2015 American Control Conference

Palmer House Hilton

July 1-3, 2015. Chicago, IL, USA

A Gradient-based Approach for Optimal Plant Controller Co-Design

Yuh-Shyang Wang and Yebin Wang

Abstract— This paper proposes a gradient-based iterative work for some specific cases. For example, the SDP method
algorithm for optimal co-design of a linear physical plant and a in [5] assumes that the system matrices are affine in the
controller. The proposed algorithm does not rely on the common plant parameters, which results in limited applicability. For
linear parameterization assumption, and thus is applicable to
a broader class of problems. The convergence of the algorithm instance, when transforming a linear second order mass-
and the verification procedure for a local minimum are given. spring-damper system into a state space form, the system
Numerical examples show that our algorithm is comparable to matrices depend on the plant parameter in a nonlinear way.
other complicated algorithms in terms of the performance, but The SDP method will not be applicable in this case unless
can deal with a more general class of problems. re-parametrization.
I. I NTRODUCTION In order to address the above-mentioned two issues, we
propose a gradient-based algorithm which is more general
The process of designing a control system can be roughly and less conservative. The algorithm can handle nonlinear
divided into two tasks: plant design and controller synthesis. parametrization, and is applicable for both state and output
In conventional design process, these two tasks are carried feedback, continuous-time and discrete-time systems. Nu-
out in a sequential manner, that is, design the plant first and merical examples suggest that our algorithm is less con-
then synthesize the controller for the designed plant model. servative than the system-equivalence-based method, and is
This separation simplifies the design process at the expense comparable to the SDP method on the specialized problem.
of performance degradation. In fact, the co-design of the The rest of this paper is organized as follows. In Section II,
plant and the controller can improve the overall system per- we formulate the optimal plant controller co-design problem
formance and robustness compared to the sequential design as an extension of the optimal control problem. An iterative
procedure. The idea of co-design has been applied to a wide algorithm is proposed in Section III to approach the co-
range of areas, including aerospace crafts [1], smart buildings design problem. In Section IV, we analyze the convergence
[2], and electric motors [3]. property of our algorithm, and extend the algorithm to
A common way to approach the plant controller co-design structured controller tuning problems. Numerical examples
problem is solving the optimal control problem and the are provided in Section V to illustrate the effectiveness of
optimal plant parameter selection problem alternately and our method. Finally, Section VI ends with conclusion and
iteratively. Although the iterative design procedure does sim- offers some future research directions.
plify the original co-design problem, tremendous difficulty
remains due to the non-convex nature of the optimal plant II. P RELIMINARIES AND P ROBLEM F ORMULATION
parameter selection problem. The non-convexity is present In this section, we introduce the common formulation
even if the plant is linearly dependent on its parameters. for H2 optimal control problem, and formulate the plant
Several methods have been proposed to address the non- controller co-design problem. Then, we present the iterative
convex constraints. For instance, work [2], [4] convexify the design procedure where the co-design problem is decom-
non-convex problem by adding an extra system-equivalence posed into an optimal control problem and a plant parameter
constraint. Work [5] tries to relax the system-equivalence selection problem, solved alternately.
constraint by establishing a set of convex constraints with
A. H2 Optimal Control Formulation
the hope to better approximate the original non-convex
constraints. The convexification usually leads to a standard Consider a generalized continuous time linear time-
semi-definite programming (SDP) problem. invariant (LTI) system
There are two primary limitations for these methods. First, ẋ = Ax + B1 w + B2 u
the additional constraint usually leads to a reduced feasible z = C1 x + D12 u
set, and conservative design. Specifically, the plant parameter
can converge to an optimal point in a convex subset, but y = C2 x + D21 w (1)
not necessary a stationary point in the original set. Second, where x ∈ Rnx is the state vector, w ∈ Rnw the external
the aforementioned convexification-based approaches only disturbance, u ∈ Rnu the control action, z ∈ Rnz the
regulated output, and y ∈ Rny the measurement. The
Y.-S. Wang is with the department of Control and Dynamical Sys-
tems, California Institute of Technology, Pasadena, CA 91125, USA
system (1) can also be written in a compact form as
([email protected]). This work was done while Y.-S. Wang was
 
A B1 B2  
an intern with Mitsubishi Electric Research Laboratories. P P12
Y. Wang is with Mitsubishi Electric Research Laboratories, Cambridge, P =  C1 0 D12  = 11
P21 P22
MA 02139, USA ([email protected]). C2 D21 0
978-1-4799-8684-2/$31.00 ©2015 AACC 3249
Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on August 21,2024 at 06:14:35 UTC from IEEE Xplore. Restrictions apply.
 where Pij = Ci (sI − A)−1 Bj + Dij . Let Twz be the closed- tunable. Equation (5) can be written in state space form by
loop transfer function from w to z. The goal of H2 optimal introducing a new state vector x = [x>s ẋ> >
s ] as
control is to find a dynamic output feedback control law    
u = Ky such that the H2 norm of Twz is minimized. 0 I 0
ẋ = x + u
Mathematically, the problem can be written as −Ms−1 Ks −Ms−1 Ds Ms−1 B
= A(θ)x + B2 (θ)u (6)
minimize ||Twz (K)||H2 
y = C 0 x

K
subject to K stabilizes P (2) = C2 x.
in which Twz (K) = P11 + P12 K(I − P22 K)−1 P21 . The Equation (6) shows that the plant model is parameterized by
problem (2) can be solved analytically, and the solution can the system parameter θ. Note that due to the inverse operation
be found in standard textbooks such as [6]. Ms−1 in (6), the matrices A(θ) and B2 (θ) usually depend
on the parameter θ in a nonlinear way. The co-design of the
B. Optimal Plant Controller Co-design mass-spring-damper system and the controller can then be
We now consider a plant model as follows formulated in the form of (4).
 
A(θ) B1 (θ) B2 (θ) C. Iterative Design Procedure
P (θ) =  C1 (θ) 0 D12 (θ)  . (3)
A heuristic to deal with (4) is solving the optimal control
C2 (θ) D21 (θ) 0
problem and the optimal parameter selection problem alter-
The plant (3) is parameterized by a vector θ ∈ Rnθ that nately and iteratively. Specifically, we solve (2) for a fixed
can be designed. We assume that all state space matrices plant parameter θ to get a controller K, and then fix K and
in (3) are differentiable with respect to θ, and θ belongs to a solve the plant parameter selection problem as
convex compact set Θ. In particular, we emphasize the fact
that the state space matrices in (3) can be any differentiable minimize ||Twz (θ, K)||H2
θ
nonlinear function of θ. Without losing much generality, we subject to θ ∈ Θ, K stabilizes P (θ). (7)
assume Θ = {θ|θmin ≤ θ ≤ θmax }. We also assume that the
pair [A(θ), B2 (θ)] is stabilizable and the pair [A(θ), C2 (θ)] We solve (2) and (7) repetitively until the iteration converges.
is detectable for all θ ∈ Θ. As mentioned in Section I, (7) is non-convex even when
The co-design objective is to determine an output feedback the system matrices are θ-affine. The common strategy in
control law u = Ky and the plant parameter θ simulta- both [4] and [5] is adding some conservative constraints
neously such that the H2 norm of the closed-loop transfer to make the modified problem convex. Then a sub-optimal
function Twz is minimized. Specifically, the optimal co- solution can be obtained by standard convex optimization
design problem can be formulated as technique. This strategy however suffers two primary limita-
tions: conservative design, and limited range of applicability.
minimize ||Twz (θ, K)||H2 Our key observation is that there is no need to solve
θ,K
subject to θ ∈ Θ, K stabilizes P (θ). (4) either (7) or its analogs exactly at each iteration. Instead, we
propose to update a new plant parameter θnew as long as it
Although problem (4) looks similar to (2), it is substantially results in significant performance improvement. With this in
more difficult to solve. Even when the state space matrices mind, we develop a plant parameter update algorithm based
in (3) are affine in θ and the controller is state feedback, on gradient computation to resolve the above-mentioned
finding a stabilizing pair (K, θ) is known to be NP-hard [7]. two limitations. Note that the gradient-based algorithm is
We now give a common engineering example that can be applicable to a more general class of problem, as long
formulated in the form of (4). as the system matrices are differentiable with respect to
Example 1: Consider a second order mass-spring-damper the plant parameter. Other methods using Hessian or its
system approximations, for instance Newton or Gaussian-Newton,
can also be applied to update the plant parameter.
Ms ẍs + Ds ẋs + Ks xs = Bu
y = Cxs (5) III. C O - DESIGN A LGORITHM
A. System Parameter Update
During the system design process, not only the con-
trol law u = Ky but also the physical plant parameters We begin with a given plant parameter θ and a fixed
(Ms , Ds , Ks ) can be chosen within some range in order stabilizing controller K. Our goal is to find a θnew such
to meet the performance specification. One simple case is that the closed-loop system performance has a sufficient
defining the system parameter vector θ by the collection improvement. Denote the objective function as f (θ) =
of all entries of the matrices (Ms , Ds , Ks ), that is, θ = ||Twz (θ, K)||2H2 for all θ. The plant parameter update is given
[{Ms }ij , {Ds }ij , {Ks }ij ] ∈ Θ. This corresponds to the by the rule
case that all the entries of (Ms , Ds , Ks ) are independently θnew = PΘ (θ + αp) (8)

3250
Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on August 21,2024 at 06:14:35 UTC from IEEE Xplore. Restrictions apply.
 where α is a positive scalar representing the step length, p where
the search direction, and PΘ (.) the projection operator on the 
A(θ) + B2 (θ)DK C2 (θ) B2 (θ)CK

set Θ. In particular, we use the steepest descent direction p = A(θ, K) =
BK C2 (θ) AK
−∇f (θ) based on the gradient of the cost function. Unlike  
the usual line search methods for nonlinear optimization, we B1 (θ) + B2 (θ)DK D21 (θ)
B(θ, K) =
only perform the update (8) once, and then turn back to BK D21 (θ)
 
solve the optimal control problem (2) with the updated plant C(θ, K) = C1 (θ) + D12 (θ)DK C2 (θ) D12 (θ)CK
parameter θnew . D(θ, K) = D12 (θ)DK D21 (θ).
We now discuss how to find an appropriate step length α
with sufficient improvement. A common way to choose the As K stabilizes P (θ), Twz is always a stable transfer func-
step length α is by the Wolfe conditions [8] given by tion and A in (12) is Hurwitz. Denote θi the i-th component
of θ and θ̂i the unit vector in the direction of θi . The gradient
f (θnew ) ≤ f (θ) + c1 α∇f (θ)> p (9) can be computed as
>
∇f (θnew ) p ≥ c2 ∇f (θ)> p (10) nθ
X ∂ < Twz , Twz >
∇f (θ) = θ̂i
for some 0 < c1 < c2 < 1. Equation (9) ensures that the ∂θi
i=1
search gives an improvement when p is a descent direction. nθ
Note that (9) will be satisfied for any small α, in which case
X ∂Twz
= 2 < Twz , > θ̂i . (13)
the improvement may be limited. Therefore, the curvature i=1
∂θi
condition (10) is added to ensure that the chosen step
where < ., . > is the inner product defined on the H2 space.
length is not too conservative. In practice, we can use the
The partial derivative in (13) can be calculated by
backtracking technique to dispense the condition (10) [8].
The idea is to set α to a large value initially, and decrease the ∂Twz ∂C(sI − A)−1 B + D
=
value until (9) is satisfied. When Θ = {θ|θmin ≤ θ ≤ θmax } ∂θi ∂θi
and p 6= 0, there exists a finite ᾱ such that PΘ (θ + αp) = ∂C ∂B
= (sI − A)−1 B + C(sI − A)−1 (14)
PΘ (θ + ᾱp) for all α ≥ ᾱ. The backtracking technique can ∂θi ∂θi
start with this initial value. The complete system parameter ∂D −1 ∂A
update algorithm with backtracking is presented as follows. + + C(sI − A) (sI − A)−1 B
∂θi ∂θi
where the last term in (14) is derived by considering the
following identity [9]
Algorithm 1: System Parameter Update
Given θ, K, f (θ) = ||Twz (θ, K)||2H2 ; ∂Y −1 ∂Y −1
= −Y −1 Y . (15)
Compute the gradient ∇f (θ) and set p = −∇f (θ) ; ∂x ∂x
if p = 0 then Note that each component in (14) can be computed from
θnew = θ ; the state space realization of Twz by simple state space
else manipulation. For instance, the state space representation for
Compute ᾱ and set α = ᾱ ; the last term of (14) is given by
Choose ρ ∈ (0, 1), c1 ∈ (0, 1) ; ∂A
 
A ∂θ 0
Set θnew = PΘ (θ + αp) ; ∂A i
C(sI − A)−1 (sI − A)−1 B =  0 A B  . (16)
while f (θnew ) > f (θ) + c1 α∇f (θ)> p do ∂θi
C 0 0
α = ρα ;
θnew = PΘ (θ + αp) ; We now exemplify how to compute the gradient with a
second order mass-spring-damper system.
Example 2: Consider the system (6). When H2 optimal
controller is used, DK = 0. The state space representation
B. Gradient Computation for Twz can be simplified into
 
We now present an efficient way to compute the gradient   A(θ) B2 (θ)CK B1
A B
∇f (θ) in Algorithm 1. We begin with a state space repre- =  BK C2 AK BK D21  .
C D
sentation of the H2 optimal controller C1 D12 CK 0
 
AK B K Note that A is the only term that depends on θ, so the partial
K= . (11)
CK DK derivative ∂T∂θi is just (16), in which case we only need
wz

∂A
The closed loop transfer function Twz can be written in state to compute ∂θi . This can be carried out by computing the
space form as partial derivative for A and B2 in (6) directly. For example,
  we have  
A(θ, K) B(θ, K) ∂B2 (θ) 0
Twz (θ, K) = (12) =
C(θ, K) D(θ, K) ∂θi Ms−1 ∂M s −1
∂θi Ms B

3251
Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on August 21,2024 at 06:14:35 UTC from IEEE Xplore. Restrictions apply.
 where the identity (15) is applied to Ms−1 . The partial IV. A NALYSIS AND E XTENSION
derivative for A can be computed in a similar way.
In this section, we analyze the property of Algorithm 2.
The last thing to do is computing the inner product in (13). In particular, we show some convergence results for the
From (14) and the fact that A is Hurwitz, we know that algorithm. We then outline the procedure to check whether
∂Twz
∂θi is stable. Assume that D = 0, the inner product can be a stationary point is a local minimum. Finally, we explain
computed in the state space form by forming a Lyapunov- how to use this algorithm to tune a structured controller.
like equation, similar to the state space method to compute
the H2 norm in [6]. Specifically, let A. Convergence Analysis
The following lemma states that the cost function is
 
∂Twz Ai Bi
= . monotonically decreasing at each step in Algorithm 2.
∂θi Ci 0
Lemma 1: We have
The inner product can be computed by
0 ≤ f (θ(k+1) , K (k+1) ) ≤ f (θ(k+1) , K (k) ) ≤ f (θ(k) , K (k) )
∂Twz (18)
< Twz , >= trace(CLCi> )
∂θi for all non-negative integer k.
Proof: There are totally three inequalities in (18). The
where L is the solution of the equation
first inequality comes from the fact that the objective (norm)
AL + LA> > is lower bounded by 0. The second inequality holds since
i + BBi = 0. (17)
K (k+1) is the optimal controller for P (θ(k+1) ). The third
Equation (17) can be solved algebraically by the technique of inequality comes from the condition (9) on plant parameter
vectorization. When the plant and controller are discrete time update.
systems, instead of (17), we solve a discrete time Lyapunov- Next, we show that the cost sequence {f (θ(k) , K (k) )}∞ k=0
like equation given by converges.
Lemma 2: The sequence {f (θ(k) , K (k) )}∞ k=0 converges to
ALA> >
i − L + BBi = 0. a value f ∗ .
Proof: From Lemma 1, {f (θ(k) , K (k) )}∞ k=0 is a mono-
C. Iterative Co-design Algorithm tonically decreasing sequence lower bounded by 0. Thus this
We now complete the iterative co-design algorithm by sequence converges to some value f ∗ .
combining optimal control problem and the parameter update If the initial cost f (θ(0) , K (0) ) is nonzero, then Lemma 2
algorithm. From now on, we use θ(k) to represent the vector implies that the sequence of relative improvement ra-
of the system parameter in the k-th iteration, and K (k) the tio {[f (θ(k) , K (k) ) − f (θ(k+1) , K (k+1) )]/f (θ(0) , K (0) )}∞
k=0
optimal controller for the system P (θ(k) ). For a given pair converges to 0. We can then infer that Algorithm 2 terminate
(θ(k) , K (k) ), we can use Algorithm 1 to compute θ(k+1) = in a finite number of steps for any  > 0.
θnew for a fixed K (k) . Denote the objective function as The sequence P∞ of the gradient have the following property.
(k)
f (θ, K) = ||Twz (θ, K)||2H2 . We assume the initial parameter Lemma 3: k=0 α ||∇f (θ(k) )||2 < ∞
θ(0) is given and the initial cost f (θ(0) , K (0) ) is finite. The Proof: Since p = −∇f (θ(k) ), we can rearrange (9)
(k)

stopping criteria is based on the relative improvement ratio and get

[f (θ(k) , K (k) )−f (θ(k+1) , K (k+1) )]/f (θ(0) , K (0) ). For other
c1 α(k) ||∇f (θ(k) )||2 ≤ f (θ(k) , K (k) ) − f (θ(k+1) , K (k) )
purpose, the stopping criteria can also be based on other
measure such as ||θ(k) − θ(k+1) ||. The proposed iterative co- ≤ f (θ(k) , K (k) ) − f (θ(k+1) , K (k+1) ). (19)
design method is summarized in Algorithm 2.
Summing over all positive integer k, the right-hand-side
of (19) is equal to f (θ(0) (0)
P,∞K (k) ) − f ∗ , which is finite. As
Algorithm 2: Co-design Algorithm c1 is nonzero, we have k=0 α ||∇f (θ(k) )||2 < ∞.
Choose a threshold  ∈ (0, 1) ; Lastly, we consider the sequence of the plant parameter
Initialize θ(0) = θ0 , K (0) the optimal controller for {θ(k) }∞
k=0 . P∞ (k+1)
P (θ(0) ) ; Lemma 4: k=0 ||θ − θ(k) ||2 < ∞
Set r = 1, k = 0 ; Proof: We have
while r >  do
||θ(k+1) − θ(k) ||2 = ||PΘ (θ(k) + α(k) p(k) ) − θ(k) ||2
Calculate θ(k+1) from (θ(k) , K (k) ) by Algorithm 1 ;
Compute the optimal controller K (k+1) for ≤ ||θ(k) + α(k) p(k) − θ(k) ||2
P (θ(k+1) ) by (2); = ||α(k) ∇f (θ(k) )||2 (20)
(k) (k) (k+1) (k+1)
f (θ ,K )−f (θ ,K )
Compute r = f (θ (0) ,K (0) )
; where the inequality holds since PΘ is a projection on the
Set k = k + 1 ; set Θ = {θ|θmin ≤ θ ≤ θmax }. As the upper bound for all
α(k) is finite, we can prove this lemma by summing over all
k on both sides of (20) and applying Lemma 3.

3252
Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on August 21,2024 at 06:14:35 UTC from IEEE Xplore. Restrictions apply.
 B. Second Order Condition and dB , dL , mB , mL , kB are system parameters that can be
In this subsection, we assume the existence of a stationary designed. The co-design objective is to improve the track
point (θ∗ , K ∗ ) such that ∇f (θ∗ ) = 0 for a fixed K ∗ , where performance during load positioning tasks. Let yd be the
K ∗ is the optimal controller for P (θ∗ ). We now outline the desired constant output. We define x1 = xL −yd , x2 = ẋL +
procedure to check whether the stationary point is a local ẋB , x3 = xB , x4 = ẋB , and introduce the state vector x =
minimum. [x1 , x2 , x3 , x4 ]> . To make fair comparison, we take dL = 10
kB dB >
To check whether (θ(∗) , K (∗) ) is a co-design local min- and define the system parameter θ = [ m1L , m1B , m B
, mB ] .
imum, we need to treat the state space matrices of K ∗ as The range of θ is given by the set Θ = {θ|θmin ≤
parameters as well. Specifically, we define the augmented θ ≤ θmax } with θmin = [0.3333, 0.04, 0.4, 0.004]> and
parameter vector θ̄ = [θ {AK }ij {BK }ij {CK }ij {DK }ij ]. θmax = [1, 0.0667, 1.3333, 0.0667]> . Choosing the same
Then, we compute the Hessian of the cost function cost function as the one in [5], system (21) can be written
∇2θ̄ f (θ∗ , K ∗ ). If the stationary point θ∗ is on the boundary of in the form of (3) as
the set Θ, we compute the bordered Hessian matrix for con-
       
0 1 0 0 −1 0
strained optimization instead. If the Hessian or the bordered  0 −10(θ1 + θ2 ) θ3 θ4 
  0  θ1 + θ2  
    
 
Hessian is a positive definite matrix, then this stationary point  0 0 0 1  0  0  
is a local minimum. The procedure of computing the Hessian
 
 0
√10θ2 −θ3  −θ4 0  −θ2  

is similar to the gradient computation in Section III-B.

 1000 0 0 0 √ 0 
 0 
C. Structured Controller Tuning
 0 0 0 0 0.1 
I4 0 0
The parameter update algorithm can also be applied to
structured controller tuning problems by treating the entries We choose ρ = 12 and c1 = 10−4 in Algorithm 1,
of (AK , BK , CK , DK ) in (12) as plant parameters. Addi- while these values do not affect the performance of our
tionally, we can impose structured constraints on the state algorithm much. The threshold value  in Algorithm 2 is
space matrices, including sparsity constraints or lower and chosen to be 10−4 and 10−8 for two different trials. The
upper bounds. proposed algorithm is compared to the system-equivalence-
Consider the static output feedback control problem. This based and the SDP methods in [2], [5]. We summarize the
corresponds to the case that AK = 0, BK = 0, and CK = 0. optimal point, total cost, and number of iterations of each
It is known that finding a stabilizing static output feedback method in Table I. It can be seen that the performance of our
controller for a fixed plant is NP-hard [10]. Assuming that gradient algorithm is comparable to the SDP method, and is
the static output feedback control problem has a stabilizing much better than the system-equivalence-based algorithm. In
solution, we can apply our algorithm to tune the stabilizing addition, our algorithm converges faster than the SDP one. In
controller and improve the system performance by treating particular, when  = 10−4 , the algorithm converges in only
the entries of DK as the plant parameter. 2 steps within a second. Finally, it should be noted that the
For a proportional-integral (PI) controller, we have AK = SDP method is developed specifically for state feedback with
0. For a completely decentralized controller, we impose the affine plant parameterization, while our gradient algorithm
appropriate sparsity patterns on (AK , BK , CK , DK ). The can handle a much broader class of problems.
tuning process is similar to that for the static output feedback.
TABLE I
V. N UMERICAL E XAMPLES C OMPARISON BETWEEN D IFFERENT A LGORITHMS
Two numerical examples are performed to illustrate the
effectiveness of our algorithm. The first one is a load-
Original S.E.-based SDP  = 10−4  = 10−8
positioning system [11]. We show that our algorithm have θ1 0.5 0.6667 1 1 1
comparable performance to the one in [5], which is devel- θ2 0.05 0.0667 0.0667 0.0667 0.0667
oped specifically for state feedback with affine plant pa- θ3 0.75 0.75 0.4393 0.4 0.4
rameterization. The second experiment is a real engineering θ4 0.025 0.025 0.0369 0.004 0.009
Cost 215.3952 193.1239 169.5836 169.5756 169.5733
problem about chip mounter plant controller co-design. Iter. - - 24 2 8

A. Load-Positioning System
The dynamics of the load-positioning system are given
B. Chip Mounter Machine
in [11] and recited as
The proposed algorithm is applied to solve a chip mounter
1 1 kB dB
ẍL = (u − dL ẋL )( + )+ xB + ẋB co-design problem as shown in Fig. 1. The control objective
mL mB mB mB is to move the head and beam of the mounter to track a
(21)
1 kB dB
ẍB = (dL ẋL − u) − xB − ẋB reference trajectory while keeping the vibration of the base
mB mB mB small. Thus, we define the objective function as a quadratic
where xL is the relative displacement of the load with function weighting tracking error, base vibration, and control
respect to the platform, xB the displacement of the platform, effort. The system dynamics are given in the form of (5),

3253
Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on August 21,2024 at 06:14:35 UTC from IEEE Xplore. Restrictions apply.
 The 10th order model

Head X-axis(t) addition, as indicated in Fig. 3, the control effort to achieve

Base Y-axis(t) the desired tracking performance is reduced as well. The
maximum overshoot of the base vibration is however slightly
Beam Y-axis(t) increase after the co-design. This is understood as the natural
Z outcome of the trade-off between tracking performance and
base vibration suppression. Basically, the system tends to
Y (t) : translation increase the base mass and stiffness when base vibration
suppression is emphasized.
Base X-axis(t)
The degree of performance improvement and the parame-
X ter selection results highly depend on the relative penalty
on tracking error, base vibration, and control effort. This
Fig. 1. Chip mounter
means the cost function shall be carefully chosen so that the
1 co-design yields meaningful results. Roughly speaking, we
0.8
can treat plant parameter selection as passive control design.
When the active control is expensive, we depend more on
Position

0.6
passive control design. So the plant parameter selection will
0.4 have a significant effect on system performance. On the
0.2 Before Co-design other hand, if active control is cheap, then there is not much
After Co-design
0
difference before and after co-design.
0 0.05 0.1 0.15 0.2
Time (sec) VI. C ONCLUSION
In this paper, we proposed a plant controller co-design
Fig. 2. Tracking performance with respect to a step signal.
algorithm for a general linear time invariant systems. The
1.5
algorithm has a wide range of applications, including state
Before Co-design and output feedback control designs, continuous time and
After Co-design
discrete time systems, and with linear or nonlinear plant
Control action

1
parameterizations. We shown some convergence properties
of the algorithm, and outlined a test to check whether
0.5
a stationary point is a local optimum. In addition to the
generality of our algorithm, simulation also verified that our
0
0 0.05 0.1 0.15 0.2 algorithm results in comparable system performance to the
Time (sec)
SDP-based method.
Fig. 3. Control action for step tracking. R EFERENCES
[1] A. L. Hale, W. Dahl, and J. Lisowski, “Optimal simultaneous structural
#10-3 and control design of maneuvering flexible spacecraft,” Journal of
1
Guidance, Control, and Dynamics, 1985.
0.5 [2] J. Lu and R. E. Skelton, “Integrating structure and control design
to achieve mixed H2 /H∞ performance,” International Journal of
Position

0
Control, 2000.
-0.5 [3] J. A. Reyer and P. Y. Papalambros, “Combined optimal design and
control with application to an electric DC motor,” Journal of Mechan-
-1 Before Co-design
ical Design, vol. 124, no. 2, pp. 183–191, 2002.
After Co-design
-1.5 [4] K. M. Grigoriadis, M. J. Carpenter, G. Zhu, and R. E. Skelton,
0 0.05 0.1 0.15 0.2 “Optimal redesign of linear systems,” in 1993 IEEE American Control
Time (sec) Conference (ACC), 1993.
[5] Y. Jiang, Y. Wang, S. A. Bortoff, and Z.-P. Jiang, “An iterative
Fig. 4. Base vibration when performing step tracking. approach to the optimal co-design of linear control systems,” Interna-
tional Journal of Control, 2014, submitted.
[6] K. Zhou, J. C. Doyle, and K. Glover, Robust and optimal control.
Prentice Hall New Jersey, 1996.
which can be further rewritten in state space form as (6). [7] O. Toker and H. Özbay, “On the np-hardness of solving bilinear
The mass, damping ratio, coefficient of friction, and the matrix inequalities and simulataneous stabilization with static output
feedback,” in 1995 IEEE American Control Conference (ACC), Jun.
stiffness of the system are free to adjust within ±25% from 1995.
its nominal value. As in (6), the state space matrices A(θ) [8] J. Nocedal and S. J. Wright, Numerical Optimization. Springer, 2006.
and B2 (θ) depend on the plant parameter in a nonlinear way. [9] K. B. Petersen and M. S. Pedersen, “The matrix cookbook,”
2012. [Online]. Available: http://www.mit.edu/∼wingated/stuff i use/
Details of the system model are proprietary information and matrix cookbook.pdf
thus omitted. [10] V. Blondel and J. N. Tsitsiklis, “NP-hardness of some linear control
Figures 2-4 show the comparison results before and after design problems,” SIAM J. Control Optim., vol. 35, no. 6, pp. 2118–
2127, 1997.
co-design. It can be seen from Fig. 2 that the tracking per- [11] V. Shilpiekandula, S. A. Bortoff, J. C. Barnwell, and K. El-Rifai, “Load
formance after co-design improves remarkably. The settling positioning in the presence of base vibrations,” in 2012 IEEE American
time is shortened to almost half of the original one. In Control Conference (ACC), Jun. 2012.

3254
Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on August 21,2024 at 06:14:35 UTC from IEEE Xplore. Restrictions apply.