Download as ps, pdf, or txt
Download as ps, pdf, or txt
You are on page 1of 6

MULTICRITERIA CONTROL SYSTEM DESIGN USING

AN INTELLIGENT EVOLUTION STRATEGY WITH


DYNAMICAL CONSTRAINTS BOUNDARIES
To Thanh Binh, Ulrich Korn,

Institute of Automation, University of Magdeburg, Germany


E-mail: binh@infaut.et.uni-magdeburg.de
E-mail: korn@infaut.et.uni-magdeburg.de

Abstract. This paper represents a multivariable controller design methodology which


is based on the solution of the multiobjective optimization problem using a multiob-
jective evolution strategy with dynamical constraints boundaries. The evolution strat-
egy allows to handle simultaneously all the types of multiple design objectives and
constraints (hard and soft), and to achieve the good approximation of the complete
tradeo set between the design objectives while satisfying all the given constraints.
This evolution strategy is implemented in the MATLABbased  environment and it
can be used to solve many optimization problems (Binh et al., 1996).
 MATLAB is the Trademark of the MathWorks, Inc.

Keywords. Mutation, Reproduction, Selection, Paretooptimal Set and Dynamical


Constraints Boundaries.

1. MOTIVATION minimized). Constraints, on the other hand, can often be


described either as the hard requirements, which must
The design of many multivariable control systems is es- be satised before the optimization of the remaining, or
sentially a multiobjective design problem, in that the the soft one, which should be maintained within specic
design objectives reect the need to achieve the eco- bounds and their temporary violations are allowed in
nomic performance of the system, specications on the order to satisfy the other objectives. In control systems,
product quality and equipment safety, etc. These prac- the closedloop system stability is an example of the
tical requirements are normally converted into mathe- hard constraints, because most performance measures
matical expressions, stated either as objectives or con- are not dened for unstable systems. In the most cases,
straints (Boyd and Barratt, 1991). The objectives  the the soft constraints can be expressed in terms of func-
functions of the variables to be optimized, for exam- tion inequalities, for example, upper or lower bounds on
ple, settling times and rise times of the control outputs, the output or the manipulated variables.
zero steadystate tracking errors, requirements on the If we describe quantitatively all these objectives as a
pole assigment in specied regions, explicit closedloop set of N design objective functions fi (x) 8i = 1; N ,
performance and robustness by the analytical optimiza- where x denotes the n-dimensional vector of design pa-
tion methods as LQG, H1 mixedsensitivity and  rameters, the design problem could be formulated as a
synthesis, need to be simultaneously achieved (normally
multiobjective optimization problem:  When the desired feasible region is reached, the op-
timization begins to search for the set of pareto
min f (x) = min (f1 (x); f2 (x);    ; fN (x)): optimal solutions in
.
x x
This evolution strategy is introduced in the next sec-
Here, the objective variable x must be in an universe
tion, and based on the main genetic mechanisms: Muta-
of the n-dimensional space Rn, which is determined by tion, Reproduction and Selection, and on the ranking ac-
the soft and hard design constraints, that means, cording to the actual concept of pareto optimality. This
\ evolution strategy guarantees the approximation of the

=
(hard)
(soft) ; paretooptimal solutions by the population in the fur-
ther next generations better than in the current gener-
where
(hard) is the feasible region for the hard con- ation. Some examples to illustrate the eciency of the
straints and can be expressed in terms of any functions evolution strategy are shown in section 3.
of the objective variable x;
(soft) is the feasible region
for the soft constraints, i. e.
2. THE MAIN ALGORITHM OF THE EVOLUTION

(soft) = fx 2 Rn : ci(lower)  ci (x)  c(upper)


i ; 8i = 1; Lg:
STRATEGY
The scalar value L is the number of the soft constraints. 2.1 Representation of the individual
By the solution of the multiobjective optimization with The major dierence between this evolution strategy
constraints the following problems can often be met: and the wellknown evolution strategy in (Binh, 1994),
 The design objectives are in conict, so that it is (Binh and Korn, 1996a), (Binh and Korn, 1996b) lies in
not possible to improve any of the objectives with- the representation of an individual.
out deteriorating at least one of the other objec- Here, an individual includes four properties as below:
tives. This is known as the concept of pareto op-
timality (Goldberg, 1989). Using this concept, the  the so-called life environment w,
solution of the multiobjective optimization can be  the objective variable x = (x1; x2;    ; xn),
seen as the set of paretooptimal solutions  the  the strategy parameter vector s = (s1 ; s2 ;    ; sn)
set of nondominated or noninferior vectors in the  the corresponding objective function vector f (x),
objective function space. The rst problem is how that means:
to get the paretooptimal set for each multiobjec-
tive optimization problem. It can be solved by us-
ing the evolution strategy with the given feasible I nd = (w; x; s; f ):
starting point for the optimization (Binh, 1994), The strategy parameter vector s has the same dimen-
(Kahlert, 1991), (Binh and Korn, 1996a). sion as the objective vector x and describes the "personal
 The second problem is concerned with the choosing experiences " of each individual by the reproducing its
a starting point for the optimization that satises osprings. It is interesting that at the begin of the evo-
all the given constraints (that means a point of an lution all individuals have very little experiences (i. e.
unknown universe
). In the most cases the soft all elements of the vector s are small),but in every gen-
constraints are so strictly that a starting point sat- eration, the better experiences of the individuals can be
isfying only the hard constraints can be chosen. The accumulated and improved (i. e. all elements of the vec-
optimization with the wellknown evolution strat- tor s are enlarged). When the global minimum or the
egy can not be started until the feasible starting paretooptimal set is found, the vector s must become
point (it is in the feasible region
) is found. small to guarantee a concentration with the high density
Both these design situations motivate the development of the current population on it.
of the new multivariable controller design methodology The life environment is the new property of the indi-
using the multiobjective evolution strategy with the dy- viduals. The addition of this property into an individ-
namical constraints boundaries. Instead of the very long ual comes from the following motivation. The subspace
searching for the feasible starting point, the optimiza-
(soft) can be seem the ideal life environment for all indi-
tion should be executed in the universe
(hard) in two viduals. But in the few rst generations the individuals
following steps: can live in the real current life environment that is more
 Search for the feasible region  the subspace of dierent to the ideal life environment. Therefore, the

(hard) . individuals usually have to nd out the way to improve
their life environment until the ideal life environment pareto optimality, i. e. choosing only the noninferior in-
will be reached. This property is used: dividuals. In another case, the better individual is the
 to make the decision if the individual is viable, individual, which satises the given softconstraints.
 to create the life environment for the current pop-
ulation, In other words, the priority by the selection is as follows:
 and to rank the current population. (1) viability,
Because only the soft constraints, expressed in terms of (2) satisfaction the given softconstraints,
the inequalities, are mentioned here, the subspace
(soft) (3) noninferiority.
can be seen like a hyperparallelogram in the space of the
soft constraints. By this way, the current life condition of
an individual is described also as a hyperparallelogram
dened by the current softconstraints boundaries. The 2.3 The main mechanism
current life condition of an individual is said to be ideal
if the given softconstraints are reached by the individ- The algorithm for the mutation, the reproduction and
ual, that means
(soft) includes its hyperparallelogram. the selection are quite same as by the wellknown evo-
Mathematically, the informations about the position of lution strategy, but the comparision between the indi-
the hyperparallelogram for an individual can be saved viduals is based on the above ranking scheme.
in the vector w consisting of the current lower and cor- The mutation allows to evolve the own strategy param-
responding upper values of the softconstraints, also: eters s of an individual so that number of its viable o-
springs is biggest and they can lie in the downclimbing
w = (c1(lower);    ; cL(lower); c(upper)
1
(upper) ):
;    ; cL direction. By this way, the current population runs very
quickly into the local minimum.
Then, the life environment of the current population The reproduction is performed on the strategy parame-
can be characterized as a hyperparallelogram, so named ters as well as on the objective variables using the well

(current)
(soft) , including all hyperparallelograms for every in- known twopointscrossover recombination.
dividual and
(soft) .
By the selection the paretooptimality in the objective
Using this representation and the ranking algorithm (see function space is used. The algorithm concludes the as-
section 2.2), the life environment of the current popu- signing rank 1 to the nondominated individuals and
lation will be improved in every next generation by in- rank 2 to the other of the current population. After some
creasingly diminishing the current hyperparallelogram generations, the current population reaches a neighbour-
into
(soft) . In other words, we have the optimization hood of the set of paretooptimal solutions and the
process with the moving or dynamical boundaries of number of the noninferior individuals is, therefore, big
constraints (namely softconstraints). enough. A new problem with the approximation of a
continuum set (namely the set of paretooptimal so-
lutions) by the nite set (the current population with
2.2 Ranking the individuals the nite population size) appears here. Generally, the
best solution for this problem is not found. This situ-
In the evolution strategy it usually needs to choose the ation leads to the great unstability of the population
better individual from two (by the mutation and the in the next generations (that means the population will
reproduction) or from many individuals (by the selec- continueosly be moving on the surface of the pareto
tion). Without loss of generality, we consider only the optimal set). From this reason, a global picture of the
ranking between two individuals in the universe
(hard) . paretooptimal solutions can not be achieved in every
It can easily be generalized for the comparision of more generation.
individuals and for the universe
. To avoid it and to create the uniform distribution of the
To do it the following selection criterions are mentioned: current population on the tradeos surface we recom-
mend the following selection algorithm:
Criterion 1. An individual is said to be viable i it sat-
ises the current softconstraints
(current)
(soft) . Algorithm 1. Let
Criterion 2. When both the viable individuals either do (min) = (min f = (f1(min) ;    ; fN(min))
not satisfy or satisfy the given softconstraints simulta- f 1 ;    ; min fN )
neuosly, the ranking is based on the actual concept of (max) = (max f ;    ; max f ) = (f (max) ;    ; f (max) );
f 1 N 1 N
where the minimum and maximum operators are per- with the hard constraints:
formed along each coordinate axes of the objective func-
tion space for all individuals of the population. 0  x1  18 and 0  x2  18
Then, the current tradeos surface is bound in the hy-
perparallelogram H dened by f (min) and f (max). and the soft constraints:
 Dividing each interval [fi(min); fi(max) ] into Npop small 0  x1  6 and 12  x2  18:
sections i , i. e.:
(max) (min) This optimization problem has the following special prop-
i = fi fi
: erties:
Npop
 the objective function has 8 local minima and the
 In the i-th coordinate axes of the objective function global minimum at the point x = (15; 3).
space, the best individual in each of NNpop  not the point (15; 3) with the smallest value (f = 1)
+ k the rst of the objective function but the point x = (3; 15)
sections is selected, where k is an integer number. with the objective function value f = 4 must be
the global solution of the above optimization. In
other words, the global minimum of the optimiza-
3. SOME DEMONSTRATIONS tion without the softconstraints does not identify
to the one with the above softconstraints.
In this section, we would like to illustrate the eciency
of the new evolution strategy by the optimization of Starting at the point x = (15; 1) lying outside and far
some mathematical functions. Because their solutions from the feasible region:
are wellknown, then it can be good understood how
to handle the multiple objectives and constraints by fx 2 R2 : 0  x1  6 and 12  x2  18g;
the evolution strategy. By the application for the mul-
tiobjective control system design, to solve the socalled the current population should know that the point x =
IFAC93 Benchmark Problem (Whidborne and etc., 1995), (15; 3) is only a temporary goal until at least an (feasi-
we have got the results which are as good as in (Binh ble) individual reachs the feasible region. The starting
and Korn, 1996a) and (Binh and Korn, 1996b) but bet- population does not satisfy the given softconstraints,
ter than in (Whidborne and etc., 1995). i. e. it is not viable. To guarantee that it can be viable,
the current softconstraints have to be enlarged to the
region H0 :
3.1 The Multi Modal Function
fx 2 R2 : 0  x1  18 and 0  x2  18g:
The objective function is described as follows:
The evolution strategy shows that some feasible individ-
8 (x uals can be found after 8 generations and the correspon-
>> 1 3)2 + (x2 3)2 + 8 dent softconstraints region H8:
>> for 0  x1  6; 0  x2  6
>> (x1 3)2 + (x2 9)2 + 5
fx 2 R2 : 0  x1  16 and 2  x2  18g
>> for 0  x1  6; 6  x2  12
>> (x1 3)2 + (x2 15)2 + 4 is smaller than H0. In the 20-th generation the feasible
>> for 0  x1  6; 12  x2  18 region is found. The search for the desired global min-
>> (x1 9)2 + (x2 3)2 + 7 imum of the optimization problem begins at the 21-th
>< for 6  x1  12; 0  x2  6 generation, and successfully ends in the 36-th generation
f (x1 ; x2) =
(x1 9)2 + (x2 9)2 + 9 (see the Figure 1).
>> for 6  x1  12; 6  x2  12
>> (x1 9)2 + (x2 15)2 + 3
>> for 6  x1  12; 12  x2  18 3.2 The biobjective optimization
>> (x1 15)2 + (x2 3)2 + 1
>> for 12  x1  18; 0  x2  6 We consider the following biobjective optimization with:
>> (x1 15)2 + (x2 9)2 + 6
>> for 12  x1  18; 6  x2  12
( ) = x21 + x22 ;
>: (x1 15)2 + (x2 15)2 + 2 f1 x1; x2
for 12  x1  18; 12  x2  18 f2 (x1; x2) = (x1 5)2 + (x2 5)2;
The starting population The 8−th iteration
18 18

16 16
+4 +3 +2 +4 +3 +2
14 14

12 12

10 10
x2

x2
+5 +9 +6 +5 +9 +6
8 8

6 6

4 4
+8 +7 +1 +8 +7 +1
2 2
Start Point Start Point
0 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
x1 x1

The 20−th iteration The 36−th iteration


18 18

16 16
+4 +3 +2 +4 +3 +2
14 14

12 12

10 10
x2

x2
+5 +9 +6 +5 +9 +6
8 8

6 6

4 4
+8 +7 +1 +8 +7 +1
2 2
Start Point Start Point
0 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
x1 x1

Fig. 1. Optimization process for the multi modal function


with the hard constraints:  to get the global survey about the structure of the
paretooptimal set.
0  f1  1100 and 0  f2  1000  to choose easily the best solution from it by using a
lot of the dialogue methods in the MATLAB-based
and the soft constraints: environment (Binh et al., 1996).
5  x1  10 and 5  x2  10 By the solution of a lot of optimization problems: both
the scalar and the multiobjective, it is well known that
The optimization proccess with the starting point at the evolution strategy possess essentially many good
x = ( 10; 30) is shown in the Figure 2. It is clear, every properties: high robustness to get a feasible region and
individual is good able to learn how to get the better so- to nd the global minimum or the set of paretooptimal
lution in each optimization situation by adaptiv chang- solutions (Binh et al., 1996).
ing its own strategy parameter. The feasible region can,
therefore, very quickly be found.
5. REFERENCES
4. CONCLUSION Binh, T.T. (1994). Eine Entwurfsstrate-
gie für Mehrgröÿensysteme zur Polgebietsvorgabe.
The new evolution strategy allows the users: PhD thesis. Institute of Automation, University of
Magdeburg, Germany.
 to overcome the missing a priori knowledge about Binh, T.T. and U. Korn (1996a). An evolution strategy
the structure of the feasible region that is very of- for the multiobjective optimization. To appear in
ten met in the system control design. Independent Mendel96 Conference, in Brno , Juni 1996.
from the starting point, the evolution strategy tries Binh, T.T. and U. Korn (1996b). Robuster Regleren-
to search the better life conditions for every popula- twurf für das IFAC93 Benchmark Problem mittels
tion and gets the ideal life condition after some gen- einer globalen Mehr-Ziel-Evolutionsstrategie. To
erations. That means the current individuals and appear in Automatisierungstechnik.
population are able to adapt themselves to various Binh, T.T., U. Korn and J. Kliche (1996). Evolution
optimization situations. Strategy Toolbox for use with MATLAB. Techni-
The starting population The 5−th iteration
1000 1000

900 900
Start Point Start Point
800 800

700 700

600 600
f2

f2
500 500

400 400

300 300

200 200

100 100

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
f1 f1

The 10−th iteration The 30−th iteration


1000 250

900
Start Point
800 200

700

600 150
f2

f2
500

400 100

300

200 50

100

0 0
0 200 400 600 800 1000 0 50 100 150 200 250
f1 f1

Fig. 2. Biobjective optimization process


cal report. Institute of Automation, University of
Magdeburg, Germany.
Boyd, S.P. and C.H. Barratt (1991). Linear Controller
Design: Limits of Performance. Prentice Hall, En-
glewood Clis, New Jersey 07632. Stanford Univer-
sity.
Goldberg, D.E. (1989). Genetic algorithms in Search,
Optimization, and Machine Learning. 1. ed..
AddisionWesley Publishing Company, Inc.. New
York, England, Bonn, Tokyo.
Kahlert, J. (1991). Vektorielle Optimierung mit Evo-
lutionsstrategien und Anwendung in der Regelung-
stechnik. Forschungsbericht VDI Reihe 8 Nr.234.
Whidborne, J.F. and etc. (1995). Robust control of an
unknown plant  the IFAC93 benchmark. Int. J.
Control 61(3), 589640.

You might also like