W. Lee, “A study of space-tapered arrays,” IEEE
method for minimization,” Computer J., vol. 6, pp. 16S168, June 1963. Trans. Antennas and Propagation, vol. AP-14, pp. 22-30, January 1966.
[ 1 1 ] W. C . Davidon, “Variable metric method for minimization,” l1 ’I C. L. Dolph, “A current distribution for broadside arrays which
AEC Research Development Rept. ANL-5990 (Rev.), 1959. optimizes the relationbetween beam width and side lobe level,” Proc. IRE,
M. J. D. Powell, “An efficient method for finding the minimum of vol. 34, pp. 33S348, June 1946.
a function of several variables without calculating derivatives,” Computer [ 1 8 ] L. S. Lasdon, D. F. Suchman, and A. D. Waren, “Nonlinear pro-
J., V O ~ .7, pp. 155-162, 1964. gramming applied to linear array design,” J. Acoust. SOC. Am., pp. 1197-
[131G. W. Stewart, “A modification of Davidon’s minimization 1200, November 1966.
method to accept difference approximations of derivatives,” J. ACM, [”I T.R.OMeara, “The cascade synthesis of crystal filters with trans-
vol. 14, pp. 72-83, January 1967. mission zeros at infinity,” IEEE Trans. Circuit Theory, vol. CT-10, pp.
[14] R. Fletcher, “Function minimization without evaluating deriva- 533-534, December 1963.
tives-A review,” Computer J.,vol. 8, pp. 3341, 1965. J.D. Schoeffler, “Insertion loss design of symmetrical lattice
[ I s 1 M. G. Andreasen, “Linear arrays with variable interelement spac- piezoelectric resonator filters with symmetrical or unsymmetrical pass-
ings,” IRE Trans. Antennas and Propagation, vol. AP-IO, pp. 137-143, bands,” I R E Trans. Circuit Theory, vol. CT-9, pp. 251-259, September
March 1962. 1962.

Computer-Aided Transistor Design,

Characterization, and Optimization
HITENDRA NATH GHOSH, MEMBER, IEEE,FRANCISCO H. DE LA MONEDA, MEMBER,
IEEE,
AND NICHOLAS R. D O N 0

Abstract-A systematic and general method of computing ac and dc total area of the emitter junction
characteristics of doublediiiused juacth tramistors oiap major process area of a given sectional emitter diode
parameters d as dimemiom of the device, d a c e concentralion, jmction collector effective area
depth, difbion time, temperature, and difiasioo coel%cii as a hmtion of
temperature and impurity concentration is dgcribed. ’zbese parameters can impurity gradient at the collector junction
be checked dming the process and, therefore, can aid the process control impurity gradient at the emitter junction
proMembypredictingtheexpectedvalwsofjrmctioedepthnndsbeetresis- collector-base breakdown voltage with the emitter
tivity.If the specified cwtrol parameters are met dnringthe fabrication,the open
ac and dc characteristics of the device w l
li be realized. ’Tbe time and/or
emitter-base breakdown voltage with the collector
frequency respome of a circuit can be computed d u g the ac and dc char-
acteristicsofthedifhseddevicesofthecircaitoa~b4fisofadistribotedor open
an equivalent lumped model. The measurement of imporhut ac parameters impurity distribution
oathebPsisoftheSemodetshnsnlsobeensimnlatedoathecomputer,thm spatial derivative of C(x)
aidirtg the cbaracterizatioa problem of the device in the integrated circuit background doping of the collector
environment. A b , the switching speed of a loaded logic net can be computed total collector capacitance
and optimized by trading & interacting parametersand relaling them back
to theoriginal difhsiooprocessparameters and dimensions of tee compweotr base region capacitance
Finally, experimental verification of the computed resalts has been acMw- emitter region capacitance
@bedand f d to be satisfactory. emitter junction capacitance
initial surface concentrationprior to drive-in
( j = 1, 2)’
NOTATION final surface concentration [see (6)] ( j = 1, 2)’
A, collector area + +
C , C j , Cd [see (22a)I
A, bottomarea of the emitter junction for the whole impurity diffusion constant ( j = 1,2)’
device electron diffusion constant
average electron diffusion constant
Manuscript received September 18, 1967. This paper is reprinted from hole diffusion constant
Solid State Electronics, vol. 10, pp. 70s-726, 1967. see Fig. 3(a)
H. N. Ghosh is with the Components Division, IBM Corporation,
Hopewell Junction, N. Y.12533. gain-bandwidth product
F. H.De La Moneda is with the Components Division, IBM Corpora- emitter half-width
tion, Hopewell Junction, N. Y. He is currently on an educational leave inverse voltage gain
of absence at the University of Florida, Gainesville, Fla.
N. R.Dono is with the Systems Development Division, IBM Corpora-
tion, Poughkeepsie, N. Y. j takes on values of 1 and 2 to denoteemitter and base diffusion cycles.