You are on page 1of 6

Robust Fuzzy Controllers

Heng-Ming Tai and Sujeet Shenoi


Center for Intelligent Systems Keplinger Hall, University of Tulsa Tulsa, OK 74104-3189, U.S.A.
email. ta@ohm. ee. utulsa.edu, s u j e e e u l e r .mcs utulsa.edu

Abstract:

Robustness is an important requirement placed on welldesigned fuzzy controllers A robust fuzzy controller can deal with significant and unpredictable process changes When designing a fuzzy controller it is partmlarly important to ensure that reasonable performance is obtained for parameter vanations within -and outside the design ranges This paper presents a systematic design methodology for tuning a class of fuzzy controllers using cell maps of dynamic systems Cell map information is incorporated in a genetic search for tuning the mle-consequent function parameters of Taka@Stigeno hzzy controllers The cffectiveness and robustness of the designed fuzzy controller are evaluated urmg the inverted pendulum control problem

I INTRODUCTION
Fwq- modeling and control techniques have seen wide commercial and industrial application. These methods can incorporate linguistic human operator experience and engineering knowledge in problem solving. The main problems in designing fuzzy controllers are: (1) the identification of fuzzy rules and the tuning of the associated membership functions; and (2) the realization of an optimal control policy and Stability. To solve the first problem, many automatic design methods based on neural networks 191, genetic algorithms [ 1 I ?121, or gradient descent method [13] have proved useful. For example. when using genetic algorithms to determine membership functions and rule-consequents, the f u z q system parameters are genetically coded and fitness functions expressing the control objectives are selected. The performance of the design method and that of the designed fizzy controllers can be improved by taking advantage of the robustness of genetic algorithms and their ability to efficienty search large problem spaces. This paper focuses on designing a stable fuzzy controller which is time-optimal and robust. This goal is aclueved using the cell mapping technique [7] and genetic algorithms [41.

A cell map conveys important feedback control knowledge about s y s t e m dynarmcs of a closed-loop system, including system trajectories and the region of stability. This information is readily incorporated into a genetic search procedure for designing timeqtimal fuzzy cantrollers with high levels ofperformance. Chen and Tsao [3] first used cell maps to analyze fuzzy systems and obtain global evaluations of the stability and performance of fuzzy controllers. However, they did not actually use cell maps in f i c r ; q controller design. Other researchers, notably Smith and Comer [14,15] and Vachtsevanos and co-workers [10,18], have applied cell mapping techniques to the design of optimal f k u y controllers. Smith and Comer's work 114,151 focuses on the synthesis of optimal Takagi-Sugeno fuzzy controllers [ 171. They employ the least-mean-square algorithm [ 191 to fine-tune the control policy embodied in a cell space approximation of an optimal control solution 161. Cell map accuracy is increased using a non-Worm quantization scheme and a variable time interval procedure for avoiding spurious equilibrium cells. The designed technique yields a minimum-time fuzq controller with good controllability and stability. The optimal design methodology developed by Vachtsevanos and his co-workers [10,18] does not construct an explicit fuzzy controller as a function of state variables. Instead, control action information is embodied in a cell map. In their design, both state and control variables are partitioned into sub-intervals called "cellgroups." The evolution of cell-groups in a cell map is obtained by simulation. A search procedure then is applied to determine the optimal control a h o n for each cell in the cell space. The methodolog is v e y versatile in that the designed controller articulates the optimal strategy encoded in the search. It can be used to design a variety of optimal fuzzy controllers such as minimumenergy, minimum-time and minimum-squareerror controllers. This paper examines the robustness of fuzzy controllers designed using cell map information in a genetic search. Section 11 reviews the Takagi-Sugeno

0-7803-2129-4/94 $3.00 8 1994 IEEE

Authorized licensed use limited to: Norges Teknisk-Naturvitenskapelige Universitet. Downloaded on December 4, 2009 at 08:30 from IEEE Xplore. Restrictions apply.

fuu?. controller formulation. a modrfied simple genetic algorithm. and Ihe control information incorporated in cell maps. Section I11 presents the genetic-algorithmbased design methodology utilizing cell maps o f closedloop systems to realize time-optimal TS fuzzy controllers. The method is applied to the inverted pendulum control problem. The robustness of the designed TS controller is tested using plant parameter variations and control forces lying outside the design ranges.
11. CELL MAPPING AND GENETIC ALCdRITHMS
The goal of this work is to develop an automated methodology for designing optimal and robust TS fuzzy controllers. This section describes the TS fuzzy controller formulation, a modified simple genetic algorithm, and the cell map information to be incorporated in the genetic search.

.4. Iakagi-Sugeno Fuzzy Controller


A typical control rule in a Takagi-Sugeno (TS) fuzzy controller [ 171 has the form: IF error ( x1) is Negative Big and error-change (x2 ) is Zero THEN control action is g, ( x r,x2) = cJlxl+ c I 2 x 2 + c J 3 . The coefficients cj, in the consequent function can be adjusted in the o p t sense to m m the control rule. The final output of the TS fuuy controller,f, is computed as heweighted average of the outputs of rules in the rule base
1 -

c;.

1H] ; P ,
-.

-.

E;=,M),

where r is the number of rules and wJ is the degree of match of the premise of the jth rule. A key advantage of the TS controller formulation is that the rule consequents are crisp (usually linear) functions of the input variables. It also is easier to optimize the consequents of the rules. i.e., to optimally tune the controller using optimization techniques such as the least-mean-square method [19]. The main task in TS controller design is to adjiist the coeflicients. c,,: to obtain the desired pnrfbrniance.
! i ..\ f(Jd$ed .?IT?/LJ/CJ~ ~ t ? f l e.;/@?/*it/ltnS ~iC

The chromosome is obtained by concatenating coded parameter values. A chromosome population is a set of solutions of varying quality. Populations evolve by reproduction, crossover and mutation. Reproduction involves the selection of strings on a probabilistic basis in proportion to their relative fitness. Two selected strings (parents) in the mating pool then undergo one-point crossover. Two new strings (offspring) are crated by swapping parent bits at the crossover point. Mutations occur as occasional random alterations of bits. The crossover and mutation rates are typically in the ranges, 0 . 6 4 . 9and 0.001-0.01, respectwely [4]. A GA is particularly suited to synthesizing optimal fuzzy controllers. It is an efficient search technique using a chromosome population enables the search to proceed in parallel. The genetic operators bring about information exchange between strings. This reduces the likelihood of termination at a local optimum. A GA also is extremely flexible as it only uses candidate solution quality to guide its search. Auxiliary problem-specific informaton, e.g., derivatives or continuity properties, is not required. The simple GA model is modified in this work by representing chromosomes as integer strings instead of binary strings. Using chromosomes with integer alleles positively impacts codingldecoding and mutation. The integer representation reduces the coding/decoding time and yields a faster genetic search. Mutation of chromosomes with integer alleles is implemented simply by incrementing or decrementing integer alleles by one. C,ontrollerdesign experiments indicate that this mutation technique produces better results than changing bit values in a binary string. Moreover, an elitist strategy is employed. It retains the best string in the current population during the reproduction process. This procedure guarantees asymptotic convergence to a global optimum [161.
C. Cell Map Information

(jenetic algarittims (GAS) can Fe used to efficienllv s e a m large paraiiieter spaces for opiiiiial or near optimal solutions. A simple genetic search [J] is implemented by randomly geiicrating sets or parameter values. Each parameter set, representing a possible solution to the search. is coded as a binap string called a chromosome.

Cell mapping is a powerfid technique for analyzing the global behavior of nonlinear dynamical systems [5,7]. It approximates a continuous phase space with a discrete cell space comprising a finite number of disjoint, equalized cells. Regions lying outside the domain of interest are lumped as a single sink cell. Cells are processed cxhaustively using an unraveling algorithm [S] to compute cell transitioris The algorithm involves calling ~ip each cell in turn aiid processing it to determine its global properties. It is assumed that d l the states in a cell transit to states in a single image cell within a fised time period. and that all cell states can be approximated by the center-point of the cell. The unraveling algorithm

86

Authorized licensed use limited to: Norges Teknisk-Naturvitenskapelige Universitet. Downloaded on December 4, 2009 at 08:30 from IEEE Xplore. Restrictions apply.

identifies all the period motions. domains of attraction and cell space trajectories in a cell map. Each non-sink cell z in a cell map is characterized by a group number G(z), a step number S ( z ) , and a periodicity number P(z). Cells in the periodic motion or domain of attraction of cell z belong to the same group G(z). The step number S(z) indicates the number of cell transitions needed to transit from cell 1: to 3 cell in a periodic motion. This information is employed to construct the fitness function used in the genetic search. For detailed information about cell mappings and the unraveling algorithm readers are referred to the original work by Hsu [5.71
111. FUZZY CONTROLLER SYNTHESIS

J = a
1+

"2

"z

Ccon(z)d(Ff ( Z ) J t ) + P -~ c m ( r ) A s s t a b ( r p ( ~ )
"r

nz

This section describes how cell map infomation can be employed in a genetic algorithm for tuning TS fuzzy controllers. Unlike other cell-map-based approaches 114,181, the GA-based design technique only uses a simple, intuitive cell-map-based objective function for the genetic search.
GA;

where nz is the number of controllable cells. These cells satisfy the predicate " 2 ) . d(FjZ),Zr)is the Euclidean &stance in the cell space between the final image cell F z ) of cell z and the target cell Zt. The predicate k a b ( z ) is satisfied by each cell z whose final image cell F j z ) lies within 6 cell of the target cell. S(z) is the number of time steps taken fiom cell z to its final image cell. a and f! are positive constants. Interested readers are referred to [8] for additional details of the fitness function and its impact on controller performance. Maximizing J favors solutions containing the maximum number of controllable initial cells which can be driven to the target cell (equilibrium state). on the average, in minimum time. Therefore time-optimal behavior can be obtained by realizing a cell map with the shortest cell trajectories. Maximal controllability is realized in a cell map containing the largest stable region. IV. RESULTS AND DISCUSSION

Control
Plant

Figure 1. Fuzzy controller design methodolon


A schematic diagram of the h q controller design methodology is shown in Figure 1. An initial population of chromosomes - representing coded TS parameter sets is randomly generated. These chromosomes are decoded lo yield the actual parameter values. The parameters then are supplied to the corresponding fuzzy control rules and to the governing equations of the system under consideration. A cell map is constructed for each parameter set, and its fitness value computed. The fitness values are utilized along with three genetic operators selection, crossover and mutation - to produce ;L new generation of chromosomes. This procedure is repeated until the new population has converged to some best fitness value producing an optimal parameter set or a preset limit is reached on the generation number. To synthesize high-perfonnance fuzzy controllers it is necessary to incorporate cell map information in the fitness function for the genetic search. This function measures solution quality and provides environmental feedback to the search algorithm. It must correctly articulate the controi goals and their impact on system dynamics. The fitness function J used in this work is given by:

The effectiveness and robustness of the fuzzy controller design methadology are assessed for the inverted pendulum control problem, a benchmark problem in classical and intelligent control. Simulation experiments indicate that the cell-map-based fuzzy controller designed with a GA out-performs controllers designed by other techniques [2,9,14,15]. The controller uses less control effort, reaches its goal more directly, and consistently produces less overshoot despite large initul deviations. This paper considers its ability to deal with plant parameters and control force 1yng outside the design ranges. The control objective of the inverted pendulum problem is to bring the p i e to the equilibrium vertical position starting from an initial non-vertical pole inclination. Many appkations of intelligent and learning control to this problem use "time to failure" to gauge the quality of the controller [1,2]. However. in this paper. three measures are used to characterize the quality of a good controller. They are (1) the time taken to achieve the control objective, (2) the range of controllable initial states, and (3) the robustness of the designed controller to parameter variations. The first two goals are achieved using the design methodology and are described in detail in 181. The robustness property is evaluated bv considering variations in the nominal parameter set. Since the initial state of the pole is more important than that of the cart in determining the inclination of the

87

Authorized licensed use limited to: Norges Teknisk-Naturvitenskapelige Universitet. Downloaded on December 4, 2009 at 08:30 from IEEE Xplore. Restrictions apply.

pole, i t is necessary to focus on balancing the pole without regard to the position and velocity of the cart. The state equations governing the pole angle (xl = 8 ) and angular velocity (x2 = 6 ) or the inverted pendulum are given by 13,151
r, = x2

- $1
~~

ICOS2 ( I l l

m p +mc

where ge = 9.8 m/sec2 is the gravitational constant, mp = 0.1 kg is the mass of the pole, me = 0.9 kg is the mass of the cart, I = 0.5 m is the half length of the pole, and f is the force applied to the cart. The required TS fkq controller is constructed with x, and x2 as the conditional variables and the force g as the control variable. The ranges used for the variables xl. x2, andfare [-1.5,1.5] radian, [-8,8] rad/= and [-50,501 Newtons, respectively. The cell space is :Formed by partitioning the phase space into 101 x 101 = 1,0201cells with qual divisions.

-1

Normalized Universe of Discourse Figure 2. Membership functions of pole angle and angular velocity.

Figure 3. Control force surface.

Triangular membership functions which overlap in their respective domains are employed (Figure 2). The resulting TS fuzzy controller thus has nine linguistic rules and 27 parameters. Each controller parameter is coded using five integer alleles. The last two digits represent the fractional part. Each chromosome in the population has length 135 (= 27 x 5). Parameters are coded and decoded by adding and subtracting 500.00 from the actual and coded values. The population size is fixed at 50. The crossover and mutation rates used are 0.8 and 0.1, respectively The chromosome with the highest fitness value is automatically advanced to the next generation. The best TS fuzzy controller derived from cell map information is obtained in the 70th GA generation. The outputfof the TS fuzzy controller is directly served as the control force. The final control surface is shown in Fig. 3. The ability of t h e designed controller to survive parameter variations successfully is tested using different pole lengths. Figure 4 shows that the TS controller designed for a pole length of 0.5 m can adequately stabilize small poles. Note that most existing approaches tawards inverted pendulum control consider bung-bung optimal control schemes, where only right and left forces of fixed magnitude can be applied. As c a n be seen in Figure 4(a), that the control force has a variable magnitude, i.e., it is a continuous function of time. A major advantage of the cell-map designed controller over other controllers (e.g., [2,9,11]) is that the range of initial states stabilized by the controller is explicitly displayed in the cell map. Graphs of the time-optimal cell and real trajectories are shown in Figure 5(a). Both trajectories start at the initial state x1 = 1.49 rad and x2 = -7.94 rad/sec. The maximum control force delivered is a good measure for assessing the quality of a control solution to the inverted pendulum problem because it determines the range of controllable initial states. Figures 5(a) and So>) show the cell map regions in which the contrdler is able to stabilize the pole for maximum deliverable forces of 50 Newtons and 10 Newtons, respectwely. In general, the larger the maximum force delivered by the actuator, the larger the controllable region. As expected, the controllable (shaded)region in the cell map in Figure 33) is reduced in size. The designed controller is robust in that it can stabilize the pole even when the maximum force delivered is less than the designed value. Time responses of the inverted pendulum system to various maximum deliverable forces are shown in Figure 6. For the initial s t a t e x1 = 1.36 rad and x2 = 0.0 rad/=, the designed controller with a maximum force larger than 50 Newtons (default value) is able to stabilize the pole. However, the controller is not successful when the maximum deliverable force is restricted to 30 Newtons.

88

Authorized licensed use limited to: Norges Teknisk-Naturvitenskapelige Universitet. Downloaded on December 4, 2009 at 08:30 from IEEE Xplore. Restrictions apply.

This phenomenon is seen in Figure 5(b) where the initial state is outside the controllable region. Figure 6 also shows that the larger the maximum deliverable force, the faster the control.

C. S. Hsu, Cell-to-Cell Mapping, Springer-Veriag, New York, 1987.


H. T. Hu, H. M. Tai and S. Shenoi, "Incorporating cell map information in fuzzy controller design," to appear in Proceedings o f the Third IEEE International Conference on Fuzzy Sjstems, 1994.
J. S. Jang, "Self-learning fuzzy controllers based on temporal back-propagation," IEEE Transactions on Neural Networks, vol. 3, pp. 714-723, 1992.

v CONCLUSIONS
Incorporating cell map infomuon in a genetic search algorithm for tuning the TS fuzzy controller results in a time-optimal controller with maximum stability. Simulation results show that this fuzzy controller also adequately sustains desired performance for parameter variations within and outside the design ranges. Cellmapdesigned controllers offer an important feature rarely obtained with other design methodologies - all the initial states that can be stabilized by the designed controller in a time-optimal manner.

H. Kang and G. Vachtsevanos, Non-linear fuzzy control based on the vector field of the phase portrait assignment algorithm, Proceedings of the American Control Conference, pp. 1179-1484, 1990.

C. L. Karr and E. J. Gentry, " F v control of pH using genetic algorithms,' IEEE Transactions on Fuzzy Systems, vol. 1, pp. 46-53, 1993.
M. Lee and H. Takagi, "Integrating design stages of fuzzy systems using genetic algorithms," Proceedings of the Second IEEF International Confrence on Fuzty Systems, pp. 6 12-617, 1993. H. Nomura, I. Hayashi and N. Wakami, "A selftuning method of fuzzy control by descent method," Proceedings o f the 4th IFSA World Congress, vol. Engineering, pp. 155-158, 1991.

ACKNOWLEDGMENT
Ths research is supported by OCAST Grants AR9010 and AR2-002. and by a grant f r o m the Oklahoma Center for Integrated Design and Manufacturing.

REFERENCES
L 11 C. W. Anderson, "Learning to control an inverted
pendulum using neural networks," IEEE Control SystemsMagazine, vol. 10, no. 3, pp. 3 1-36: 1989.
121 H. R. Berenji and P. Khedkar, "Learning and tuning f k z y logic controllers through reinforcements," IEEE Transactions on Neural Networks, vol 7, pp. 724-740, 1992.
I31 Y. Y. Chen and T. C. Two, " A description of the

S. M. Smith and D. J. Comer, "Automated calibration of a fuzzy logic controller using a cell state space algorithm, " IEEE Control Systems Magazine, vol. 11, no. 8, pp. 18-28, 1991.
S . M. Smith and D. J. Comer, "An algorithm for automated fuzzy logic controller tuning," Proceedings o f the First IEEE International Confireme on Fuzzy Systems, pp. 61S-622, 1992.

dynamical behavior of fuzzy systems," lEEE Transactionson Svstems, Man and Cybernetics, vol. 19. pp. 745-755; 1989. 141 D. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, Reading, Massachusetts, 1989.

G. Rudolph, "Convergence analysis of canonical genetic algorithms," IEEE Transactions on Neural Networks, vol. 5, pp 96-101, 1994.
T. Takagi and M. Sugeno, 'Fuzzy identification of systems and its applications to modeling and control," IEEE Transactions on Systems, Man and Cybernefics,vol. 15, pp. 116-132, 1985.

I51 C . S. Hsu and R. Cuttalu. "An unravelling algorithm for global analysis of dynamical systems: An
application of cell-to-cell mappings," ASME Joumal ofApplied hfechanics.vol. 47: pp. 93 1-939, 1980.

I61

C. S . Hsu. "A &%rete method of optimal control based upon the cell state space concept," Journal qf Optimization Theow and Applications. vol. 46. pp.
547-569; 1985.

G. Vachtsevanos, S. S . Farinwata and D. K. Pirovolou, "Fuzzy logic control of an automotive engine," IEEE Control Systems Magazine, vol. 13, no. 3, pp. 62-68, 1993,

B. Widrow and S . Steams, Adaptive Signal Processing, hentice-Hall, Englewood Cliffs, NJ, 1985.

89

Authorized licensed use limited to: Norges Teknisk-Naturvitenskapelige Universitet. Downloaded on December 4, 2009 at 08:30 from IEEE Xplore. Restrictions apply.

50

I . . . . . . . . ..I.. ..................

..........
................................

' U

. :

c,

.............

*;

I.

j .
j

* . :

-50

...............

1
0

......................
05

1 -

0.5

(c)

PolrMglc(nd)

(d)

time (sec)

Figure 4. Time responses for the inverted pendulum system. (a) Force: (bl Pole angle: (c) Phase plane trajectories: (d) Angular velocity. (Solid line I = 0 5 m: Dashdotted line: I = 0.25 m: Dashed line: 1 = 0.125 m.)

(b)

Figure 5 . Controllable regions for different maximum force magnitudes. ((a) 50 newtons: (b) 10 Newtons.)
50

li
Y

.......... i... ........ .....

.......................

&

....... 0
(a)

................... ::v _.-.

05 rim( s e c )

0
(b)

05 timt (sec)

1
(c)

0 pokaIlg*(rad)

0
(d)

0.5
lime (=)

Figure 6. Time responses for the merted pendulum system (a) Force, (b) Pole angle; (c) Phase plane trajectories: (d) Angular velocity (Solid line:./ = 50 Nenlons: Dashdotted line: f = 70 Newtons: Dashed line. f = 30 Newtons.)

90

Authorized licensed use limited to: Norges Teknisk-Naturvitenskapelige Universitet. Downloaded on December 4, 2009 at 08:30 from IEEE Xplore. Restrictions apply.

You might also like