Multi-objective optimization

for control design

James Whidborne
Centre for Aeronautics
Cranfield University

Sensors & their Application XVIII

Queen Mary University of London
12–13 September 2016
Outline

• Conflicts and Trade-offs in Control Systems

• Multi-objective Optimization Problem
• Method of Inequalities (MOI)
• Multi-Objective Genetic Algorithms (MOGA)
• Mixed Optimization
• Convex Multi-objective Optimization
• Design Example - Maglev Suspension Controller Design
• References

Slides will be available from publi . ranfield.a .uk/eh3081/

Multi-objective optimization for control design — 1/63

Abstract

Control system design solutions require the right balance between

conflicting requirements such as cost, complexity, robustness and
performance. Hence to quantify the design process, it is often
appropriate to formulate control system design problems as
multi-objective optimization problems. There are a number of
different approaches to solving such problems. The presentation will
introduce the concept of multiobjective optimization, and outline
several approaches familiar to the presenter, including the Method of
Inequalities, the Multi-Objective Genetic Algorithm and
multiobjective convex optimization. Various tradeoffs in control
design will also be explored, in particular those between the quality
of the sensors and the control system performance. The idea of
mixed-optimization is also introduced. The methods are illustrated
with examples from flight control and maglev suspension control.

Multi-objective optimization for control design — 2/63

Conflicts and Trade-offs in Control Systems

• In general, engineering design consists of obtaining the right balance

between conflicting cost, design and performance requirements
• There are trade-offs to be made between conflicting requirements
• Hence to quantify design, it is appropriate to formulate problem as a
multi-objective problem
• There exist trade-offs in control systems design

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 3/63
Example - simple servo system

Consider a simple servo system with a proportional controller k with negative

feedback

plant/motor
controller
✲
reference ✲ k ✲ 1 output
✲
+ s(s+a)
✻−

Design parameter is controller gain k

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 4/63
Example - simple servo system

Plant: Step Response

From: U(1)
1.5

1
G(s) = s(s+a)

Proportional controller k gives

closed loop system: 1

Amplitude
k

To: Y(1)
T (s) = s2 +as+k

Fractional overshoot to step 0.5

response:
√
4k −a2
Mp = e−aπ/
0
0 5 10 15 20 25
Time-to-peak Time (sec.)

Typical step response of second order system showing

tp = √ 2π for k > (a2 /4)
4k −a2 peak overshoot and time-to-peak

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 5/63
Example - simple servo system
There is a trade-off between Mp and tp

1 10

0.8 8

0.6 6

tp (sec)
Mp

0.4 4

0.2 2

0 0
−1 0 1 2
10 10 10 10
k (log-scale)
Overshoot Mp and time-to-peak tp against k
Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 6/63
Example - simple servo system

10

6
tp (sec)

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Mp
Trade-off curve showing Mp against tp
Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 7/63
Sensor Noise versus Disturbance Trade-off

• Most feedback control problems can be cast into architecture below

• Considers effects of reference input, plant disturbance & sensor noise

D(s)
disturbance

+
U (s✲
) ❄
R (s) ✲
+ ✲ K (s) G(s) ✲
+ ✲ Y (s)
reference output
✻
−
+
controller plant
❄
✛ M (s)
+ sensor
noise
measured
output

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 8/63
The fundamental conflict D (s )
+
R (s ) ✲
+ ✲ K (s) U (✲
s) + ❄
✲ ✲ Y (s )
G (s )
−
✻ +
❄
✛+ M (s )

• Error is difference between reference and output

E (T ) = R (T ) − Y (T )

• Define sensitivity function S (s) & closed-loop transfer function T (s)

1 G(s)K (s)
S (s) = T (s) =
1 + G(s)K (s) 1 + G(s)K (s)

giving
E (s) = S (s) [R (s) − D(s)] + T (s)M (s)

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 9/63
The fundamental conflict
• Taking magnitudes

|E (s)| = |S (s)| (|R (s)| + |D(s)|) + |T (s)||M (s)|

• To reduce error cause by R (s) and D(s) make |S (s)| small (in some sense)
• To reduce error cause by M (s) make |T (s)| small (in some sense)
• BUT
1 G(s)K (s)
S (s) = 1+G(s)K (s)
and T (s) = 1+G(s)K (s)
⇒ T (s) + S (s) = 1

• Hence
|T (j ω)| + |S (j ω)| ≥ 1
Im

T (s ) S (s )

Re
1
Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 10/63
Fundamental conflict — frequency response
• For design in frequency domain

|T (j ω)| + |S (j ω)| ≥ 1

• If |S (j ω)| is made nearly zero, |T (j ω)| becomes nearly unity (often greater
than unity)
• Conversely, if |T (j ω)| is nearly zero, |S (j ω)| must be at least nearly unity or
greater
• Unavoidable trade-off between attenuating plant disturbances, D(s), and
filtering out measurement error, M (s)
• Can also be shown that making |T (jw )| small means stability robustness &
small control effort

plant disturbance rejection & reference tracking

versus
measurement noise filtering & stability robustness & control effort

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 11/63
Fundamental conflict example - simple servo system

2.5

|S(j ω)|
|T(j ω)|
2
|S(j ω)|+|T(j ω)|
magnitude, | · |

1.5

0.5

0
10-2 10-1 100 101 102
frequency, ω (rad/s)
Sensitivities for Servo System (with 10% overshoot)
Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 12/63
Another Control Design Trade-off — Waterbed effects

log |S (j ω)|
Bode’s Sensitivity Integral states ✻ pop up
that the average of the logarithm
✻ ✻
of the sensitivity is conserved. ✍
0 ✲ω
If the sensitivity function is
decreased at low frequencies, it
R
must be traded-off for a larger ❘ log |S (j ω)|d ω ≥ 0
sensitivity function at higher
frequencies ◆
❄ push down

Multi-objective optimization for control design — Conflicts and Trade-offs in Control Systems 13/63
Multi-objective Optimization

• Express design aims quantitatively as a set of n design objective functions

{φi (p) : i = 1 . . . n}, where p denotes the design parameter vector chosen by
the designer
• Formulate design problem as a multi-objective optimization problem

Problem

min {φi (p), for i = 1 . . . n}

p∈P

where P denotes the set of possible design parameters p

Multi-objective optimization for control design — Multi-objective Optimization 14/63

Pareto-optimal solutions

• In most cases, objective functions, φi , are in conflict, so the reduction of one

objective function leads to increase in another
• Result of multi-objective optimization known as a Pareto-optimal solution

• Pareto-optimal solution has

φ2 (p) ✻
property that it is not possible
to reduce any one φi without
Attainable set
increasing at least one other φi
• Point lying in the interior of the
attainable set is sub-optimal,
since both φ1 and φ2 can be
both reduced Pareto ✻
optimal set
• A solution to a multi-objective ✲
optimization problem is hence φ1 (p)
generally not unique

Multi-objective optimization for control design — Multi-objective Optimization 15/63

Multi-Objective Design Approaches

• Many approaches suggested in several disciplines

• Non-linear programming — many approaches proposed, in particular, work of
Polak, Mayne and coauthors
• Interactive multi-objective programming
• Convex optimization
• Some early software
• DELIGHT — many of Polak and Mayne approaches
• ANDECS — whole CACSD environment
• QDES —- Boyd & Barratt approach to convex optimization
• MOPS — Multi-Objective Parameter Synthesis — from DLR
• MODCONS — mixed optimization — MATLAB Toolbox
• MATLAB Optimization Toolbox (Goal Attainment method)
• CONVEX CONTROL DESIGN TOOLBOX — MATLAB Toolbox from ONERA
• Q-Synthese — convex optimization MATLAB Toolbox from Universität Kassel
• Many modern packages have been developed for general problems

Multi-objective optimization for control design — Multi-objective Optimization 16/63

Weighted Sum Gap
To solve problem, it is common to convert to a single objective optimization
problem by use of a weighted sum objective function & use non-linear
programming Pareto φ1
optimal
n ✻ curve
Φ(p) =
X
wi φi (p)
✾
i =1
weighted sum
solutions
• If some φi are not convex, then ✙
parts of the Pareto optimal set may
not be found for any wi
❃
• Convex problems pose few duality gap
difficulties in converting to
single-objective problems
✌

❄
✲
φ2

Multi-objective optimization for control design — Multi-objective Optimization 17/63

Minimax Problem

Also common to use a max objective function

Φ(p) = max wi φi (p)

i

• Objective function is not everywhere differentiable — can cause problems

• Many other formulations have been proposed — some maintain the genuine
multi-objective nature of problem

Multi-objective optimization for control design — Multi-objective Optimization 18/63

Interactive Multi-objective Programming (IMOPS)

• Generally, the multi-objective design process is interactive, with computer

providing information to designer about conflicting design requirements, and
designer adjusting problem specification to explore the various possible
solutions to problem
• The design process is thus a two way process, with
1. computer providing information to designer about conflicting design requirements
2. and designer making decisions about ‘trade-offs’ between design requirements
based on this information as well as on the designer’s knowledge, experience
and intuition about the particular problem
• Designer supported by various graphical displays which provide information
about progress of the search algorithm and about conflicting design
requirements

Multi-objective optimization for control design — Multi-objective Optimization 19/63

Convex optimization

• Many multi-objective problems can be posed as convex optimization

problems by using a Youla parameterization (Boyd & Barratt)
• Efficient methods have been developed to solve such problems
• Some multi-objective problems can be posed as linear matrix inequalities
(LMI’s) — some others as linear programming problems (Elia & Dahleh)

Multi-objective optimization for control design — Multi-objective Optimization 20/63

Method of Inequalities (MOI)
Problem is expressed as a set of algebraic inequalities which need to be satisfied
for a successful design

Problem
Find p such that inequalities

φi (p) ≤ εi for i = 1 . . . n

are satisfied, where εi are design goals chosen by the designer and represent the
largest tolerable values of the objective functions φi

• The aim of the design is to find a p that simultaneously satisfies the set of
inequalities
• Generally used interactively in an IMOPS environment
• Designer iteratively tightens and relaxes εi with the aid of graphical tools
• In the limit, a Pareto-optimal solution is obtained — but in practise,
near-optimal solutions work better (see John Doyle’s work on Highly
Optimized Systems and Robust Yet Fragile systems, for example)
Multi-objective optimization for control design — Method of Inequalities (MOI) 21/63
MOI — Admissible Set

φ1 Pareto
optimal
✻ curve
✾

ε1
Any solution to the problem lies in the ②
admissible set (the set of all solutions):
admissible
set
{p : φi (p) ≤ εi , i = 1 . . . n}

✲
ε2
φ2

Multi-objective optimization for control design — Method of Inequalities (MOI) 22/63

MOI — IMOPS Interface

Multi-objective optimization for control design — Method of Inequalities (MOI) 23/63

Multi-Objective Genetic Algorithms (MOGA)

• Design philosophy of MOGA differs φ1 Pareto

from MOI, in that a set of optimal
✻ curve
simultaneous solutions is sought, ✾
and designer then selects best × ×
solution from the set × ×
• Idea is to develop a population of
Pareto-optimal or near
×
Pareto-optimal solutions ×

× dominated solution
non-dominated solution
✲
φ2

Multi-objective optimization for control design — Multi-Objective Genetic Algorithms (MOGA) 24/63
Multi-Objective Genetic Algorithms (MOGA)

• Design philosophy of MOGA differs φ1 Pareto

from MOI, in that a set of optimal
✻ curve
simultaneous solutions is sought, ✾
and designer then selects best × ×
solution from the set × ×
• Idea is to develop a population of ε1
Pareto-optimal or near
×
Pareto-optimal solutions ×
• However, to restrict the size of the
near Pareto-optimal set design
goals are also incorporated in a solution
similar way to the MOI × dominated solution
non-dominated solution
• This formulation maintains the ✲
genuine multi-objective nature of ε2
φ2
the problem

Multi-objective optimization for control design — Multi-Objective Genetic Algorithms (MOGA) 25/63
MOGA Ranking

φ1
• The MOGA is set into a
multi-objective context by means of a

the fitness function b

• Individuals are ranked on the basis

of the number of other individuals c
d
they are dominated by for the ε1
e
unsatisfied inequalities f
g
• GAs are naturally parallel and
hence lend themselves well to h
multi-objective settings i

• Also work well on non-smooth j

objective functions – it is very easy

to extend GA’s to solve mixed
ε2 φ2
continuous/integer problems

Multi-objective optimization for control design — Multi-Objective Genetic Algorithms (MOGA) 26/63
Trade-off Diagram

Cost

1 2 3 4 5 6 7
Objective

Multi-objective optimization for control design — Multi-Objective Genetic Algorithms (MOGA) 27/63
Example Solution

140

120
Total number of bits for implementation

100

80

60

2
40
1

20

0
−5 −4 −3 −2 −1 0
10 10 10 10 10 10
||R−R || (log scale)
q ∞

Multi-objective optimization for control design — Multi-Objective Genetic Algorithms (MOGA) 28/63
Mixed Optimization

Analytical optimization techniques Parameter optimization based methods

(e.g. H∞ , LQG) generally (based on hill-climbing, GA’s etc)
1. have non-explicit closed-loop 1. have explicit closed-loop performance,
performance, 2. are often multi-objective,
2. are single-objective, 3. are not implicitly robustly stable,
3. are robustly stable, 4. provide simple controllers,
4. provide high-order controllers, 5. are flexible,
5. are not very flexible, 6. are often non-convex resulting in local
6. provide a global optimum, minima,
7. can deal with relatively large 7. can deal with small problems only,
multivariable problems; 8. may have difficulty stabilizing the
system.

Multi-objective optimization for control design — Mixed Optimization 29/63

Mixed Optimization

• A combination of analytical optimization and parameter search methods may

be able to overcome some of the limitations of using just one approach.
• The basic idea is simple, and that is to use search methods to automate the
iterative procedure of designing the weighting functions and other design
parameters to achieve the performance specifications for the system.

Multi-objective optimization for control design — Mixed Optimization 30/63

Mixed Optimization – Overview

• Design parameters in analytical optimization are often weighting functions —

e.g. LQR cost function:
Z ∞
(x ′ (t )Qx (t ) + u ′ (t )Ru (t )) dt
0

design parameters are Q and R

• Parameterize weighting functions by p and use MOI or MOGA to design p in
combination with optimal controller synthesis in mixed optimization approach
• e.g. for nominal plant G(s) augmented by a set of nw weighting functions
W (s) = (W1 (s), W2 (s) . . . Wnw ) — controller Kmin (s, G, W ) which is optimal
in some sense can be synthesized — set of closed-loop performance
functions φ of optimal control system can be computed
• Parameterize weighting functions by the design vector p formulate as for MOI
or MOGA and use search algorithm to design p such that set of performance
criteria are satisfied

Multi-objective optimization for control design — Mixed Optimization 31/63

Mixed Optimization - Early Examples
• J.S. Baras, M.K.H. Fan, W.T. Nye, and A.L. Tits. DELIGHT.LQG: A CAD system for control
systems design using LQG controller structure. In 18th Annual Conf. on Inf. Sci and Syst.,
Princeton, NJ, 1984.
• J.F. Whidborne, I. Postlethwaite, and D.-W. Gu. Robust controller design using H∞ loop-shaping
and the method of inequalities. Technical Report 92-33, Leicester University Engg Dept,
Leicester, U.K., 1992.
• W. Tych. Multi-objective optimisation technique for mapping the technical design objectives into
the values of the weighting matrices in the linear quadratic regulator design problem. Technical
Report TR-117, Univ. Lancaster Centre for Research on Environmental Systems and Statistics,
Lancaster, UK, 1994.
• J.E. Paddison, R.M. Goodall, J. Bals, and G. Grübel. Multi-objective design study for a maglev
suspension controller using the databased ANDECS-MATLAB environment. In Proc. IEEE/IFAC
Symp. on Comp. Aided Contr. Syst. Design (CACSD’94), pages 239–246, Tuscon, USA, March
1994.
• D. Haessig. Selection of LQR/LTR weighting matrices through constrained optimisation. In Proc.
1995 Amer. Contr. Conf., pages 458–460, Seattle, USA, 1995.
• N.V. Dakev, J.F. Whidborne, and A.J. Chipperfield. H∞ design of an EMS control system for a
maglev vehicle using evolutionary algorithms. In Proc. GALESIA 95, pages 226–231, Sheffield
U.K., September 1995.
• J. King and B.A. White. Robust controller design using H∞ loop shaping with a genetic algorithm
multiobjective optimization. In Proc. 13th IFAC World Congress, page , San Fransisco, CA, June
1996.
Multi-objective optimization for control design — Mixed Optimization 32/63
Mixed Optimization — Conclusions

• Resulting controllers are moderately high order — controller order reduction

stage could be introduced
• Problem is non-linear — local minima only found
• Hill-climbing methods require accurate calculation of indices
• GA’s inefficient but indices can be approximate (problem can be non-smooth)
e.g. mixed sensitivity H∞ -optimal

Multi-objective optimization for control design — Mixed Optimization 33/63

Convex Optimization
A function φ(p) is convex if for all p1 , p2 ∈ P
φ(λ1 p1 + λ2 p2 ) ≤ λ1 φ(p1 ) + λ2 φ(p2 ), λ1 + λ2 = 1,
φ
• Convex function
✻
optimization problems can
be solved relatively λ1 φ(p1 )
efficiently +λ2 φ(p2 )

• Norms of convex functions

are also convex
• Hence multiobjective
problems that are convex φ(λ1 p1
can be solved efficiently (by +λ2 p2 )
weighted sums, minimax,
weighted squares etc)

✲
p1 p2 p

Multi-objective optimization for control design — Convex Multi-objective Optimization 34/63

Convex Problems

Many performance/robustness indices are convex

1.4

e.g. 1.2
• H∞ -norms,
H2 -norms, 1

L1 -norms, . . .
0.8
• Time domain
y(t)

specifications 0.6 y 1 (t)

envelope y 2 (t)
0.4
• Frequency domain (y1 (t)+y 2 (t))/2
specifications 0.2 bounds
envelope
0
0 5 10 15
t

Multi-objective optimization for control design — Convex Multi-objective Optimization 35/63

Q-parametrization

• Standard closed loop control

arrangement w ✲ P11 P12 ✲ z
z = P11 w + P12 u
e = P21 w + P22 u u ✲ P21 P22 y

• Closed loop transfer function:

Hzw = P11 + P12 K (I − P22 K )−1 P21 K ✛

• Not convex unless P22 = 0

• Convexify by Q-parametrization

Multi-objective optimization for control design — Convex Multi-objective Optimization 36/63

Q-parametrization

• stabilizing nominal controller Knom is T̃ (s)

added
w ✲ ✲z
• Controller F (s) is given as
P (s)
✲
F (s) = LFTl (Knom , Q̃ )
u y
• But to make the problem convex, ✛
transfer function Hev from v to ẽ Knom (s)
✲
must be zero
• Plant and nominal controller are v ẽ
combined into a new transfer matrix
  Q̃ (s) ✛
T T2
T̃ = LFTl (P , Knom ) =: 1
T3 T4

Multi-objective optimization for control design — Convex Multi-objective Optimization 37/63

Convex Q-parametrization
T (s)

w ✲ ✲z
✲ T̃ (s)
• Replace lower dotted block with
ẽ ❄
Q = (I − Q̃T4 )−1 Q̃ = Q̃ (I − T4 Q̃ )−1
✲ −T4 (s) ✲ Σ
• Final structure is an LFT,
Hzw = LFTl (T , Q ), and because T4 v e
is zero, the closed loop transfer
matrix is T4 (s) ✲ Σ ✛

Hzw = T1 + T2 QT3

and is convex in Q Q̃ (s) ✛

Q (s)

Multi-objective optimization for control design — Convex Multi-objective Optimization 38/63

Observer structure Q-parametrization

T (s)

w ✲ ✲z
✲ T̃ (s)

ẽ ❄
✲ −T4 (s) ✲ Σ

v e

T4 (s) ✲ Σ ✛

Q̃ (s) ✛

Q (s)

Multi-objective optimization for control design — Convex Multi-objective Optimization 39/63

Convex Control Problem

Problem

min kΦk
Q

where Φ = [φ1 , φ2 , . . . , φn ]

All that remains is to parameterize Q (s) so that the convexity is maintained. For
example
∞
X
Q (s, p) = pi ψi (s)
1

can characterize all stable Q (s) (e.g. Ritz-Galerkin approximation) where {ψi (s)}
provides an orthogonal basis. A finite truncation of the sequence can be used, i.e.
m
X
Q (s, p) = pi ψi (s)
1

Multi-objective optimization for control design — Convex Multi-objective Optimization 40/63

Convex Optimization — Conclusions

• Resulting controllers are very high order (in theory infinite dimensional) and
often not practical
• Performance indices φi must be calculated to very high accuracy — time
consuming
• Method provides “Limitations of Performance”

Multi-objective optimization for control design — Convex Multi-objective Optimization 41/63

Some References

• Boyd, S. & Barratt, C. Linear Controller Design: Limits of Performance,

Prentice-Hall, 1991 (convex MO optimization for control systems – can
download from Boyd’s website)
• Stephen Boyd and Lieven Vandenberghe Convex Optimization Cambridge
University Press 2004 (can download from Boyd’s website)
• A. Linnemann. Convergent Ritz approximations of the set of stabilizing
controllers. Syst. Control Lett., 36:151–156, 1999.
• G. Ferreres and G. Puyou. Flight control law design for a flexible aircraft:
Limits of performance. J. Guid. Control Dyn., 29(4):870–870, 2006.
• C. Scherer, P. Gahinet, and M. Chilali. Multiobjective output-feedback control
via LMI optimization. IEEE Trans. Autom. Control, 42(7):896–911, 1997.

Multi-objective optimization for control design — Convex Multi-objective Optimization 42/63

Example — Maglev Suspension Controller Design

• DC electro-magnetic
suspension uses attractive
forces of electro-magnets
acting upwards to levitate
vehicle towards steel
guideway
• High speed frictionless
transport — up to 500
km/hour
• Several operating Maglevs
including
Shanghai-Pudong, Aichi
Linimo and Incheon Airport

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 43/63
Control aims

passenger cabin
• EMS inherently unstable —
needs active control secondary
suspension
• must maintain airgap
chassis
between vehicle and
guideway guidance guideway
magnet
• ensure quality of ride
• avoid actuator saturation
levitation
magnet

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 44/63
Maglev Model

passenger
m2 ❄
x2 • dc electromagnet force:
cabin
 2
K i (t )
F (i , z , t ) =
guideway 2 z (t )
c k
z ✻ ❄
h i is current
❄ F
• Control voltage:
m1 ❄
x1
chassis  
d i (t )
v (t ) = Ri (t ) + K .
dt z (t )
mg
❄
R is total resistance
• Secondary suspension consisting of airsprings and hydraulic shock
absorbers — assumed linear
• h(t ) is disturbance resulting from variations in guideway profile

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 45/63
Performance for Maglev Suspension

Major disturbance from variations in guideway height:

 
D = sup ḣ(t ) : t ≥ 0 = 30 mm/s
 
For all possible h(t ) such that sup{ḣ(t ) : t ≥ 0} ≤ D
Z ∞
sup {|yi (t , h)| : t ≥ 0} = D |yi (τ, 1)| d τ
0

yi (τ, 1) is unit step response of ith output of linearised closed-loop system

Nominal performance functions are defined on the airgap, passenger acceleration
and control voltage

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 46/63
Performance for Maglev Suspension

For non-linear system performance, responses of airgap, passenger acceleration

and control voltage to test input, htest (t ), are calculated, and maximum absolute
values determined

200
Guideway disturbance (mm)

150

100

50

−50
0 2 4 6 8 10 12
Time (sec)
Test input htest

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 47/63
Performance for Maglev Suspension

A maximum power spectral density Φmax (ω) of passenger cabin acceleration has
been recommended by US Department of Transportation as minimum ride quality
standard — performance functional is defined based on Φmax (ω)
5
10

• The p.s.d. of track variations:

4
10

Φhh (ω) = Av /ω 2

Acceleration psd
3
10

A depends on track quality, v is

speed 2
10

• p.s.d. of passenger cabin: 1

10

2 Av
Φẍ2 ẍ2 (ω) = |Tẍ2 h (ω)| ω2 . 0
10 0 1 2
10 10 10
Frequency (rad/sec)

Max permitted power spectral density

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 48/63
The Design

• Air gap measurement z, and passenger cabin acceleration, ẍ2 , used for
feedback
• Weighting function configurations :

(s2 +p s+p )
W1 = p1 (s2 +p42 s+p53 )
 
(s+p )
W2 = diag p6 (s+p78 ) , p9

• Secondary suspension stiffness and damping factors c and d are included as

design parameters p10 and p11

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 49/63
Objective functions

Seven performance objective functions Performance goals set:

defined:
Z ∞
εγ = 5
φ1 = D |z (τ, 1)|d τ ε1 = ε4 = 5 mm
Z0 ∞ ε2 = ε5 = 500 mm/s2
φ2 = D |ẍ2 (τ, 1)|d τ ε3 = ε6 = 600 V
Z0 ∞ ε7 = 0
φ3 = D |v (τ, 1)|d τ
0
φ4 = max{|z ′ (t , htest )|}
t

φ5 = max{|ẍ2′ (t , htest )|}

t

φ6 = max{|v ′ (t , htest )|}

t

φ7 = max{Φẍ2 ẍ2 (ω) − Φmax (ω)}

ω

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 50/63
Problem

✲ W1 u✲ G y✲
W2 ✲
Problem for mixed optimization design of
Gs
EMS control system is to find p satisfying
the inequalities:
✲ W1 u✲ G y✲
W2
γ0 (p) ≤ ǫγ
Gs
and Ks ✛

φi (p) ≤ εi , i = 1, 2, . . . , 7 u ✲ G y

where γ0 is optimal H∞ -norm used by

McFarlane & Glover’s LSDP W1 ✛ Ks ✛ W2 ✛
K

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 51/63
Solution

• Multi-objective simulated annealing was used to solve problem

• Design vector found which met all design goals except φ1 and φ7 , which were
only marginally exceeded
• Final performance objective functions:

γ0 = 2.64 φ1 = 5.06 mm
2
φ2 = 391.4 mm/s φ3 = 201.9 V
φ4 = 4.38 mm φ5 = 291.1 mm/s2
φ6 = 33.6 V φ7 = 0.075

• Designed weighting functions:

 
(s2 +169.6s+485.4) (s+256.2)
W1 = 903.2 (s2 +69.5s+488.3) W2 = diag 426.9 (s+379.2) , 3.23

Secondary suspension stiffness c = 90.3

Secondary suspension damping factors d = 20.0

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 52/63
Nonlinear system responses to test input htest
Test input
200
Guideway disturbance (mm)
150

100

50

−50
0 2 4 6 8 10 12
Time (sec)
Airgap response to test input
4

2
Airgap (mm)

−2

−4

−6
0 2 4 6 8 10 12
Time (sec)
Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 53/63
Nonlinear system responses to test input htest

Passenger acceleration (mm/sec/sec)

Passenger acceleration response to test input
200

100

−100

−200

−300
0 2 4 6 8 10 12
Time (sec)
Control voltage response to test input
60
Control voltage (volts)

40

20

−20
0 2 4 6 8 10 12
Time (sec)
Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 54/63
Power Spectral Densities
5
10

4
10
Acceleration psd

3
10

2
10

1
10

0
10 0 1 2
10 10 10
Frequency (rad/sec)

Passenger cabin psd (—–) and ride quality standard psd (· · · )

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 55/63
Conclusions

Analytical optimization combined with MOI (mixed optimization):

• Explicit closed loop performance
• Multi-objective
• Robustly stable
• Flexible / interactive / trade-offs easily made
• High order controllers

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 56/63
Example References

• Dakev, N.V., J.F. Whidborne, A.J. Chipperfield & P.J. Fleming. 1997. “H∞
design of an EMS control system for a maglev vehicle using evolutionary
algorithms.” Proc. IMechE, Part I: J. Syst. & Contr. 311(4):345–355.
• J.F. Whidborne and S. Arunsawatwong. Design of a critical control system
using simulated annealing and mixed optimization. In 2nd Asian Control
Conf., pages I288–I289, Seoul, S. Korea, July 1997.

Multi-objective optimization for control design — Design Example — Maglev Suspension Controller 57/63
Multi-objective optimization – References
• R.G. Becker, A.J. Heunis, and D.Q. Mayne. Computer-aided design of control systems via
optimization. IEE Proc.-D, 126(6):573–578, 1979.
• S. Boyd, C. Barratt, and S. Norman. Linear controller design: Limits of performance via convex
optimization. Proc. IEEE, 78(3):529–574, 1990.
• G.P. Liu, J.B. Yang, and J.F. Whidborne. Multiobjective Optimisation and Control. Research
Studies Press, Baldock, UK, 2002.
• Fonseca, C.M. & P.J. Fleming. 1995. “Multiobjective optimization and multiple constraint
handling with evolutionary algorithms – part I: a unified formulation.” IEEE Trans. Syst. Man &
Cybernetics – A 28(1):26–37.
• F.W. Gembicki and Y.Y. Haimes. Approach to performance and sensitivity multiobjective
optimization: the goal attainment method. IEEE Trans. Autom. Control, 20(8):821–830, 1975.
• D.P. Giesy. Calculation of Pareto-optimal solutions to multiple-objective problems using
threshold-of-acceptability constraints. IEEE Trans. Autom. Control, 23(6):1114–1115, 1978.
• G. Grübel, H.-D. Joos, M. Otter, and R. Finsterwalder. The ANDECS design environment for
control engineering. Proc. 12th IFAC World Congress, volume 6, pages 447–454, Sydney, 1993.
• J.G. Lin. Mutiple-objective problems: Pareto-optimal soultions by method of proper equality
constraints. IEEE Trans. Autom. Control, 21(5):641–650, 1976.
• A. Gambier & M. Jipp. Multi-objective Optimal Control: An introduction. Proc. 8th Asian Control
Conference (ASCC), Kaohsiung, Taiwan, 2011.
• K.F. Man, K.S. Tang, S. Kwong & W.A. Halang. 1997. Genetic Algorithms for Control and Signal
Processing. Advances in Industrial Control London, U.K.: Springer.
Multi-objective optimization for control design — References 58/63
Further references — theory

• A. Molina-Cristobal, I.A. Griffin, P.J. Fleming, and D.H. Owens. Linear matrix inequalities and
evolutionary optimisation in multiobjective control. Int. J. Systems Sci., 37(8):513 – 522, 2006.
• W.Y. Ng. Interactive Multi-Objective Programming as a Framework for Computer-Aided Control
System Design, volume 132 of Lect. Notes Control & Inf. Sci. Springer-Verlag, Berlin, 1989.
• I. Postlethwaite, J.F. Whidborne, G. Murad, and D.-W. Gu. Robust control of the benchmark
problem using H∞ methods and numerical optimization techniques. Automatica, 30(4):615–619,
1994.
• C.-C. Sun, H.-Y. Chung, and W.-J. Chang. H2 /H∞ robust static output feedback control design
via mixed genetic algorithm and linear matrix inequalities. ASME J. Dyn. Syst. Meas. & Control,
127(4):715–722, 2005.
• J.F. Whidborne, D.-W. Gu, and I. Postlethwaite. Simulated annealing for multi-objective control
system design. IEE Proc. Control Theory and Appl., 144(6):582–588, 1996.
• J.F. Whidborne, I. Postlethwaite, and D.-W. Gu. Robust controller design using H∞ loop-shaping
and the method of inequalities. IEEE Trans. Control Syst. Technology, 2(4):455–461, 1994.
• L.A. Zadeh. Optimality and non-scalar-valued performance criteria. IEEE Trans. Autom. Control,
8(1):59–60, 1963.
• V. Zakian and U. Al-Naib. Design of dynamical and control systems by the method of
inequalities. Proc. IEE, 120(11):1421–1427, 1973.

Multi-objective optimization for control design — References 59/63

Applications references

• O. Cifdaloz, M. Shayeb, R. Metzger, Y.-L. Yi, and A. Rodriguez. MIMO control system design for
aircraft via convex optimization. In Proc. 2003 Amer. Contr. Conf., pages 987–992, Denver, CO,
June 2003.
• B. Clement, G. Duc, and S. Mauffrey. Aerospace launch vehicle control: a gain scheduling
approach. Control Engineering Practice, 13(3):333–347, March 2005.
• T.R. Crossley and A.M.S. Dahshan. Design of a longitudinal ride-control systems by Zakian’s
method of inequalities. J. Aircraft, 19(9):730–738, 1982.
• G. Ferreres and G. Puyou. Flight control law design for a flexible aircraft: Limits of performance.
J. Guid. Control Dyn., 29(4):870–870, 2006.
• K. Fu and J.K. Mills. Integrated design of a quarter-car semi-active suspension system using a
convex integrated design method. Int. J. Vehicle Design, 42:328–347, 2006.
• E.H. Fung, Y. Wong, H.H.T. Liu, and Y. Li. Design of longitudinal system for a nonlinear F-16
fighter using MSS method. In Proc. 16th IFAC World Congress, Prague, July 2005.
• E.J. Hughes, A. Tsourdos, and B.A. White. Multiobjective design of a fuzzy controller for a
nonlinear missile autopilot. In Proc. IEEE Symp. on Comp. Aided Contr. Syst. Design (CACSD
2002), pages 15–20, Glasgow, UK, September 2002.

Multi-objective optimization for control design — References 60/63

Applications references

• M. Jeanneau, J. Lamolie, G. Puyou, and N. Aversa. AWIATOR’s design of multi-objectives

control laws. In Proc. 16th IFAC World Congress, Prague, July 2005.
• H.-D. Joos and R. Finsterwalder. Multiobjective design assessment and control law synthesis
tuning for flight control development. In Proc. IEEE Symp. on Comp. Aided Contr. Syst. Design
(CACSD’99), pages 433–438, Kohala Coast, HI, August 1999.
• G. Kreisselmeier and R. Steinhauser. Application of vector performance optimization to a robust
control loop design for a fighter aircraft. Int. J. Control, 37(2):251–284, 1983.
• G.P. Liu, R. Dixon, and S. Daley. Multi-objective optimal-tuning proportional-integral controller
design for the ALSTOM gasifier problem. Proc. Inst. Mech. Eng. Part I-J Syst Control Eng.,
214(16):395–404, 2000.
• H.H.T. Liu. Multiobjective design for an integrated flight control system: a combination with model
reduction approach. In Proc. IEEE Symp. on Comp. Aided Contr. Syst. Design (CACSD 2002),
pages 21–26, Glasgow, UK, September 2002.
• G. Puyou, G. Ferreres, C. Chiappa, and P. Menard. Multiobjective method for flight control law
design. In Proc. AIAA Guidance, Navigation, and Control Conference, 2004.

Multi-objective optimization for control design — References 61/63

Applications references

• D. Tabak, A.A. Schy, D.P. Giesy, and K.G. Johnson. Application of multiobjective optimization in
aircraft control systems design. Automatica, 15:595–600, 1979.
• O. Taiwo. Design of multivariable controller for a high-order turbofan engine model by Zakian’s
method of inequalities. IEEE Trans. Autom. Control, AC-23(5):926–928, 1978.
• O. Taiwo. Design of multivariable controllers for an advanced turbofan engine by Zakian’s
method of inequalities. ASME J. Dyn. Syst. Meas. & Control, 101:299–307, 1979.
• O. Voinot, D. Alazard, P. Apkarian, S. Mauffrey, and B. Clement. Launcher attitude control:
discrete-time robust design and gain-scheduling. Control Engineering Practice,
11(11):1243–1252, November 2003.

Multi-objective optimization for control design — References 62/63

Summary

• Considered the problem of trade-off in control design

• Leads to design being formulated as a multi-objective optimization problems
• Problem can be tackled using several methods including MOI, MOGA,
Convex Optimization & Mixed Optimization
• Example of Mixed Optimization given

Slides will be available from publi . ranfield.a .uk/eh3081/

Multi-objective optimization for control design — References 63/63