Multiobject ive Optimization of Fuzzy Logic Scheduled Controllers

for Missile Autopilot Design

Anna Blumel, Brian White

Department of Aerospace, Power and Sensors, Cranfield University
Shrivenham, Wilts, SN6 8LA, UK
A.L.Blumel@rmcs.cranfield.ac.uk,B.A.White@rmcs.cranfield.ac.uk

Abstract solutions that have significant overshoot. Hence

there is a need to determine the trade-off mecha-
This paper describes the design of a F’uzzy Logic nism for such conflicting requirements and to pro-
Scheduled Controllers used for side-slip velocity vide a method for designing such systems. This
control of a missile autopilot design. A multiob- implies that any method of tuning the trajectory
jective evolutionary algorithm is used to determine control parameters should consider the four objec-
the membership function distribution of the fuzzy tives simultaneously: the usual single-objective op-
trajectory controller within an outer loop control timization problem has become multi-objective op-
system. Scaling factors of the FLC inputs and out- timization. This will produce multiple solutions
puts for each required demand are obtained by us- rather than a single solution associated with a single
ing a polynomial fit for a large range of multiple ve- objective. There is also considerable uncertainty in
locity demands (lg,5g,lOg,lSg-lateral acceleration the dyanmics associated with the missile airframe
equivalent). The design meets mulitple objectives which has to be taken into account in the design
related to closed loop performance such as: steady process. In previous research 111 a combination of
state error, overshoot, settling and rise time. Mul- an input/output linearisation technique (nonlinear
tiple solutions are obtained simultaneously by using control law) and a fuzzy logic trajectory controller
nondominated sorting for forming the Pareto front, has been considered solving the scalar optimization
combined with a reference point approach to incor- problem. This format is retained but the fuzzy logic
porate preference information into the GA to direct trajectory controller is now designed using the mul-
the search towards feasible desirable areas which tiple objective approach.
satisfy specific values of the objectivesSimulationn
results are presented showing the fuzzy gain surface The paper will look at designing an autopilot for
and extreme models in the multip[le model popula- a highly nonlinear missile over large ranges of inci-
tion. dence. The technique will obtain multiple solutions
for sideslip velocity control of the missile for a large
range of multiple demands, thus exercising the non-
linear model over its entire range. Sideslip veloc-
1 Introduction ity demands are considered for both pitch and yaw
planes, using the missile rudder and elevator as con-
Many problems involve simultaneous optimization trol surfaces hence yielding a system with 2 inputs
of multiple objectives and the design of a missile and 2 controlled outputs. The autopilot design uses
autopilot is one such application area. One of the an evolutionary algorithm optimisation approach to
main requirements for an autopilot design is to yield a multiple model description of the airframe aero-
a fast response with the minimum of overshoot with dynamics. This is used to determine the member-
good accuracy. Fast rise time or settling time can ship function distribution within the outer fuzzy
measure the speed of response, and overshoot and loop control system by using a multi-objective e v e
steady state level can then complete the measure- lutionary algorithm that meets objectives related
ment of the time response. These measurements to closed loop performance such as: rising and set-
or objective functions will produce conflicting re- tling time, steady state error, and overshoot. Fuzzy
quirements as fast responses tend to produce over- scheduled controllershaving scaling factors on both
shoot and damped responses slow down the speed inputs and outputs by using polynomial fit to allow
of response. Good steady state error performance for a large range of required demands.
usually demands high gains or the inclusion of in-
tegrat0rs.n Both of these will produce less stable

(C)u)olIEEE
0-7803-7078-3/0V$l0.00 Page: 1758
 Figure 1. Tmjectory control design

2 f i z z y Autopilot Design
The autopilot design process is shown in figure 1.
The autopilot consists of an inner linearisation loop,
together with an out fuzzy controller to improve the
robustness of the design. The missile dynamics are
given by: where the aerodynamic coefficients C,, and Cyc
are functions of Mach number M and incidence
= f ( 4+ A f ( 4 + (dz)+ A g ( 4 b (1) u. A state feedback controller has been devel-
oped which results in linear, decoupled closed loop
Y = h(4 input/output behavior, and is given by the non-
which are non-linear and have significant uncer- linear control law U = y, derived using the feed-
tainty Af(z) and Ag(z). The description of the back linearisation technique [3]. A specified track-
model is obtained from data supplied by Matra- ing performance for lateral velocity control is thus
BAE and detailed in the Horton report [2]. cheiveded but the design assumes a nominal model
obtained from wind tunnel tests with exact knowl-
The angular and translational equations of motion edge of aerodynamic coefficients. In practice how-
of the missile airframe are given by: ever, this assumption is not valid as each missile will
1 1 be close to, but not the same as, the wind tunnel
4 = zT,;'PVoSd( 2dCmpq + Cmww + VoCrnq~) model. The aerodynamical functions will be differ-
1 ent from their nominal values by the uncertainties
w = --pvos(cz,w
2m
+ V0Czvq)+ uq (2) Af(z) and Ag(x). The feedback linearisation con-
1 troller will no longer meet the desired performance
i. =
1
~I;'pVoSd(~dCnrr+ Cnwv+ VoCnci) specifications and hence will not be robust to these
1 uncertainties. The size of the uncertainties Af(s)
-2m + pv of c y c o - Ur
P~o~(~~u (3) and Ag(z) can be assessed by scoping the size of
the uncertainties in the aerodynamic coefficient C,,
where the body angular velocties are ( T , q ) (rad/sec) and Cy(, together with the static margin X,,. Un-
and the lateral translational velocities are (U,eo) certainty ranges of f 2 5 % change for C,,,f15% for
(m/sec) . Cy,, and the most sensitive coefficient, X,,, has an
uncertainty range offl.5% have been used in this
Lateral control of the missile is usually accom-
paper as representive uncertainties in the respecive
plished by controlling either the lateral acceleration
coefficients.
or the lateral velocity. For this paper, the control
of the lateral velocity will be considered as it does In order to achieve robust performance against un-
not have the same deficiency in relative degree that certainties, a fuzzy logic outer loop controller (tra-
control of the lateral acceleration exhibits [2]. The jectory outer loop) suppliments the feedback lin-
lateral velocity in equation (3) can be expanded us- earisation controller. Two fuzzy logic trajectory
ing the parametric relationships detailed in [2], to controllers are used in the outer loop for the hor-
give: izontal U, and vertical w channels respectively. In
order to capture the uncertainty structure a poly-
v = V0(C,,v + VoCyc() - U r tope of models is calculated that represent the ver-

Page: 1759
 tex models for the parametric uncertainites in the may be found in [5], and is based on several layers
aerodynamic coefficients. This family is used within of classificationsof the individuals. To classify the
a GA optimisation procedure to tune the member- individuals, the population is ranked on the basis
ship functions and the rules of the fuzzy logic con- of non-domination: all non dominated individuals
troller. are classified into one category (with a dummy fit-
ness value, which is proportional to the population
The feedback linearisation controller designed using
size, to provide an equal reproductive potential for
the nominal model [3]produces a natural fiequency
these individuals). To maintain the diversity of the
of w,, = 50(rad/sec) and a damping factor C = 0.7
population, these classified individuals are shared
for the closed loop system and a fast 250[rads/sec]
using their dummy fitness values. Each category
second order linear actuator is also included. The
of non dominated individuals is considered in turn
closed loop performance is twice as fast as the open
and the process continues until all individuals in
loop airframe, hence will produce a challenging ro-
the population are classified.
bustness problem.
The evolutionary algorithm behavior in terms of
The fuzzy logic trajectory controller has been de-
convergence and searching through feasible regions
signed based on a fuzzy inference engine, as a two
for acceptable solutions will depend on how the
input - one output system with five membership
multiple objectives are used . In a previous pa-
functions for each variable. The membership func-
per [l]we have used the surrogate additive function
tions’ positions and the rule parameters are gener-
which transforms the vectorised multi-objective
ated using an evolutionary algorithm.
problem into a scalar optimisation problem. In this
paper we have defined the closed loop performance
criteria as four competing objectives by using the
3 Multi-objective Optimiza- reference point approach [SI. Other researchers [7]
have applied similar ideas to a gas turbine engine
t ion model.
Competing objectives force a trade-off solution to
The proposed framework maintains a population of be sought. The tradeoff information generated by
fuzzy rule sets with their membership functions and the evolutionary algorithm can contribute to a bet-
uses the evolutionary algorithm to automatically ter understanding of the complexity of the problem.
derive the resulting fuzzy knowledge base. A hy- Generally, the objective criteria are not directly
brid real valued/binary chromosome has been used comparable and their numerical values may differ
to define each individual fuzzy system. The evo- considerably. This can make the tradeoff strategies
lutionary algorithm [4] follows the usual format of difficult to define. A procedure for normalisation
ranking, selection, crossover, mutation and evalua- has been used to convert the criteria yj(X) into a
tion but the real and binary parts of the chromo- dimensionless function ~ ( x E) [0,1] has been used
somes are processed separately. A multi-objective to make the trade-off problem easier. One such nor-
approach was used to identify good solutions. A malisa5tion procedure is the optimistic reference
population of 100 individuals was maintained by point approach [SI which redefines the objectives
the algorithm and at each generation, 20 individu- relative to an ideal values 9;:
als were selected for breeding. Crossover was per-
formed at a rate of 0.9, with intermediate crossover
being used for the real values and uniform multi-
point crossover for the binary part. A mutation
rate of 2/137 was used. Selective pressure (SP) of
1.7 is used. The high crossover and low selective If the ideal value g; is small or + 0, an alterna-
pressure is designed to slow convergence in order to tive form can be used:
prevent local optimum being selected. The twenty
new individuals were evaluated and then concate-
nated to the old population, forming a set of 120
individuals. Non-dominated ranking was then ap-
plied to this set and the best 100 were taken for the
where gjmaZand gj,,,in are respectively the max-
next generation.
imum and minimum values of the criterion yj(x)
The non-dominated ranking method was used in in x. The optimistic reference point method is ap-
the evolutionary algorithm to allow the multi- plied to all four closed loop performance criteria:
objective problem to be handled easily. A detailed rise time, steady state error, overshoot and settling
description of the non-dominated ranking process time, defined as follows:

(C)UIol IEEE
0-7803-7078-3/0U$l0~00 Page: 1760
 0 Side-slip velocity steady state error: tuning of a PID controller. In [9Jthe firing of the
rules in a fuzzy controller is described with different
ET; - E r j ( x ) values for the scaling factors. This demonstrates
, j E [ l , . . . ,m].
%(‘) = Er,,,, - Et-jmin the fact that the adjustment of the factors is equiva-
(7) lent to the reconstruction of the membership func-
tions in the rule-base.
Overshoot:
The three scaling factors (error,derivative of error
- osj ( x )
os; and output) for each required lateral acceleration
, j E [ l , . . ,m].
vj(2) = osjmoz - osj,in demand lg, 29, . . . ,159 have been determined via
steady state analysis of the model polytope using
the fixed gain trajectory controller around the nom-
0 Rise time: inal model. Then a polynomial fitting has then been
evaluated to interpolate between the specific veloc-
ity demands. The scaling factor for the velocity
error is lStorder:
Settling time:
T s -~ T s ~ ( x )
, j E [ l , . .. , m ] . and the polynomial curve for the velocity error scal-
qj(x) = Tsjmaz- Tsjmin ing factor is shown in figure 2. The scaling factor for

Table 1 defines the reference point parameters.

4 Fuzzy gain scheduled con-

trollers for multiple de-
mands
The fuzzy controllers for the horizontal and verti-
cal channels are shown in figure 1. The inputs to
the controller consist of the sideslip velicity error
and its derivative, and the output is the input to
the derivative of error is of a 3rd order polynomial:
the feedback linearisation control loops. Each input
variable is covered using 5 membership functions.
SCv-erd =f(vd) = + P2.i + plvd +PO (12)
The problem of finding a single set of rules for the
polytope of models is compounded by the fact that and the output scaling factor is of ldtorder poly-
the nonlinear aerodynamics produce a large range nomial.
in model characteristics. It is not possible for a s c o u t = f ( v d ) = q l v d + @J (13)
single controller to satifactorily control all models where p o , . .. ,p4 and qo, q1 are the polynomial fit
to the required specification. In order to solve this coefficients for each scaling factors respectively.
problem, a scaling factor that is a function of the
parametric variables is considered. The greatest
variation is seen in the change of incidence cr. For
a contstant forward speed V, this translates into 5 Results
a side-slip velocity dependency. ‘ihs fuzzy logic
controller inputs and outputs are thus scaled by a The optimization procedure produces multiple so-
scaling factor, and the rule parameters are treated lutions and has been able to tune the membership
as constants to be determined by the evolutionary function parameters simultaneously for multiple de-
algorithm. This concept has been used by other mands. In a typical run, about 90% of the solutions
workers [SI to improve the performance and stabil- in the final population are non-dominated. Fig-
ity of fuzzy controllers. In Bonissone’s chapterQ[S], ure ?? shows a detailed trade-off for the nondomi-
scaling factors of an FLC have been tuned by GA’s. nated solutions at the last generation. Two alter-
Some time scaling factors are used to fine tune the native solutions have been presented in figure 6 to
performance of the system in a similar way to the show the effect of the interpolative mechanism of

(C)U)Ol IEEE
0-7803-7~8-3/0u$l0.00 Page: 1761
 I I I
Reference points
Ideal point
Stkady State Error
I Er: = 0.0 (%) I
Settling time
Ts: = 0.15 (sec) I Rising time
Trl = 0.08 (sec)
Overshoot
I 0s: = 4.5 (%)

Table I . Closed loop performance criteria

the fuzzy scheduled controllers when multiple de- been exercised for a large range of velocity demands
mands are used. Solution (b) has very little over- (lg,5g,lOg,l5g-lateral acceleration equivalent).
shoot, almost no steady state error and good set-
tling time. Solution (c) has a very slow rise time,
very little overshoot, a long settling settling time References
but a good steady state error. Solution (b) is cho-
sen as the final controller. On the left side of fig-
ure 6(a) is the nonlinear control surface found by (11 E. H. A. Blumel, B. A. White, “Designof robust
fuzzy controllers for aerospace applications,”
the GA’s of the FLC. in Proceedings of the 18 NAFIPS Conference,
vol. 1, pp. 438-442, 1998.
0.41 . . . . 7
B. A. White, “Non-linear autopilot of the hor-
ton and generic missile,” tech. rep., Dept.
of Aerospace, Power and Sensors, Cranfield
0’ I University-RMCS, Shrivenham, Wilts, SN6
o 0.1 0.2 0.3 0.4
&bra BLA, October 15 1998.
1,

A. L. B. A. Tsourdos and B. A. White, “’Ikajec-

tory control of a nonlinear homing missile,” in
Proc. of the 14th IFAC Symp. on Aut. Control
0 02
c
04
__ *
06 08
in Aerospace, 1998. Korea.
A. Zalzalaand P. J. Fleming, eds., Genetic algo-
r i t h m in engineering systems. The Institution
of Electrical Engineers, 1997.
N. Srinivas and K. Deb, “Multiobjective opti-
0 OS 1 15 2 25
obumme4 mization using nondominated sorting in genetic
algorithms,” Evolutionary Computation, vol. 2,
Figure 3. Dade-ofl between each objectiues at no. 3, pp. 221-248,1995.
last gen 250
H. Weistroffer, “A combined over and under-
achievement programming approach to multiple
objectives decision making,” Journal of Large
Scale Systems, vol. 7, pp. 47-58, May 1984.
Conclusions C. M. Fonseca and P. Fleming, “Multiobjective
optimization and multiple constraint handling
Evolutionary algorithms have been successfully ap- with evolutionary algorithms - part 2: Appli-
plied to a multimodal search space in the context of cation example,” IEEE Zhnsactions on Sys-
determining the control parameters for the missile tems,Man and Cybernetics, vol. 28, pp. 38-47,
autopilot. The optimization has used multiple ob- January 1998.
jectives related to closed loop performance: steady
C. C. Bonivento and R.Rovatti, eds., Fuzzy
state error, overshoot, settling and rise time. Mul-
Logic Control Advances in Methodology. World
tiple solutions have been obtained using a Pareto Scientific Publishing CO, 1998.
based approach combined with a reference point ap-
proach to incorporate preference information into Y.-Y. Chen and C. Pemg, “Input scaling fac-
the GA to direct the search towards feasible de- tors in fuzzy control systems,” in Pmc. IEEE,
sired area from which the designer can choose solu- pp. 166&1670,1994.
tions that satrisfy the performance specifications in
the face os parametric uncertainty. The design has

(C)U#)lIEEE.
0-7803-7@78-3/0U$l0.00 Page: 1762
 4 ...................... -.... .........................................................
7

-,Y -im . ,.

........

+-7.-.7.- - 7 7............... "" ......_........... '" ........

a) h z z y control surface for 15g [150m/sec"] Side-slip velocity and lateral acceleration for 15g

b) Best Solution for steady state error Phase portrait for steady state solution

c) Slow rise time solution

Figure 4. Multiple solutions for multiple de-

man&

(C)U)ol IEEE.
0-7803-7078-3/0l/$10.00 Page: 1763