Control Proxy Functions for

Sequential Design and Control

Diane L. Peters1
e-mail: [email protected] Optimization
P. Y. Papalambros Optimal system design of “smart” products requires optimization of both the artifact and
e-mail: [email protected] its controller. When the artifact and the controller designs are independent, the system
solution is straightforward through sequential optimization. When the designs are
A. G. Ulsoy coupled, combined simultaneous optimization can produce system-optimal results, but
e-mail: [email protected] presents significant computational and organizational complexity. This paper presents a
method that produces results comparable with those found with a simultaneous solution
University of Michigan, strategy, but with the simplicity of the sequential strategy. The artifact objective function
2250 G. G. Brown Building, is augmented by a control proxy function (CPF), representing the artifact’s ease of con-
2350 Hayward Street, trol. The key to successful use of this method is the selection of an appropriate CPF.
Ann Arbor, MI 48109 Four theorems that govern the choice and evaluation of a CPF are given. Each theorem
is illustrated using a simple mathematical example. Specific CPFs are then presented for
particular problem formulations, and the method is applied to the optimal design and
control of a micro-electrical mechanical system actuator. [DOI: 10.1115/1.4004792]

Keywords: co-design, optimization, coupling, control

1 Introduction tures subject to active control (e.g., Ref. [13]); other such systems
exists, including an elevator with gain-scheduled control [14].
Optimal design of “smart” products and systems requires opti-
Coupling in these problems is described as unidirectional. In the
mization of both the physical system or artifact, and its controller.
present work, only unidirectional coupling is considered.
This codesign problem can present challenges when the design of
A variety of measures have been proposed to quantify the
artifact and controller is dependent on each other. In general, the
strength of coupling [1,15–17]. These measures have been shown
artifact objective function, fa, and inequality and equality con-
to be related, though most are not commensurate with one another
straints, ga and ha, may all be functions of both artifact and con-
[18,19]. In problems with unidirectional coupling, a useful mea-
trol variables, da and dc, respectively. Likewise, the control
sure is the coupling vector, Cv, which will be used in this work,
objective function, fc, and inequality and equality constraints, gc
and is defined as Refs. [16, 20, 21]:
and hc, may all be functions of both da and dc; i.e., fa ¼ fa(da, dc),
ga ¼ ga(da, dc), ha ¼ ha(da, dc), fc ¼ fc(da, dc), gc ¼ gc(da, dc), and  
wc @fc @fc @d c
hc ¼ hc(da, dc). When this interdependence exists, the solution of Cv ¼ þ (6)
the bi-objective optimization problem given by Eqs.(1)–(5) is a Par- wa @d a @d c @d a
eto set, with the various Pareto points found by varying the weights
wa and wc, where wa, wc > 0, and the problem is said to be coupled. This vector is valid only at an optimal solution; however, if a
point is not known to be an optimal solution, an estimate can be
min wa fa þ wc fc (1) computed. The equation for the estimated coupling vector,
d a ;d c
denoted as C b v , is identical to Eq. (6), but does not require the solu-
subject to ga  0 (2) tion of Eqs. (1)–(5). Of course, if the point does happen to be on
the Pareto frontier, then C b v ¼ Cv .
ha ¼ 0 (3) It has been shown that simple sequential optimization, in which
gc  0 (4) the artifact is first optimized and then the optimal control is found
for that artifact, does not necessarily find system-optimal solutions
hc ¼ 0 (5) [21,22]. Combined optimization methods such as a simultaneous
strategy, in which both the artifact and the control are optimized
Many such coupled systems have been reported in the literature. together, will produce system-optimal solutions [21,22]. These
These include structural systems with active control (e.g., Refs. methods require combining expertise from more than one disci-
[1–3]), micro-electrical mechanical systems (MEMS, e.g., Refs. pline to formulate and solve the full optimization problem. This
[4, 5]), and robotics and mechatronics (e.g., Refs. [6–8]). The presents organizational challenges, since such expertise is typi-
experimental and analytical studies have shown that coupling cally found in different individuals or different groups within an
must be considered and addressed and that doing so can present organization. Furthermore, specialized techniques developed for
significant challenges (e.g., Refs. [9–12]). optimal control can no longer be used when the problem is not
When all of the objective and constraint functions depend on formulated as a purely optimal control one. The combined solu-
both da and dc, coupling is described as bi-directional. However, tion presents also a computational challenge.
there are many engineering problems in which the artifact objec- In this paper, the new concept of a control proxy function
tive and constraints are not functions of dc, i.e., fa ¼ fa(da), (CPF) is introduced. A combined problem including the CPF and
ga ¼ ga(da), and ha ¼ ha(da). Such problems include many struc- the artifact objective function is formulated, followed by the opti-
mization of the control problem. This allows an effective sequen-
1
Corresponding author. tial solution strategy to be implemented for the combined
Contributed by the Design Automation Committee of ASME for publication in problem. The CPF concept is introduced in Sec. 2. In Sec. 3, four
the JOURNAL OF MECHANICAL DESIGN. Manuscript received May 25, 2010; final manu-
script received June 17, 2011; published online September 15, 2011. Assoc. Editor: theorems are presented governing the choice and the evaluation of
Timothy W. Simpson. appropriate CPFs, and simple mathematical examples are used to

Journal of Mechanical Design Copyright V

C 2011 by ASME SEPTEMBER 2011, Vol. 133 / 091007-1

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 demonstrate the concepts. In Sec. 4, the use of a CPF is shown to neous formulation; a poorly chosen CPF will yield solutions far
provide optimal or near-optimal solutions to the codesign problem from system optimality. The question then arises of how one
without the disadvantages seen in the combined optimization tech- should choose an appropriate CPF, and how one may evaluate a
niques. A set of CPFs are derived for the specific problem formu- proposed CPF to determine whether it results in optimal or near-
lations. In Sec. 5, the method is demonstrated on a MEMS optimal solutions. Of course, the closeness of the CPF solution
actuator, and Sec. 6 presents concluding remarks. should be determined without solving the simultaneous formula-
tion in Eqs. (1)–(5) for Pareto optimal points, since the motivation
for the CPF formulation is to eliminate the need to solve the
2 Control Proxy Function Problem Formulation simultaneous problem.
In order to preserve the functional decomposition of the code-
sign problem while realizing optimal or near-optimal solutions, a 3 Characteristics of Effective Control Proxy
modified sequential optimization strategy is proposed. In this Functions
strategy, the original artifact objective function, fa, is augmented Four theorems are presented that govern the characteristics of
with a CPF, representing the system’s ease of control. The CPF, an appropriate CPF. First, a “perfect” CPF is defined, and a condi-
denoted as v, is a function of only the artifact design variables da; tion which ensures that a CPF is perfect is given as Theorem 1.
it is independent of the control design variables dc and of the con- Next, for a CPF that is not perfect, a measure of its “closeness” to
trol architecture itself. The bi-objective artifact design problem the Pareto frontier is defined and characterized in Theorem 2.
with the two functions fa(da) and v(da) may be solved in a variety Finally, Theorems 3 and 4 relate an effective CPF to the mathe-
of ways, as discussed in the multi-objective optimization literature matical form of the control objective function, fc. For Theorems 2
(e.g., Refs. [23–26]). In this paper, a simple weighted linear com- through 4, it is assumed that all functions are locally convex.
bination will be used to demonstrate the proposed method. The
optimization problem is then formulated as follows, see Fig. 1: 3.1 Characterization of a Perfect CPF. A CPF is described
as perfect if every solution of the CPF problem is also a solution
min w1 fa ðd a Þ þ w2 vðd a Þ (7) to the simultaneous problem given in Eqs. (1)–(5), i.e., every CPF
da
point will coincide with the Pareto frontier. Such points satisfy the
subject to ga ðd a Þ  0 (8) condition given in Theorem 1.
Theorem 1. If Cv k rv for all solutions to the CPF problem
ha ðd a Þ ¼ 0 (9) given in Eqs. (7)–(9), then all solutions to the CPF problem will
also be solutions to the simultaneous problem given in Eq. (1)–(5).
where the positive weights w1 and w2 represent the relative impor- Proof. For the simultaneous problem stated in Eqs. (1)–(5), the
tance of the artifact objective and the CPF. This is followed by the Karush-Kuhn-Tucker (KKT) conditions [27] can be stated as
control design problem:
2  3 2 @h 3
  @fa wc @fc @fc @d c a
min fc d a ; d c (10) þ þ
6 @d a wa @d a @d c @d a 7
dc 6 7 þ kT 6 @d
4 a5
7
  4 wc @fc 5 @hc
subject to gc d a ; d c  0 (11)
w @d @d c
  2 3a c
hc d a ; d c ¼ 0 (12) @ga
6 @d 7
þ lT 4 a 5 ¼ 0 (13)
where d a ¼ argminðw1 fa ðd a Þþ w2 vðd a ÞÞ subject to ga(da)  0, @gc
ha(da) ¼ 0 is the solution to Eqs. (7)–(9). @d c
The success of the method, in terms of reproducing the results  
g
of the simultaneous formulation of Eqs. (1)–(5), depends on the lT a ¼ 0; k 6¼ 0; l  0 (14)
gc
selection of an appropriate CPF. A well-chosen CPF, which effec-
tively captures the fundamental physical limitations that deter-
mine the attainable control performance of the system, will result and, for the CPF problem stated in Eqs. (7)–(12), the KKT condi-
in solutions close to the Pareto-optimal points found by a simulta- tions are
2 3 2 3 2 3
@fa w2 @v @ha @ga
6 @d þ 7 6 7 6 7
6 a w1 @d a 7 þ kT 6 @d a 7 þ lT 6 @d a 7 ¼ 0 (15)
4 w2 @fc 5 4 @hc 5 4 @gc 5
w1 @d c @d c @d c
 
g
lT a ¼ 0; k 6¼ 0; l0 (16)
gc

Assume that, for every set of weights wa and Wc, there exists
some set of weights w1 and w2 such that the two formulations will
have identical solutions. Then it can be shown that

@fa @fa w2 @v
þ Cv ¼ þ (17)
@d a @d a w1 @d a

and, consequently, that

w2
Cv ¼ rv (18)
w1

Such a set of weights will exist and the modified sequential prob-
Fig. 1 Control proxy function problem formulation lem will produce the Pareto-optimal solutions when the gradient

091007-2 / Vol. 133, SEPTEMBER 2011 Transactions of the ASME

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
  
of the CPF, rv, is parallel to the coupling vector Cv. Thus, the e ¼ d a  d a 2 (28)
theorem is proven. h
Example. To illustrate the relationship between Cv and rv, con- where da is the vector of design variables and d a denotes the vec-
sider the following coupled optimization problem: tor of design variables at an optimal solution to the codesign
problem.
min wa fa ðd a Þ þ wc fc ðd a ; d c Þ (19) Since it has been specified that coupling is unidirectional, it is
subject to ga ðd a Þ ¼ 4da21 þ da22  900  0 (20) possible to express the optimal control design variables dc as a
function of the artifact design variables as follows:
where  
dc ¼ d c d a (29)
fa ¼ 0:5da21 þ da22  da1 da2  7da1  7da2 (21)
which allows the control objective fc to be transformed into a
1 function of only da. This shall be used later in the proof in order
fc ¼ ðda1  da2 Þ þ ðda1 þ da2 þ d c  10Þ2 þ ðd c  5Þ2
2
(22)
9 to find gradients of fc in the da-space. It is possible to incorporate
any active controller constraints gc(da, dc), hc(da, dc) into Eq. (29).
The CPF This codesign problem is also formulated as the CPF problem

vðd a Þ ¼ 11da21 þ 11da22  18da1 da2  10da1 min w1 fa ðd a Þ þ w2 vðd a Þ (30)

da
 10da2 þ 25 (23)
followed by
is chosen, and the system is optimized both sequentially and
simultaneously. min fc ðd c Þ (31)
dc
The coupling vector Cv is given by the relation
subject to gc ðd c Þ  0 (32)
 
2wc 10da1  8da2 þ dc  10 T hc ðd c Þ ¼ 0 (33)
Cv ¼    (24)
9wa 8da1 þ 10da2 þ dc  10
The functions v(da), fa(da) and the reduced-space function fc(da)
where da1 , da2 , and dc are values of these variables evaluated at an are all assumed to be convex functions. Consider a point A which
optimal solution. The CPF has the gradient solves the CPF problem. At this point, with d a ¼ d Aa
 T
22da1  18da2  10 w1 rfaA ¼ w2 rvA (34)
rv ¼ (25)
18da1 þ 22da2  10
The point B is chosen such that it is the nearest Pareto-optimal so-
Since there are no controller constraints gc(da, dc) or hc(da, dc), lution to point A. Its location in the artifact design variable space,
@fc =@d c ¼ 0 at an optimal solution, and therefore therefore, satisfies the relation d Aa ¼ d Ba þ ev, where v is a unit
vector normal to rfaB and e is the distance between A and B, as
dc ¼ 5:5  0:1da1  0:1da2 (26) shown in Fig. 2.
Using a Taylor series expansion of about B, the artifact and the
controller objective functions at A can be expressed as
Substituting into Eq. (24)
  1
2wc 1:1da1  0:9da2  0:5 T faA ¼ faB þ rfaB ev þ e2 vT r2 faB v þ higher order terms (35)
Cv ¼   (27) 2
wa 0:9da1 þ 1:1da2  0:5 1
fcA ¼ fcB þ rfcB ev þ e2 vT r2 fcB v þ higher order terms (36)
2
Comparing Eqs. (25) and (27) at an optimal solution to the simul-
taneous problem, it can be seen that Cv ¼ 10wwac rv; therefore, the
use of this CPF will duplicate the Pareto frontier.

3.2 Quantification of the “Closeness” of a CPF Point to

the Pareto Frontier. As stated above, a perfect CPF is character-
ized by rv k Cv , i.e., when the angle n between the vectors rv
and Cv is zero. This suggests that the angle n may serve as a
means of evaluating the fidelity of a given CPF in modeling the
behavior of fc. The theorem specifies that no artifact constraints
are active. However, since constraints reduce the degrees of free-
dom of a system, it is conjectured that if the problem is convex
and the point in question is near the Pareto frontier, the angle n
will represent an upper bound on the “distance” between a CPF
point and the Pareto frontier when constraints are active.
Theorem 2. If a codesign problem, as given in Eqs. (1)–(5), is con-
vex and no artifact constraints ga(da), ha(da) are active, then the angle
n between rv and the estimated coupling vector C b v at a CPF point
will be monotonically related to , the distance between that CPF
point and the nearest Pareto optimal point, measured in the da-space.
Proof. The distance from optimality has been defined in several
ways [28]; here, it is defined as the distance between a given
point, da, and the nearest Pareto-optimal point, d a , in the da-space Fig. 2 Gradients at Points A and B

Journal of Mechanical Design SEPTEMBER 2011, Vol. 133 / 091007-3

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 and, therefore, neglecting higher-order terms, the gradients of fa
and fc at A can be expressed in terms of the gradients at B as

rfaA ¼ rfaB þ evT r2 faB (37)

rfcA ¼ rfcB þ ev T
r2 fcB (38)

The estimate of the coupling vector, computed from Eq. (6), is

given by

^ A ¼ w2 rf A
C (39)
v
w1 c

and the angle n can be found from the relation

CA  rvA rf A  rvA
cos n ¼  Av  ¼ c (40)
C krvA k rf A krvA k
v c
Fig. 3 Comparison of angle n and distance e
From optimality conditions, it is known that

wa rfaB ¼ wc rfcB (41) always increasing or always decreasing with respect to a given
variable, e.g., if the partial derivative of a continuous function
w1 rfaA ¼ w2 rvA (42) does not change sign [28]. Monotonicity analysis is a useful tool
in optimization problems. For example, it can be used to deter-
Using Eq. (42), it is possible to rewrite Eq. (40) as mine constraint activity, to study the behavior of composite func-
tions, and to give insight into the tradeoffs present in optimization
rfcA  rfaA problems [28]. Here, monotonicity is used to characterize effec-
cos n ¼   
rf A rf A  (43) tive CPFs.
c a
If a controller objective fc(da, dc) is monotonic with respect to
an element of da, then it seems appropriate that v(da) should have
By following the mathematical procedure detailed in Ref. [19], it the same monotonicity with respect to that element of da in order
can be shown that to effectively model the behavior of fc.
  Theorem 3. If fc(da, dc) is monotonic with respect to some ele-
T 1 wc BT  T 2 B 
cos2 n r2 faB v  rfc v r fc v ment of da, and that element of da does not appear in any active
e wa constraint, then a CPF with the same monotonicity will produce
 
T 1 T  T 2 B  solutions closer to the optimum than a CPF with the opposite
¼ r2 faB v  rfcB v r fc v (44) monotonicity.
e
Proof. Assume that, in a codesign problem, the controller
Since the functions have been specified as convex, vT r2 fcB v  0 objective function fc(da, dc) is monotonic with respect to the jth
and vT r2 fc B v  0. As the value of  increases, it is evident that component of the n-dimensional vector of artifact design variables
cos2n must decrease, and, therefore, n is increasing. Given that n da. Two CPFs will be used to solve this problem, denoted as
has the same monotonicity as e, the angle n is an appropriate mea- v1(da) and v2(da). The two CPFs are selected such that they have
sure of the distance between a CPF solution and the unknown Par- opposite monotonicity with respect to the jth component of da,
eto frontier, and the theorem is proven. h i.e., one CPF is increasing with respect to daj , while the other is
Example. Consider the following problem: decreasing with respect to daj [28]. Mathematically, this can be
stated as follows:
min wa fa ðd a Þ þ wc fc ðd a ; d c Þ (45)
@v1 @v
¼ 2 8 fi : i 6¼ j; 1  i  ng (49)
where @dai @dai
@v1 @v
fa ¼ 0:5da21 þ da22  da1 da2  7da1  7da2 (46) ¼ 2 (50)
@daj @daj
1 ! !
fc ¼ ðda1  da2 Þ2 þ ðda1 þ da2 þ d c  10Þ2 þ ðd c  5Þ2 (47) @v1 @fc
9 sgn ¼ sgn (51)
@daj @daj
A CPF of
Let the point A in the da-space be a solution to the CPF problem
v1 ðd a Þ ¼ ðda1  5Þ2 þ da22  25 (48) using v1(da), and let point B in the da-space be the Pareto optimal
solution to the codesign problem that is nearest to point A.
is chosen, and the system is optimized both sequentially and Assume that, at point B, daj does not appear in any active control-
simultaneously. The angle n is compared with the distance  to the ler constraints gc(da, dc) or hc(da, dc). Choose point C such that it
nearest point on the true Pareto frontier in Fig. 3, and it can be is a solution to the CPF problem using v2(da) and such that point
seen that, for this example, this angle is an effective measure of B is the Pareto-optimal point nearest to it, as shown in Fig. 4.
the distance to the frontier. Using this measure, the accuracy of a If the distances from point B to points A and C, denoted as eA
CPF can be evaluated without knowing the true Pareto frontier. and eC, respectively, are sufficiently small, then the functions fc,
v1, and v2 can each be represented by first-order Taylor series
approximations. Expanding about B,
3.3 Monotonicity of Controller Objective and CPF. A
function is said to be coordinate-wise monotonic if it is either fcA ¼ fcB þ eA rfcB v1 (52)

091007-4 / Vol. 133, SEPTEMBER 2011 Transactions of the ASME

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 to the conclusion that nA < nC. Therefore, from Theorem 2, it can
be stated that v1(da) will produce solutions closer to the Pareto opti-
mal points than v2(da) will. Since v1(da) shares the same monoto-
nicity as fc(da, dc), the theorem is proven. h
Example. Consider the following problem:

min wa fa ðd a Þ þ wc fc ðd a ; d c Þ (66)
subject to ga ðd a Þ ¼ 10  da1  da2  0 (67)
 
dc 2
gc ðd a ; d c Þ ¼ da2   25  0 (68)
2

where

fa ¼ 0:5da21 þ da22  da1 da2  7da1  7da2 (69)

fc ¼ da1 þ da2  d c (70)
Fig. 4 Pareto-optimal Point B and CPF Points A and C
and da ¼ {da : da  0}.
A CPF of

fcC ¼ fcB þ eC rfcB v2 (53) v1 ðd a Þ ¼ da1 þ 0:25da0:5

2
(71)
vA1 ¼ vB1 þe A
rvB1 v1 (54)
is chosen. It is evident that fc and v are both monotonically
vC2 ¼ vB2 þe C
rvB2 v1 (55) increasing with respect to da1 and da2 . Solving both the simultane-
ous optimization and the CPF problem, it can be seen in Fig. 5
where v1 and v2 are unit vectors such that v2 ¼ 6v1. that the CPF solution models the tradeoff between fa and fc,
By taking gradients of Eqs. (52) and (53), it can be stated that though imperfectly.
In contrast, consider a CPF of
rfcA ¼ rfcB (56)
v2 ðd a Þ ¼ da1 þ 0:25da0:5 (72)
rfcC ¼ rfcB (57) 2

In this case, the monotonicities of v and fc with respect to da1 do

The coupling vector estimates at A and C are given by not match. Solving the new CPF problem, it can be seen in Fig. 5
A
that this CPF does not model the tradeoff between fa and fc.
b A ¼ w2 rf A
C (58)
v
wA1 c 3.4 Locations of Unconstrained Minima of fc and v. In
wC2 Sec. 3.3, fc was assumed to be monotonic with respect to some
bC ¼
C rfcA (59) element of da. Here, the case where fc is not monotonic, but rather
v
wC1 has an unconstrained minimum, is considered. Intuitively, it
seems to be likely that the values taken by da at the minimum of
where wA1 and wA2 are the weights which produce point A, and wC1 fc should also minimize v, and that a CPF will become less effec-
and wC2 are the weights which produce point C. Substituting Eqs. tive if the minimum of v is farther from the minimum of fc in the
(56) and (57) into Eqs. (58) and (59) da-space. This can be proven for an unconstrained problem, and
A
it is conjectured that it will also be true for constrained problems.
b A ¼ w2 rf B
C (60)
v
wA1 c
C
b C ¼ w2 rf B
C (61)
v
wC1 c

The gradients rvA1 and rvC2 are given by the relations

rvA1 ¼ rvB1 (62)

rvC2 ¼ rvB2 (63)

Using Eqs. (58)–(63), n can be found for A and C

b A  rvA
C
cos nA ¼  v
 b A 
1
 (64)
Cv rvA1 

b C  rvA
C
cos nC ¼  v
 b C 
2
 (65)
Cv rvC2 

Substituting into Eqs. (64) and (65), it is possible to relate nA and Fig. 5 Comparison of simultaneous and CPF solutions for
nC, as shown in Ref. [19], and to show that cos nA > cos nC, leading appropriate and inappropriate monotonicity

Journal of Mechanical Design SEPTEMBER 2011, Vol. 133 / 091007-5

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Theorem 4. Assume that fc(da, dc) has an unconstrained mini- The unit vector s can be expressed in terms of the unit vector n by
mum, and that v(da) is chosen such that it has an unconstrained means of a rotation matrix, R, as s ¼ Rn; substituting this relation
minimum. Then, the distance between a CPF point and the Pareto and Eqs. (75) and (76). Following the mathematical steps given in
frontier will increase as the distance increases between the min- Ref. [19], it is possible to establish the monotonicity of cos nA
ima of fc and v. with respect to d. It can be shown that, regardless of the value of
Proof. Assume that in a codesign problem the control objective d, an increase in d will always result in a decrease in cos nA; thus,
function fc(da, dc) has an unconstrained minimum in the da-space, we see that an increase in the distance d between the minima of fc
denoted as point D. The CPF v(da) is chosen such that it has an and v will result in an increase in the angle nA. From Theorem 2,
unconstrained minimum in the da-space. The minimum of v(da), it can be stated that an increase in d will result in CPF solutions
denoted as point C, is located at a distance d from point D, as that are farther from the Pareto-optimal solutions to the codesign
shown in Fig. 6. problem, and thus, the theorem is proven. h
Let point A be a solution to the CPF problem using v(da). The Example. Consider the following problem:
distance from point D to point A is denoted as r, and the distance
from point C to point A is denoted as b. The vectors n, r, and s in min wa fa ðd a Þ þ wc fc ðd a ; d c Þ (79)
Fig. 6 are unit vectors.
subject to ga ðd a Þ ¼ 4da21 þ da22  900  0 (80)
The function fc evaluated at point A shall be represented by a
second-order Taylor series expansion about its minimum, point D, 1 1
as follows: gc ðd a ; d c Þ ¼ da21 þ da22  48  0 (81)
50 5
1 where
fcA ¼ fcD þ r2 sT r2 fcD s (73)
2
fa ¼ 0:5da21 þ da22  da1 da2  7da1  7da2 (82)
The function v evaluated at point A shall be represented by a sec-
ond-order Taylor series expansion about its minimum, point C, as 1
fc ¼ ðda1  da2 Þ2 þ ðda1 þ da2 þ d c  10Þ2 þ ðd c  5Þ2 (83)
follows: 9

1 This problem is solved twice, with two different CPFs, which are
vA ¼ vC þ b2 r T r2 vC r (74)
2 then compared

It is then possible to find the gradients of fc and v, evaluated at v1 ðd a Þ ¼ ðda1  5Þ2 þ da22  25 (84)
point A, in the da-space.
v2 ðd a Þ ¼ ðda1  1Þ2 þ ðda2 þ 10Þ2  10 (85)
rfcA ¼ rs T
r2 fcD (75)
The unconstrained minima of the functions fa, fc, v1, and v2 are
rvA ¼ br T r2 vC (76) given in Table 1. The minimum of v1 is located at a distance of
3.54 from the minimum of fc, significantly closer than the mini-
The unit vector r can be expressed as r ¼ b1ðrs  dnÞ, and, there- mum of v2, which is located at a distance of 12.59 from the mini-
fore, Eq. (76) can be rewritten as mum of fc. The solutions found by solving the two CPF sequential
problems are shown in Fig. 7. It can be seen that v1, which obtains
rvA ¼ ðrs  dnÞT r2 vC (77) its minimum closer to that of fc than does v2, produces a closer
match to the simultaneous solution than v2.
From Eq. (40), it is known that It is important to note that, in the case of Theorem 1, no assump-
tions were made about the functional forms of fa, fc, or v. For Theo-
rvA  rfcA rems 2 through 4, the functions were assumed to be convex in the
cos nA ¼   (78) region of interest. This assumption allowed higher-order terms to
krvA krfcA 
be neglected without changing the sign of the Taylor series expan-
sion. If a function was nonconvex, then the results of Theorems 2
through 4 would not necessarily be applicable.

4 Control Proxy Functions for Specific Codesign

Problem Formulations
In Sec. 3, the characteristics of effective CPFs have been
shown. The four theorems presented were based purely on the
mathematics of the functions involved in the problem, with no
consideration for what they might physically represent or what
specific problems could be formulated. Here, specific codesign
problem formulations are considered, and the types of CPFs that
may apply to those problems are studied.
Previous work has shown that, in some cases, the natural fre-
quency of a system can be used as an effective proxy for that

Table 1 Comparison of minima of objective functions and con-

trol proxy functions

fa fc v1 v2

da1 21 2.5 5 1
da2 14 2.5 0 10
Fig. 6 Unconstrained minima of fc and v

091007-6 / Vol. 133, SEPTEMBER 2011 Transactions of the ASME

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 2 1 3T
T @W c @xf
6 xf @da xf þ 2xTf W1
6 1
c
@da1 7
7
6 @W1 @xf 7
6 T
wc 6 x f c
xf þ 2xTf W1 7
Cv ¼ 6 @da2 c
@da2 7
7 (88)
wa 6 .. 7
6 7
6 . 7
4 @W1 @x f
5
xTf c
xf þ 2xTf W1
c
@dan @dan

For vðd a Þ ¼ xTf Wc 1 xf , Cv ¼ wwac rv, and thus, the CPF is perfect.
There are situations, however, when a simpler CPF would suffice.
Consider the case where the final state of the system, xf, is a parameter.
One might expect that when this is the case, a CPF based only on Wc
may be effective, and in some situations this is true, as shown below.
Consider a situation in which the parameter xf has as its only
nonzero component the jth element, i.e., xfj 6¼ 0, xfq ¼ 0 8 q 6¼ j;
this situation corresponds to problems, e.g., in which a system is
to be moved to a final location where it is at rest. A CPF of
 
Fig. 7 Comparison of simultaneous and CPF solutions for two 1
  Wcjj tf
choices of CPF v ¼ Wcjj tf ¼ (89)
detðWc Þ
   
is chosen, where Wc tf  isthe adjoint matrix of Wc(tf), and Wcjj tf is
the (jj)th element of Wc tf . Note that, if the (jj)th element of Wc  tf
system’s ease of control (e.g., Refs. [29–33]), and the natural fre-
 of da, then this is equivalent to maximizing the deter-
is not a function
quency has been successfully used in many structural problems, in
minant of Wc tf . It will be shown here that this is a perfect CPF.
which minimizing vibration is critical. Furthermore, it has been
The gradient of Eq. (89) is given by
demonstrated that, in certain cases, a CPF based on natural fre-
quency is guaranteed to produce the system-optimal results [34]. " 1      #
@W cjj tf @W1 cjj tf @W1 cjj tf
Such a CPF is not always effective, however; in particular, when rv ¼ ; ;…; (90)
the system’s response to a forcing function is sensitive to the arti- @da1 @da2 @dan
fact design variables, a CPF based on natural frequency will not
be effective since it does not capture this behavior. Therefore, The coupling vector Cv is found from Eq. (88)
additional CPFs are necessary to cover a wide range of problems  
of interest. Here, the controllability Grammian matrix, Wc, will be wc 2 @W1 cjj ðt f Þ cjj ðt f Þ
@W 1
Cv ¼ x 2
; …; xfj @da (91)
considered as the basis for a CPF. Previous work has shown that, wa fj @ da1 n

for some problem formulations, there is a relationship between Wc

and the coupling vector Cv [35]. Since there is also a relationship which can be rewritten as
between Cv and an effective CPF, this suggests that a CPF based
on Wc will be effective for some problems. wc 2
Cv ¼ x rv (92)
This section will examine situations in which a CPF based on wa fj
either the time-dependent controllability Grammian matrix, Wc(tf)
or the steady-state controllability Grammian matrix, W c 1 , is per- Thus, as stated, Eq. (89) will be a perfect CPF for this situation.
fect. In all cases, it is assumed that the codesign problem is formu- It has been shown that the coupling vector is related to the con-
lated as in Eqs. (1)–(5), and that the system dynamics are linear trollability Grammian matrix for the case where the control objec-
and can be described in state-space form, i.e., x_ ¼ Ax þ Bu, tive function is the final time and a constraint on control effort is
where A describes the system’s free response, B characterizes the active [35]. The coupling vector for that problem formulation was
system’s response to a forcing function, u is the control input, and found to be parallel to the coupling vector found when the control
x is the vector of system states. objective function is control effort. Therefore, it can then be seen
that the CPF developed above will also apply to such problems.
4.1 Control Proxy Function for the Case of Control Effort 4.2 Control Proxy Function for the Case of Linear Quad-
as Objective. Consider the case in which the control objective is ratic Regulator (LQR): In the LQR problem, a cost function J is
to minimize control effort, defined as to be minimized, where J is given in Eq. (93) below and the
ð tf weighting matrices Q and R denote the relative cost of errors and
fc ¼ ðuðtÞÞ2 dt (86) of control, respectively, [36]
0 ð1

J¼ xðtÞT QxðtÞ þ uðtÞT RuðtÞ dt (93)

It is known from control theory that there exists a minimum possi- 0
ble value for the control effort, regardless of the controller archi-
tecture chosen, and that this limit is dependent on the The coupling vector for the LQR problem can be related to the
controllability Grammian matrix, Wc [36]. Therefore, if an opti- controllability Grammian for a specific choice of the state weight-
mal controller is chosen, then the control objective function fc will ing matrix Q ¼ cBBT [35], where c is a positive scalar value. In
be given by the most general, such LQR case, the initial state x0, W c 1 , and A
all depend on da, and a perfect CPF would take the form of
fc ¼ xTf W1
c xf (87)
vðd a Þ ¼ xT0 AT AW1
c x0 (94)
This expression for fc depends only on da and thus could serve as
a CPF. It has been shown that the coupling vector Cv is related to As in previous cases, however, there are special circumstances
Wc [35], and for the control objective in Eq. (87) is given by where a simpler CPF may be chosen. This may be desirable in

Journal of Mechanical Design SEPTEMBER 2011, Vol. 133 / 091007-7

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 large problems where computational effort must be reduced. Con-
sider the situation where the matrix A is not a function of da. Such
a situation, while not common, may occur when considering the
problem of locating an actuator for canceling vibrations. If the ac-
tuator mass is small, its position will not affect the free response
of the system; however, it will affect the forced response of the
system. Furthermore, assume that the initial state x0 is a parameter
with exactly one nonzero component, i.e., x0j 6¼ 0,
x0q ¼ 0 8 q 6¼ j; 0  q  p, and the matrix W c 1 is diagonal,
which will be the case when a balanced realization is utilized
[36].
For the system described above, select the CPF

vðd a Þ ¼ W 1
cjj (95)

Its gradient is
  Fig. 8 MEMS actuator configuration
@W 1
cjj @W1
cjj @W1
cjj
rv ¼ ; ; …; (96)
@da1 @da2 @dan
displacements and could be useful in a variety of applications,
such as a confocal scanning microscope [37]. The actuator utilizes
and the coupling vector is given by the relation [35]
four electrostatic comb-drive actuators to produce this out-of-
2 1 3T plane displacement. The actuator can be used to produce an angu-
1 2 2 @Wcjj lar deflection of the platform as well, but, here, only the vertical
6 x A 7
6 det A 0j jj @da1 7 displacement of the platform is considered. In order to produce
6 @W 17
this displacement, each of the four comb drives is excited with a
6 1 2 2 cjj 7
wc 6 x0j Ajj 7 voltage, V, resulting in horizontal (in-plane) movement (DX) of
Cv ¼ c6
6 det A @d a 2 7
7 (97)
wa 6 .. 7 the silicon shuttles. The microhinges on the polydimethyl siloxane
6 . 7 (PDMS) platform bend, and the platform moves vertically, or out-
6 17
4 1 2 2 @Wcjj 5 of-plane (DZ). The amount of movement resulting from the comb
x A
det A 0j jj @dan drives’ actuation depends on both the applied voltage, V, and the
physical dimensions of the actuator. Changing the actuator’s
This leads to the relation dimensions results in a different output displacement for the same
applied voltage.
wc c 2 2 The displacement of the actuator, DZ, is given by the equation
Cv ¼ x A rv (98)
wa det A 0j jj
DZ ¼ ðh1 þ h2 Þð1  cos DhÞ þ ðt þ pÞ sin Dh (99)
and therefore, the CPF is perfect for this problem.
In this section, it has been shown that a CPF based on the con- where p, t, h1, and h2 are the hinge dimensions shown in Fig. 9,
trollability Grammian matrix, Wc(tf), can be effective for many and Dh is the angular displacement of the hinge.
problems, in which either fc or an active constraint is dependent The angular displacement Dh can be found from the differential
on control effort. This includes some problems in which the con- equation
trol objective, fc, is the response time of the system. For the spe-
cific problem formulations investigated here, a perfect CPF can be MDh€ þ CDh_ þ KDh ¼ AðDhÞV 2 (100)
formulated based on either the time-dependent or the steady-state
controllability Grammian matrix. As previously stated, these where M, C, K, and A(Dh) are functions of the actuator geometry.
problem formulations are not exhaustive. It is anticipated that Derivations, and the equations for the various masses and stiff-
future work will show that a CPF based on the controllability nesses of the system components, are given in Ref. [37] and repro-
Grammian matrix will be effective for additional codesign prob- duced in Ref. [19]. Alternatively, the system dynamics may be
lem formulations, particularly, those in which the objective func- h iT
tion or constraints are dependent on control effort. written in state-space form, with x ¼ Dh Dh_ , as
However, it may not always be possible to formulate a perfect
CPF based on the control Grammian. For example, consider the
case where fc is the maximum control signal, rather than control
effort. A control signal with a high peak that quickly decays may
result in a lower control effort than a signal with a lower peak that
does not decay as quickly. The relationship between the maximum
control signal and control effort, and the choice of an appropriate
CPF for problems in which the control objective, fc, is based on
the maximum control signal, requires further investigation. It is
expected that a CPF based on the controllability Grammian will
produce results that are near-optimal for a variety of problems,
since it provides a measure of how easily a system is controlled.

5 MEMS Actuator and Controller Optimization

The MEMS actuator considered in this case study was origi-
nally designed by Tung and Kurabayashi [37] and is shown in
Fig. 8. This actuator has been designed to produce out-of-plane Fig. 9 Microhinge structure

091007-8 / Vol. 133, SEPTEMBER 2011 Transactions of the ASME

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 neo ðh1 þ h2 ÞVss2
ga3 ¼ 
d
(107)
kb p2 EPDMS wð2h1 þ h2 Þ3
0
12p2
 
lSi lf
ga4 ¼ DXss  pffiffiffi  o  0 (108)
2 2
EPDMS h1 Dhss
ga5 ¼  rPDMSmax  0 (109)
2p
Fig. 10 Control architecture and system dynamics 3DXss ESi b
ga6 ¼  rSimax  0 (110)
       4l2Si
Dh_ 0 1 Dh 0
¼ þ V2 (101) 1 lm  p  1000 lm (111)
Dh€ K=M C=M Dh_ AðDhÞ=M
1 lm  t  1000 lm (112)
An integral controller with state feedback is applied to the system, 100 lm  l1  1000 lm (113)
as shown in Fig. 10. It is assumed that the angle Dh and the angu-
lar velocity Dh_ can be measured, and that the angle Dh is to be Here, n ¼ 50 is the number of fingers in the comb drive,
controlled. The dynamics of the closed-loop system can then be eo ¼ 8.854e – 12 F=m is the permittivity of vacuum, d ¼ 3 lm is
written as the width of a finger, b ¼ 3 lm is the thickness of the silicon leaf
spring, w ¼ 100 lm is the depth of the micro-hinge, lp ¼ 350 lm
MDh€ þ ðC þ K2 AðDhÞÞDh_ þ ðK þ K1 AðDhÞÞDh is the length of the platform, kb ¼ 0.25 is the beam end condition
ðt
coefficient, EPDMS ¼ 750 kPa is Young’s modulus for PDMS,
 Ki AðDhÞ ðDhr  DhÞds ¼ 0 (102) lSi ¼ 500 lm is the length of the silicon springs, lfo ¼ 50 lm is the
0
initial finger engagement, ESi ¼ 190 GPa is Young’s modulus for
silicon, and rPDMSmax ¼ 2:24MPa and rSimax ¼ 1:5GPa are maxi-
Note that the controller output is u ¼ V2, and that the coefficient A
mum allowable stresses in PDMS and silicon, respectively. The
in Eq. (100) is a function of Dh. Thus, the resulting controller
steady-state actuator displacement, DZss, can be found from Eq.
design problem is nonlinear.
(99). The final time is a parameter, tf ¼ 0.25 ms.
It is possible to define a variety of possible measures of “ease
5.1 Optimization Problem Formulation. The artifact objec- of control,” which might serve as a CPF. Such measures could be
tive is to maximize the final displacement of the actuator, DZf, at a based on natural frequency, the controllability Grammian matrix,
given time tf, where DZf is the peak displacement and is 5% higher the reference signal used to produce the displacement, or a variety
than the steady-state displacement, DZss, see Eq. (103). This of other measures. The choice of which possible CPF to use is
objective function is chosen with the assumption that, while the based on knowledge of the control objective function and on the
final position is important, there may be times in which the actua- fundamental control limitations present in the system. In this case,
tor’s steady-state displacement is important, and, therefore, the the CPF problem to be solved is
relationship between the two is relevant to the problem. Artifact
constraints based on manufacturability, kinematics, mechanical  
DZf2 Wc22 tf
and electrical stability, and stress, are given in Eqs. (105)–(110). min w1 fa þ w2 v ¼ w1 fa þ w2    (114)
The artifact design variables, da, are p, t, and the shuttle length l1, p;t;l1 det W c tf
with the bounds in Eqs. (111)–(113). The control objective is to
minimize the actuation control effort, see Eq. (104). The control subject to the constraints in Eqs. (105)–(110); followed by
design variables are the gains K1, K2, and Ki. There are no control ð tf
constraints. These variables are summarized in Table 2.
min E ¼ ðV ðtÞÞ4 dt (115)
K1 ;K2 ;Ki 0
fa ¼ DZf ¼ 1:05DZss (103)
ð tf ð tf
fc ¼ ðuðtÞÞ2 dt ¼ ðV ðtÞÞ4 dt (104)
0 0

ga1 ¼ t  5h1  0 (105)

lp DXss
ga2 ¼ 910  l1   2t þ 0 (106)
2 2

Table 2 Design and control variables for MEMS actuator

optimization

da p Microhinge length
t Microhinge width
l1 Shuttle length

dc K1 Controller gain for Dh

K2 Controller gain for Dh_
Ki Integral controller gain
Fig. 11 CPF points for optimization of MEMS actuator

Journal of Mechanical Design SEPTEMBER 2011, Vol. 133 / 091007-9

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 to empirically construct a CPF. Such problems are of interest, and
should be a subject of future work.
One can discover CPFs for specific codesign problem formula-
tions. Examples are CPFs based on natural frequency [34] and on
the controllability Grammian matrix presented here. It is antici-
pated that a variety of CPFs could be formulated for a wide range
of problems, motivated by the work presented.

Acknowledgment
This work was partially supported by NSF Grant No. 0625060
and by the Automotive Research Center (ARC). This support is
gratefully acknowledged.

Nomenclature
A ¼ state coefficient matrix determining the unforced
response of a system
A(Dh) ¼ coefficient describing response of MEMS actuator to
an applied voltage
Fig. 12 Displacement and voltage profile for sample MEMS b ¼ thickness of silicon leaf springs in MEMS actuator
actuator design B ¼ state coefficient matrix determining the forced
response of a system
Optimal values of  DZf and E are shown in Fig. 11, with simula- C ¼ generalized damping term for MEMS actuator
tion results for one point shown in Fig. 12. Once a design is d ¼ width of finger in comb drive for MEMS actuator
selected, since the closed-loop system is nonlinear, we can conduct da ¼ vector of artifact design variables
linearization and a local stability analysis to ensure closed-loop dc ¼ vector of control design variables
stability. At each of the points shown, the angle n is calculated in EPDMS ¼ Young’s modulus for PDMS
order to determine whether the point is optimal or near-optimal. ESi ¼ Young’s modulus for silicon
For each point, n ¼ 0, indicating that the CPF points are Pareto fa ¼ artifact objective function
optimal. However, the CPF points require less computational fc ¼ control objective function
effort. Using MATLAB, a typical CPF point required 7 s of computa- ga ¼ artifact inequality constraints
tional time, while a simultaneous solution for that point required gc ¼ control inequality constraints
577 s of computational time. h1 ¼ thickness of micro-hinge
The accuracy of this CPF can be ascertained by examining fc, h2 ¼ height of micro-hinge
whose value is found from Eq. (87), where xf ¼ ½ DZf ; 0  .
T ha ¼ artifact equality constraints
Since for an optimal controller, fc ¼ v, the tradeoff between dis- hc ¼ control equality constraints
placement and control effort can be evaluated using v, prior to the kb ¼ beam end condition coefficient for micro-hinge in
formulation of the control optimization problem. MEMS actuator
K ¼ generalized spring constant for MEMS actuator
K1, K2, Ki ¼ controller gains for MEMS actuator
l1 ¼ length of comb drive in MEMS actuator
6 Concluding Remarks lfo ¼ initial finger engagement in MEMS actuator
A new sequential method for optimization of codesign problems lp ¼ length of PDMS platform in MEMS actuator
was introduced using a CPF. The method provides solutions that lSi ¼ length of silicon springs in MEMS actuator
are identical or close to the Pareto optimal solutions to the codesign M ¼ generalized inertia for MEMS actuator
problem, while the decomposed problem is easier to solve sequen- n ¼ number of fingers in comb drive in MEMS actuator
tially. The method’s effectiveness depends on the choice of CPF. p ¼ length of micro-hinge in MEMS actuator
Guidelines for choosing a CPF and a metric to evaluate the close- t ¼ width of micro-hinge in MEMS actuator
ness of the solutions were developed and illustrated with examples. tf ¼ final time for interval of interest
While the method is presented under some assumptions, such as the u ¼ control input applied to a system
presence of unidirectional coupling, these assumptions are not V ¼ voltage applied to comb drive in MEMS actuator
unique to the codesign problem. Any bi-objective problem exhibit- w ¼ depth of micro-hinge in MEMS actuator
ing unidirectional coupling could be successfully solved for optimal w1 ¼ weight attached to artifact objective function, fa, in
or near-optimal solutions with an appropriate proxy function that CPF problem formulation
satisfies conditions such as those set forth in this work. w2 ¼ weight attached to Control Proxy Function, v, in
Some limitations were specified in the development of the CPF problem formulation
theory. A particularly limiting assumption is that no constraints be wa ¼ Weight attached to artifact objective function, fa, in
active in certain cases. It is expected that the unconstrained case simultaneous problem formulation
provides a bound, and that the performance of a CPF for a con- wc ¼ weight attached to control objective function, fc, in
strained problem will be no worse than the performance of that simultaneous problem formulation
CPF if the constraints were not present. Proof of this conjecture Wc(tf) ¼ Controllability Grammian matrix, evaluated at tf
should be a subject of future work. A further limitation is that the Wc1 ¼ steady-state controllability Grammian matrix
assumption was made that the functional form of the control x ¼ state vector describing a system
objective is known, e.g., that fc can be expressed in terms of da x0 ¼ state vector describing a system at t ¼ 0
and dc. In the case of so-called “black-box” controllers, this is not xf ¼ state vector describing a system at t ¼ tf
true. In these cases, it may be possible to formulate a CPF based Cv ¼ coupling vector
on knowledge of the underlying physics that limits controller per- Cb v ¼ estimate of coupling vector
formance. If this approach is not possible, then an experimental e ¼ distance between a given point in the da-space and
approach could be taken, in which a design of experiments is used the nearest Pareto optimal point

091007-10 / Vol. 133, SEPTEMBER 2011 Transactions of the ASME

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 eo ¼ permittivity of vacuum [16] Fathy, H., Papalambros, P. Y., and Ulsoy, A. G., 2004, “On Combined Plant
and Control Optimization,” In 8th Cairo University International Conference on
f ¼ empirical parameter used in characterization of Mechanical Design and Production, Cairo University.
MEMS actuator [17] Alyaqout, S. F., Papalambros, P. Y., and Ulsoy, A. G., 2005, “Quantification
DX ¼ horizontal displacement of comb drive in MEMS and use of System Coupling in Decomposed Design Optimization Problems,”
actuator Proceedings of the ASME IMECE 2005, ASME, pp. 95–103, Paper No.
IMECE2005-81364.
DZ ¼ vertical displacement of MEMS actuator [18] Peters, D. L., Papalambros, P. Y., and Ulsoy, A. G., 2009, “On Measures of
Dh ¼ angular displacement of micro-hinge in MEMS Coupling Between the Artifact and Controller Optimal Design Problems,” Pro-
actuator ceedings of the ASME Design Engineering Technical Conference and Com-
k ¼ vector of Lagrange multipliers associated with equal- puters in Engineering Conference, ASME, Paper No. DETC 2009-86868.
[19] Peters, D. L., 2010, “Coupling and Controllability in Optimal Design and Con-
ity constraints trol,” Ph.D. thesis, University of Michigan, Ann Arbor, MI.
l ¼ vector of Lagrange multipliers associated with in- [20] Fathy, H., Papalambros, P. Y., Reyer, J., and Ulsoy, A. G., 2001, “On the Plant
equality constraints and Controller Optimization Problems,” Proceedings of the American Control
n ¼ angle between coupling vector and gradient of Con- Conference, ACC.
[21] Fathy, H., 2003, “Combined Plant and Control Optimization: Theory, Strategies
trol Proxy Function and Applications,” Ph.D. thesis, University of Michigan, Ann Arbor, MI.
rPDMSmax ¼ maximum allowable stress in PDMS [22] Reyer, J., 2000, “Combined Embodiment Design and Control Optimization:
rSimax ¼ maximum allowable stress in silicon Effects of Cross-Disciplinary Coupling,” PhD Thesis, University of Michigan,
v ¼ Control Proxy Function (CPF) Ann Arbor, MI, April.
[23] Das, I., and Dennis, J., 1997, “A Closer Look at Drawbacks of Minimizing
References Weighted Sums of Objectives for Pareto Set Generation in Multicriteria Opti-
mization Problems,” Struct. Optim., 14, pp. 63–69.
[1] Haftka, R., Martinovic, Z., Jr., W. H., and Schamel, G., 1986, “An Analytical [24] Messac, A., and Puemi-Sukam, C., 2000, “Aggregate Objective Functions and
and Experimental Study of a Control System’s Sensitivity to Structural Mod- Pareto Frontiers: Required Relationships and Practical Implications,” Optim.
ifications,” AIAA J., 25, pp. 310–315. Eng., 1, pp. 171–188.
[2] Rao, S., and Pan, T., 1990, “Modeling, Control, and Design of Flexible Struc- [25] Kitayama, S., Yamazaki, K., Arakawa, M., and Yamakawa, H., 2009,
tures: A Survey,” Appl. Mech. Rev., 43, pp. 99–117. “Quantitative Trade-off Analysis and Its Application to the Compromise Solu-
[3] Albers, A., and Ottnad, J., 2010, “Integrated Structural and Controller Optimi- tion in the Multi-Objective Design Optimization,” Proceedings of the 8th World
zation in Dynamic Mechatronic Systems,” ASME J. Mech. Des., 132, 041008. Congress on Structural and Multidisciplinary Optimization, WCSMO.
[4] Carley, L., Ganger, G., Guillou, D., and Nagle, D., 2001, “System Design Con- [26] Steuer, R. E., 1986, Multiple Criteria Optimization: Theory, Computation and
siderations for MEMS-Actuated Magnetic-Probe-Based Mass Storage,” IEEE Application, John Wiley and Sons, Inc., New York.
Trans. Magn., 37, pp. 657–662. [27] Kuhn, H. W., and Tucker, A. W., 1950, “Nonlinear Programming,” Proceed-
[5] Oldham, K., Huang, X., Chahwan, A., and Horowitz, R., 2005, “Design, Fabri- ings of the Second Berkeley Symposium on Mathematical Statistics and Proba-
cation and Control of a High-Aspect Ratio Microactuator for Vibration Sup- bility, University of California, pp. 481–492.
pression in a Hard Disk Drive,” Proceedings of the IFAC World Congress. [28] Papalambros, P., and Wilde, D., 2000, Principles of Optimal Design: Modeling
[6] Ravichandran, T., Wang, D., and Heppler, G., 2006, “Simultaneous Plant- and Computation, Cambridge University Press, Cambridge, NY.
Controller Design Optimization of a Two-Link Planar Manipulator,” Mecha- [29] Peters, D. L., Kurabayashi, K., Papalambros, P. Y., and Ulsoy, A. G., 2008,
tronics, 16, pp. 233–242. “Co-Design of a MEMS Actuator and Its Controller Using Frequency Con-
[7] Zhu, Y., Qiu, J., and Tani, J., 2001, “Simultaneous Optimization of a Two-Link straints,” Proceedings of the ASME Dynamic Systems and Control Conferen-
Flexible Robot Arm,” J. Rob. Syst., 18(1), pp. 29–38. ces, ASME. Paper No. DSCC 2008-2212.
[8] Udengaard, M., and Iagnemma, K., 2009, “Analysis, Design, and Control of an [30] Hale, A., Lisowski, R., and Dahl, W., 1985, “Optimal Simultaneous Structural
Omnidirectional Mobile Robot in Rough Terrain,” ASME J. Mech. Des., 131, and Control Design of Maneuvering Flexible Spacecraft,” J. Guid., Control
121002. Dyn., 8, pp. 86–93.
[9] Ou, J., and Kikuchi, N., 1996, “Integrated Optimal Structural and Vibration [31] Khot, N., and Abhyankar, N., 1993, “Integrated Optimum Structural and Con-
Control Design,” Struct. Optim., 12, pp. 209–216. trol Design,” In Structural Optimization: Status and Promise, M. P. Kamat, ed.
[10] Kosut, R., Kabuli, G., Morrison, S., and Harn, Y., 1990, “Simultaneous Control [32] Bodden, D., and Junkins, J., 1985, “Eigenvalue Optimization Algorithms for
and Structure Design for Large Space Structures,” Proceedings of the American Structure=Controller Design Iterations,” J. Guid., Control Dyn., 8, pp.
Control Conference, IEEE, pp. 860–865 697–706.
[11] Milman, M., Salama, M., Scheid, R., Bruno, R., and Gibson, J., 1991, [33] Yee, E., and Tsuei, Y., 1991, “Method for Shifting Natural Frequencies of
“Combined Control-Structural Optimization,” Comput. Mech., 8, pp. 1–18. Damped Mechanical Systems,” AIAA J., 29(11), pp. 1973–1977.
[12] Sobieszczanski-Sobieski, J., and Haftka, R., 1997, “Multidisciplinary Aero- [34] Peters, D. L., Papalambros, P. Y., and Ulsoy, A. G., 2010, “Sequential Co-
space Design Optimization: Survey of Recent Developments,” Struct. Optim., Design of an Artifact and Its Controller via Control Proxy Functions,” Proceed-
14, pp. 1–23. ings of the 5th IFAC Symposium on Mechatronic Systems, IFAC.
[13] Haftka, R., 1990, “Integrated Structure=Control Optimization of Space [35] Peters, D. L., Papalambros, P. Y., and Ulsoy, A. G., 2010, “Relationship
Structures,” Proceedings of the AIAA Dynamics Specialists Conference, AIAA, Between Coupling and the Controllability Grammian in Co-Design Problems,”
pp. 1–9 Proceedings of the American Control Conference, ASME.
[14] Fathy, H., Bortoff, S., Copeland, G., Papalambros, P. Y., and Ulsoy, A. G., [36] Skogestad, S., and Postlethwaite, I., 2005, Multivariable Feedback Control:
2002, “Nested Optimization of an Elevator and its Gain-Scheduled LQG Con- Analysis and Design, John Wiley and Sons, Ltd., West Sussex, UK.
troller,” Proceedings of the ASME IMECE 2002, ASME, pp. 119–126, Paper [37] Tung, Y., and Kurabayashi, K., 2005, “A Single-Layer PDMS-on-Silicon
No. IMECE2002-34273. Hybrid Microactuator With Multi-Axis Out-of-Plane Motion Capabilities—Part
[15] Bloebaum, C., 1995, “Coupling Strength-Based System Reduction for Complex I: Design and Analysis,” J. Microelectromech. Syst., 14(3), pp. 548–557.
Engineering Design,” Struct.Optim., 10, pp. 113–121.

Journal of Mechanical Design SEPTEMBER 2011, Vol. 133 / 091007-11

Downloaded 16 Sep 2011 to 141.213.155.99. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm