Professional Documents
Culture Documents
Introduction
In the past decades, H optimal control theories have been developed very quickly and widely applied to treat
control system problems, such as min max model matching, [3],[10],[11], min-max disturbance rejection [2],[5],[9], and
robust optimization [2],[3],[5],[11]. In general, the H optimal output feedback control problem can be solved by two
methods. One is based on Youlas parameterization with Nehari theorem or so called A-A-K technique from the
frequency domain viewpoint [2],[3]. The other is based on observer based state feedback with game theory to solve the
Riccati - like equations from time domain [5]. However, based on the above approaches the order of the derived H
optimal controller is much higher than that the plant. Hence it is not easy to implement the H optimal controller in
practical applications.
On the other hand Genetic Algorithms (GAs) are searching algorithms based on simulation of the human trial -and error procedure using the survival of the fittest Darwinian principle [1] and they have been successfully applied in
many optimization problems [4]. Further more multi objective GAs (MOGAs) have been proven as a very effective
tool in searching poorly understood and complex space for optimization in engineering applications [14],[19]. MOGAs
incorporate the concept of Paretos domination to evolve a family of non-dominated solutions along the pareto optimal
frontier.
This paper introduces a blended target function which incorporates the pareto rank [17] into a specific function and
then with the use of genetic evolution tries to minimize it towards to feasible solutions.
At first criteria, based on plant stability and plant maximum singular values are evaluated, then ranking based
coefficient is produced using those criteria, and finally a new target function having the form of Kreisselmeier [8] is
created with the use of the rank of each individual and the relative criteria values.
The above mention procedure is applied to design a controller for an ill-conditioned plant. By ill-conditioned, we
mean that the plant gain is strongly depended on the input direction, or equivalently that the plant has a high condition
number.
(1)
J ( p ) = [ J1 ( p ),..., J m ( p )]
The performance specification is satisfied if J(p) < c for some threshold vector c.
The distinctive feature of Kreisselmeier [8] approach is the way in which they replace the multi-objective problem
min { Ji ( p ) ci } , i = 1,...., m by the min -max problem
p
J i ( p)
ci
and finally approximate this by a smooth, unconstrained problem:
min a ( p) = max
p
min a( p ) =
p
J ( p )
m
ln exp i
>0
i =1
ci
(2)
(3)
3. The -framework.
This section gives a very brief introduction to -framework and defines some of nomenclature used in the rest of the
paper. For further reading consult [6],[7]. The H norm of a transfer function matrix M(s) is the peak value of the
maximum singular value over all frequencies.
(4)
M ( s) sup (M ( j ))
The left block diagram in Fig. 1(a) shows the general problem formulation for controller synthesis. It consists of an
augmented plant P with three pairs of input /output variables including the performance and stability requirements on
the nominal to be controlled system G. The first pair consists of the measured outputs y(t) and control inputs u(t). The
second pair consists of performance variables e(t) and external input signals d(t) and the third pair consists of output
signals z(t) and w(t) through which unit norm perturbations are fed back into system. A controller K and a (blockdiagonal) perturbation matrix =diag{1,,n} representing the uncertainty that includes all possible perturbations of
the system and it is usually normalized in such a way that 1 .
By grouping P and K we get N obtained from the lower LFT [6] with
(5)
To evaluate the perturbed (uncertain) transfer function from external inputs d to performance variable e we use to
close the upper loop N Fig 1 (b) resulting the upper LFT
N 22 + N 21 ( I N11 )1 12
F = Fu ( N , )
(6)
with % = diag{, P } . is computed frequency by frequency through upper and lower bounds. Here we only consider
the upper bound % ( F ( P, K ) ) min ( DF ( P, K )( j ) D 1 ) where D is a scale for the plant with D = {D | D% = % D}
D D
At present there is no direct method to synthesize an -optimal controller, however -synthesis which combines analysis and H -synthesis often yields good results. The idea is to attempt to solve
min inf DFl ( P, K )( j ) D 1
(8)
K
DD
by alternating between minimizing for either K or D while holding the other fixed. In this work we keep the D(s) fixed
at a given, stable, minimum phase, real rational function and then we solve the optimization problem
min Fl ( PD , K )( j ) which is an H synthesis.(PD is the D- scaled plant).
K
stabilizing
Coding Scheme
The structured - specified controller has the following form:
N ( S ) m s m + m 1s m 1 + L + 0
(9)
C (s) = C
=
DC ( S )
s n + n 1s n 1 + L + 0
The controller is assigned with some desired order m and n to minimize the performance criterion (8) and the solution
will be found using genetic algorithms. Thus for integrating the GAs with the H optimal control problem first we
define the structure - specified controller parameter vector , which corresponds to (9) and has the following form:
= [ m , m 1 ,K , 1, 0 , n 1 , n 2 ,K , 1 , 0 ]
(10)
Further more for each i we specify an upper and lower bound and thus the controllers parameter space can be
defined as:
= { i | iL i iU , i = 1,.., n}
(11)
iU , iL are the upper and lower bounds of each i , where i denotes the ith element of
Thus for solving the problem and finding a feasible solution, in parameter vector, which minimizes (8) using GAs, , is
coded to a binary chromosome. In binary coding method every element i of the parameter vector is coded as a
string of length l i , which consists of 0 and 1 for the desired resolution Ri . In general we have Ri =
iU iL
2l i 1
Pareto domination
A vector x is partially less than y, symbolically x < py when the following condition holds:
( x < py ) ( i )( xi yi ) ( i )( xi < yi )
Under these circumstances we say that point x dominates point y [4].
Multiobjective Rank
An individual xi , at generation t, which is dominated, by pi(t ) individuals in the current population receives the
following rank ( xi , t ) = 1 + pi( t ) [17]
rank ( xi , t ) m
Nind
GOV ( xi , t ) =
J k ( xi )
ln exp
Nind k =1
rank ( xi , t ) + 1
Nind is the number of individuals at the current population.
J k ( xi ) is the k-th objective and xi = i for our control optimization problem.
(12)
1.
2.
3.
4.
100
103 ,103
5.
if J1 ( i ) 0.0
(13)
else
6. To convert the GOV to the fitness value an exponential ranking [18] with appropriate selective pressure is applied
Optimization Procedure
The optimization procedure follows the next steps:
1.
2.
3.
4.
5.
6.
For a given plant P ( s ) specify the weighting function matrices, the controller structure C ( s ) , the genetic
parameters pc , pm , Nind and the desired resolution R .
Specify upper and lower bounds for the parameter vector (parameter space) and then calculate the required
chromosome length.
Generate randomly a population of binary strings
Calculate objective values for each chromosome.
Start Genetic Algorithm
Select parents by Stochastic Universal Sampling
Generate new chromosomes via double point crossover and mutation
Calculate the objective and fitness values of the new chromosomes
Reinsert the new chromosomes into the population via fitness-based reinsertion
Terminate genetic algorithm if maximum generations have reached.
Terminate genetic algorithm if maximum generations have reached
1.082 1.096
2 s
75s + 1
k2e
In physical terms this equivalent to a gain uncertainty of 20% and the delay of up to 1 min in each input channel. The
set of possible plants defined by (14) is denoted as and the following specifications must be achieved for every plant
G% .
Design Specifications
The following specifications S1 to S3 should be fulfilled for every plant G
S1 Closed loop stability
S2 For a unit step demand in channel 1 at t=0 the plant outputs y1 (tracking) and y2 (interaction) should satisfy:
0.99 y1 () 1.01
0.01 y2 () 0.01
Corresponding requirements hold for a unit step demand in channel 2.
S3 ( KS ) 316, . This specification is mainly added to avoid saturation of the plant inputs.
% ) < 1, 150. , which is indirectly satisfied with S3
S4 (GK
0.667 1.042
and the nominal model becomes
0.693
1 0.878 0.864 1
1 0.30171 -0.29183
G ( s) =
1.082 1.096 0.667 1.042 =
75 s + 1
Controller Structure
As a controller structure we will use a diagonal controller and then with Singular Value Decomposition it will transform
to a full order one. Hovd et al [12] found that the SVD-controller structure is optimal in a some cases e.g. for plants
consisting of symmetrically interconnected subsystems.
T
So the controller will be K ( s) = V0 K GA (s )U 0 where V0 , U0 are obtained from the singular value decomposition of
G0 = U 0 0V0T at a given frequency 0 and KGA is the diagonal controller which will be produced by the GA.
In this work we set 0=0 rad/sec and thus we have
T
-0.78198 0.67279
0
0.30171 -0.29183 0.6233
0.68745 -0.72623
G0 =
=
0.6233 0
0.023646 -0.72623 -0.68745
0.35097 -0.39221 1444
0.78198
424444
3 1444
424444
3 1444424444
3
U
2
s (b12 s + b11 s + b10 )
K GA =
0
A2
VT
(16)
with
(17)
= [a12 , a11 , a10 , b12 , b11 , b10 , a22 , a21 , a20 , b22 , b21 , b20 , A1 , A2 ]
as the parameter vector.
The parameter space region is = { i | 0 i 400 i = 1...14} and the resolution for each parameter will be 10-4 .
Controller synthesis
The controller may be synthesized using the GA-H synthesis with the following D-scales [13] for One Degree of
Freedom Controller and the control problem structure is pictured in fig 2
D(s ) = diag{d ( s ), d ( s), I 22 }
where
2 10 5 (s + 1000)( s + 0.25)(s + 0.054)
d ( s) =
( s + 067 + j 0.56)( s + 0.67 j 0.56)( s + 0.013)
and weights
1.1s + 0.2
10s+1
W (s ) =
I 2 2 ,We (s)=0.5
I 22
0.5s + 1
10s+10 -5
10
10
1/W P
1/W P
10
10
-1
-1
sv
10
sv
10
-2
-2
10
10
-3
-3
10
10
-4
10
-4
-3
10
-2
10
-1
10
10
Frequency
10
10
10
10
-3
10
-2
10
-1
10
10
Frequency
10
10
10
10
10
10
10
10
sv
sv
10
10
10
10
10
-1
10
-1
-3
10
-2
-1
10
10
(b)
10
Frequency
10
10
10
10
-3
10
-controller
-2
10
-1
10
10
Frequency
10
10
10
Figure 4 Maximum and minimum singular values of K S% for the two under comparison controllers
The specification S3 for both controllers is presented in figure 4. We can see from there that the GA produced controller
achieves the bound of 50db (316) and thus the plant input never saturates. The -controller violates the bound and
increases the amplitude in higher frequencies.
In fig 5 we can see the step response for the four different plants and we can see that both controller violate the
specifications presented above.
Opt. Controller
GA Controller
1
--..
-.
--
0.5
k1 k2
1.2 1.2
0.8 1.2
1.2 0.8
0.8 0.8
0.5
0
0
50
Time: min
100
0.5
0.5
0
0
50
Time: min
100
50
Time: min
100
50
Time: min
100
7. Conclusions
In this paper we have developed a method based on genetic algorithms to design structured specified H controller with
-sub optimal characteristics. The proposed scheme is found beneficial in designing low degree controllers and the
control of an ill-condition plant outlines the usefulness of the procedure.
The GOV function guides very well the the genetic search over large parameter spaces and finds feasible solutions very
soon. Finally it is easy to change the objectives and apply the scheme to 2 or other robust control problems.
References
[1] J.H.Holland, Adaptation in natural and artificial systems (The University of Michifan Press, Ann Ardor MI, 1975).
[2] M.Vidyasagar, Control System Synthesis : A factorization approach (Cambridge MA. MIT Press, 1985).
[3] B.A.Francis, A course in H Control Theory (Berlin Germany, Springer - Verlag, 1987).
[4] D.E.Goldberg, Genetic Algorithms in Search Optimization and Machine Learning (Addison Wesley, Pub Co. Inc, 1989).
[5] A.Stoorvogel, The H Control Problem: A State Space Approach (Prentice Hall, 1992).
[6] G.Balas, J.Doyle, K.Glover, A.Packard, and R.Smith, -Analysis and Synthesis Toolbox (The Mathworks Inc, 1998).
[7] S.Skogestad and I.Postlethwaite, Multivariable Feedback Control : Analysis and Design (John Willey and Sons, 2000)
[8] G.Kreisselmeier and et al, Application of vector performance optimization to a robust control loop design for a fighter
aircraft, Int.J.Control, (1983), 37, (2), 251-284.
[9] B.S Chen, Controller synthesis of optimal sensitivity : Multivariable Case, Proc. Inst. Elect. Eng, (1984), 131, 47-51.
[10] Kwakernaak H., Minimax frequency domain performance and robustness optimization of linear feedback systems, IEEE
Trans.Aut.Cont., (1985), 10, (AC-30), 994-1004.
[11] B.S Chen, Robust controller for multivariable model matching system, Int.J.Control, (1989), 50, 1717-1730.
[12] M.Hovd, R.D.Braatz, and S.Skogestad, SVD controllers for H2 - H and m-optimal, Automatica, (1997), 33, 433-439.
[13] P.Lundstrom, S.Skogestad, and J.Doyle, Two Degree of Freedom Controller Design for an Ill-Conditioned Distillation
Process using -Synthesis, IEEE Trans.Control Sys. Tech, (1999), 7, (1), 12-21.
[14] S Obayashi, D.Sasak, Y Takeguchi, and N.Hirose, Multiobjective Evolutionary Computation for supersonic wing -shape
optimization, IEEE Trans.Evol.Comp.,(2000), 4, (2), 182-188.
[15] D.J.N.Limmebeer, The specification and purpose of a controller design study, Proc.IEEE Conf.Decision Contr., 1991, 15791580.
[16] O.Yaniv and I.Horowitz, Ill-conditioned plants: A case study, Proc.IEEE Conf.Decision Contr., 1991, 1596-1600.
[17] C.M.Fonseca and P.J.Fleming, Genetic Algorithms for Multiobjective Optimization: Formulation, Discussion and
Generalization, Proc.5th Inter.Conf.on Genetic Algorithms, 1993, San Mateo, California, 416-423.
[18] A.J.Chipperfield and P.J.Fleming, The matlab genetic algorithm toolbox, IEE Col.on Applied Control Tech.Using MATLAB,
1995.
[19] Kay Chen Tan and Yun Li, Multiobjective Genetic Algorithm based time and frequency domain design unification of linear
systems, IFAC/IEEE Int.Symp.on Art.Intel.in Real-Time Control, Sept.1997, Kuala Lumpur Malaysia, 61-66.