You are on page 1of 10

Contents

1 Introduction 1.1 Why fuzzy control . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Fuzzy Based Techniques . . . . . . . . . . . . . . . . . . . . . . . 2 Fuzzy Systems 2.1 Mamdani fuzzy systems . . . . . . . . . . . . . . . . . . . . . . . 2.2 Takagi-Sugeno fuzzy systems . . . . . . . . . . . . . . . . . . . . 2.3 Comparison between Mamdani & Takagi Sugeno systems . . . . 3 Properties of the Takagi Sugeno Model 3.1 Fuzzy Systems Are Universal Approximators . . . . . . . . . . . 4 Takagi-Sugeno fuzzy controllers 4.1 The Takagi-Sugeno open loop fuzzy model . . . . . . . . . . . . . 4.2 The Takagi-Sugeno fuzzy controller TSFC-1 . . . . . . . . . . . . 4.3 The Takagi-Sugeno closed loop fuzzy model . . . . . . . . . . . . 5 Conclusion 2 2 3 4 4 5 5 6 6 6 7 8 9 10

Takagi Sugeno Model for Fuzzy Control Systems


Ashar Ahmad
Group - Modelling & Simulation

November 28, 2011


Abstract This paper attempts to present the basics of fuzzy modelling principles.Strarting from the notion of fuzzy control as oppsosed to conventional control , we delve more into the two mechanisms for fuzzy inferences viz. Mamdani & Takagi Sugeno exploring the advantages and disadvantages of both. We look at the mathematical formulation of the Mamdani and Takagi Sugeno systems. Then some key properties of the Tagaki Sugeno Model are presented including the Universal Approximator property of fuzzy sytems.Finally we look at the system dynamics of a closed loop Takagi Sugeno Fuzzy controller which has two components - a system modelling part and a controller part , both based on Tagaki Sugeno formulation.

1
1.1

Introduction
Why fuzzy control

What is traditional control? The main objective of traditional control theory is to nd the control that is the best (in the sense of some well-dened criterion). As a result of applying the traditional control theory, we get a control strategy, i.e., an algorithm that transforms the sensor inputs x1 , . . . , xn into the corresponding control values u(x1 , . . . , xn ). Methods of traditional control theory are often not applicable. The methods of traditional control theory are based on the assumption that we know exactly the behavior of the system, i.e., that we know exactly how the system will react to dierent controls. In many real-life control situations, we do not have the exact knowledge of the system that we want to control: for example, when we drive or park a car, we may not know exactly the conditions of the road, the location of other cars, etc. Then, methods of traditional control theory are often not applicable.

In some other situations, we know the systems behavior exactly, but it is so complicated, that no methods have yet been invented for controlling such systems. In such situations, the only information that we can use to automate control is the experience of human controllers. This experience rarely comes in the form of a precise control strategy, it is usually formulated in terms of imprecise rules like if an obstacle is close, slow down a little bit, rules that use (vague) words from natural language. What is fuzzy control. A methodology for translating expert rules into the precise control strategy is called fuzzy control.

1.2

Fuzzy Based Techniques

Lot Zadeh rst introduced fuzzy logic in the 1960s in an attempt to nd a more natural way to handle imprecise information [1]. The theory of fuzzy logic attempts to enable machines to deal with imprecise language used by humans in order to describe data that may not be exact or crisp. A fuzzy system can be represented by the following components,

Figure 1: Fuzzy Control System Fuzzication Fuzzication is the process by which crisp data is converted to non-crisp linguistic terms. Fuzzication attempts to nd the degree matching between the input, xi to a set of linguistic terms dened by A = {Ai,1 , , Aik }. For example, the feature of age may be described as {very young, young, middle-age, old}. Each of these linguistic terms can be represented by a set of values of the input xi . Fuzzication replaces xi with a set of Ki numbers which range from 0 to 1 and reect how well xi matches each linguistic term. The degree of matching is represented by a membership degree dened Ai ,1 (xi ), , Ai ,Ki (xi ) for Ki numbers. Inference An inference system uses a set of fuzzy if-then rules called a rule base. Each rule contains an antecedent part and a consequent part. The antecedent may consist of Boolean expression that contains simple clauses, for example, if x is small. The consequent may be an output represented by linguistic terms, such as then y is big (Mamdani-Assilian model), or 3

a function of the input x, then y = fx (x), (Takagi-Sugeno Kang model) [2]. Defuzzication Defuzzication involves calculating a single value to represent the fuzzy set. This is done in order to nd the systems output, y. There are many methods in which defuzzication occur. One common technique is the centre of gravity method which evaluates y by calculating the centre of gravity of the area under the fuzzy set.

2
2.1

Fuzzy Systems
Mamdani fuzzy systems

Mamdani fuzzy systems, also known as linguistics fuzzy systems[3], are composed by four dierent elements: a fuzzier, a rule base, an inference engine, and a defuzzier. The inputs are transformed in fuzzy numbers by the fuzzier, which converts numeric values x into fuzzy sets A (x). Usually singleton fuzziers are the predominant ones, since their use simplies the computations in the fuzzy system considerably. This means that the inputs of the fuzzy system are real numbers and not fuzzy sets. The fuzzy rule base consists of fuzzy IF-THEN rules and membership functions, characterizing the fuzzy set. More precisely a fuzzy rule base, R, is a set of rules R(l) , l = 1, 2, . . . , K of the form: R(l) : IF x1 IS A1 AND . . . AND xn IS An THEN y IS B (l) where xi are the input (antecedent) variables, and y Y is the output (l) (consequent). Ai and B (l) are linguistic terms (labels) dened by fuzzy sets A(l) (xi ) : [0, 1] and B (l) (y) : Y [0, 1]. The inference engine takes i the input of the fuzzy system, and uses the rule based systems to calculate the output of the system. Each rule is evaluated and the output suggested by the rule is inferred using fuzzy operators. Then the predictions of each rules are aggregated together in a fuzzy set. The defuzzier has the task of taking transforming this fuzzy set into a numeric value which is the output of the system. These type of fuzzy systems are well suited for encoding imprecise knowledge expressed in the form of IF-THEN rules. These rules are easily understood by the user since they do not imply any mathematical representation. The type of interpolation between the rules depends on their shape and and the choice of the fuzzy logic operators and defuzzication procedures . The class denes methods for adding, and removing rules. It is possible to choose the shape of the membership functions by composition of simpler shapes. In this way it is possible to select gaussian, sigmoidal, triangular, and several variations of rectangular membership functions. Freedom about the defuzzication method is also granted. Center of gravity, center average, and maximum defuzziers are implemented in this class.
(l) (l)

2.2

Takagi-Sugeno fuzzy systems

The Takagi-Sugeno fuzzy systems are a dierent type of fuzzy systems which use rules of the form [2]: R(l) : IF x1 IS A1 AND . . . AND xn IS An THEN y = h(l) (x) where xi are the inputs (antecedent) variables, and y is the out(l) put (consequent). Ai are antecedent fuzzy sets of the lth rule, dened by a membership function A(l) (xi ) : [0, 1]
i

(l)

(l)

In many cases the output function h(l) (x) is a linear combination of the input variables plus a constant term. h(l) (x) = a0 + a1 x1 + + a(l) xn n In the special case when all the l1 . . . ln = 0, and therefore the consequents are constant functions, this becomes a special case of the Mandami system with the consequent fuzzy sets being fuzzy singletons. The system output is a weighted average of the individual rule outputs, similar to the fuzzy-mean defuzzication formula: M A(l) (x) h(l) (x) y= M A(k) (x) k=1 l=1 where the weights A(l) (x) are computed according to
n (l) (l)

A(l) (x) =
i=1

A(l) (xi )
i

This approach allows to model a system by means of the decomposition of a nonlinear system into a collection of local linear models. It allows a more accurate representation of systems, since the rules are usually simple linear subsystems, and not constant values like in the case of the Mamdani fuzzy systems. Since this approach imposes to structure the problem in a series of local models, Takagi-Sugeno models can be more easily constructed from numerical data than Mamdani models, which are essentially structure free. The Features of this class are similar to the ones dened in the mamdani class. It is possible to add and delete rules, and to dene their shape. Since these type of models are usually constructed from numerical data several identication procedures are available for performing the parameter identication procedure.

2.3

Comparison between Mamdani & Takagi Sugeno systems

Fixed Fuzzy sets vs Moving singleton sets : Mamdani-style inference, requires us to nd the centroid of a two-dimensional shape by integrating 5

across a continuously varying function. In general, this process is not computationally ecient. Michio Sugeno suggested to use a single spike, a singleton, as the membership function of the rule consequent. A singleton, or more precisely a fuzzy singleton, is a fuzzy set with a membership function that is unity at a single particular point on the universe of discourse and zero everywhere else. Computational eort : Mamdani method is widely accepted for capturing expert knowledge. It allows us to describe the expertise in more intuitive, more human-like manner. However, Mamdani-type fuzzy inference entails a substantial computational burden.On the other hand, Sugeno method is computationally eective especially for multivariableand works well with optimisation and adaptive techniques, which makes it very attractive in control problems, particularly for dynamic nonlinear systems

Properties of the Takagi Sugeno Model

TS models are basically non-linear interpolators of linear models. They can also be viewed as interpolator between linear mappings.Model identication is an important part of modelling systems as TS systems.In their paper [2] Takagi and Sugeno have proposed an algorithm for identication of Input Variables Fuzzy sets to which the input variables belong The parameters of the rule consequents They used the method of least squares to determine the value of the matrix for parameters.

3.1

Fuzzy Systems Are Universal Approximators

Mathematical formulation of universal approximation. In mathematical terms, we want to guarantee that, given an arbitrary function u(x1 , . . . , xn ), and an arbitrary positive real number > 0, we will be able to nd a function u(x1 , . . . , xn ) generated by fuzzy control methodology that is -close to the original function u(x1 , . . . , xn ). In mathematics, this ability of approximate an arbitrary function with an arbitrary accuracy is called a universal approximation property. Fuzzy control methodology is indeed a universal approximator and Tagaki Sugeno models have the universal approximaton property [4].

Takagi-Sugeno fuzzy controllers

The model based design of Takagi-Sugeno fuzzy controllers (TSFC-1) requires a fuzzy model of the nonlinear autonomous open loop system under control. Such

a fuzzy model was rst proposed in 1985 by Takagi and Sugeno [2] and is the result of identication from observed input-output data. The TSFC-1 is then intended to solve the stabilization problem in the context of a given (identied) fuzzy model. Relevant publication we used for the fuzzy model based design and analysis of the TSFC-1 is Tanaka and Sugeno [5].

4.1

The Takagi-Sugeno open loop fuzzy model

The open loop fuzzy model consists of a set of fuzzy rules where the IF-part of each fuzzy rule describes a particular fuzzy region in the state space and its corresponding THEN-part contains a linear open loop model. Such a fuzzy rule is written as IF where
i i i LXi = (LX1 , . . . , LXn )T , where LXk denotes the fuzzy value which xk i takes in the i-th fuzzy region. Each LXk is determined by a fuzzy set i LXk (xk )/xk of a standard triangular, trapezoidal, or bell-shaped type. X The membership functions of any one of the previously mentioned types are nonlinear functions of xk ,

x = LXi

THEN

x = Ai x + Bi u,

(1)

x = Ai x + Bi u is a linear autonomous model corresponding to the i-th fuzzy region of the state space. The state vector x is a n 1 linear vector function of time and its components x1 , x2 , , xn denote time dependent state variables. The control input vector u is a m 1 vector function of time with components u1 , u2 , , un . The entries of the matrices Ai (n n) and Bi (n m) are constant. The open loop systems dynamics corresponding to computation with a single fuzzy rule is given as x = LXi (x) (Ai x + Bi u). (2)

The results computed for each individual rule via Eq. (2) are aggregated by simply taking their average, that is, LXi (x) (Ai x + Bi u) x=
i

LXi (x)
i

(3)

Eq. (3) can be simplied by normalizing the degrees of satisfaction LXi (x) and using instead their normalized counterparts wi (x) that are obtained as follows: for any LX1 (x), LX2 (x), . . . , LXM (x) wi (x) = LXi (x) . LXi (x)
i

(4)

Thus,

wi (x) = 1 and Eq. (3) can be rewritten in the more simple form, x=
i

wi (x) (Ai x + Bi u).

(5)

From Eq. (5) it is easily seen that the set of all fuzzy rules denes linear dynamics for all points that belong to the center of an arbitrary fuzzy region in the fuzzy state space. Take, for example, the i-th fuzzy region. For every point that belongs to its center, i.e., x = xi , we have that wi (x) = wi (xi ) = 1 while for each j = i we have that wj (x) = 0, since i wi (x) = 1. Thus we have that Eq. (5) becomes (for each x = xi ), x = wi (xi ) (Ai xi + Bi u) = Ai xi + Bi u. On the other hand, Eq. (5) denes nonlinear dynamics for all points x = xi . This is so, because in this case there is no wi (x) = 1, and thus each linear part (Ai xi + Bi u) in Eq. (5) is multiplied by the nonlinear function wi (x).

4.2

The Takagi-Sugeno fuzzy controller TSFC-1

The TSFC-1 is used for the stabilization problem in the context of the fuzzy model given by Eq. (5). The TSFC-1 for the stabilization problem is given as a set of fuzzy rules where each fuzzy rule is of the form IF where
j j j LXj = (LX1 , . . . , LXn )T , where LXk denotes the fuzzy value which xk j takes in the j-th fuzzy region. Each LXk is determined by a fuzzy set j LXk (xk )/xk of a standard triangular, trapezoidal, or bell-shaped type. X

x = LXj

THEN u = Kj x,

(6)

u = Kj x is a linear autonomous control law corresponding to the j-th fuzzy region of the state space. The matrix Kj is of dimension m n , its elements are constant, and it is called the gain matrix of the j-th fuzzy region. The control laws dynamics corresponding to computation with a single fuzzy rule is given as u = LXj (x) Kj x. (7)

The results computed for each individual rule via Eq. (7) are aggregated by simply taking their average, that is, LXj (x) Kj x u=
j

LXj (x)
j

(8)

Eq. (8) can again be simplied by normalizing the degrees of satisfaction LXj (x) and using instead their normalized counterparts wj (x). that are obtained according to Eq. (4). Since j wj (x) = 1, Eq. (8) can be rewritten in the simpler form u=
j

wj (x) Kj x.

(9)

From Eq. (9) it is easily seen that the set of all fuzzy rules denes a control law with linear dynamics for all points that belong to the center of an arbitrary fuzzy region in the fuzzy state space. Take, for example, the j-th fuzzy region. For every point that belongs to its center,i.e., x = xj , we have that wj (x) = wj (xj ) = 1 while for each i = j we have that wi (x) = 0, since j wj (x) = 1. Thus we have that Eq. (9) becomes (for each x = xj ), u = wj (xj ) Kj xj = Kj xj . On the other hand, Eq. (9) denes a control law with nonlinear dynamics for all points x = xj . This is so, because in this case there is no wj (x) = 1, and thus each linear expression Kj xj in Eq. (9) is multiplied by the nonlinear function wj (x).

4.3

The Takagi-Sugeno closed loop fuzzy model

So far, by using the fuzzy rules of the form IF and IF x = LXi THEN u = Ki x, we have described the nonlinear autonomous (time-invariant) open loop fuzzy model and the nonlinear TSFC-1. Furthermore, we saw that computation with the open loop fuzzy model is equivalent to x=
i

x = LXi

THEN

x = Ai x + Bi u,

wi (x) (Ai x + Bi u),

and computation with the TSFC-1 is equivalent to u=


j

wj (x) Kj x.

Thus, replacing u from the former equation with its equivalent term from the latter equation, we obtain the expression for the nonlinear autonomous closed loop system: x=
i j

wi (x) wj (x) (Ai + Bi Kj ) x,

(10)

Denoting, (Ai + Bi Kj ) by Aij , we can write x=


i j

wi (x) wj (x) Aij x.

(11)

From the above equation it is easily seen that the dynamics of the autonomous closed loop system is nonlinear since both wi (x) and wj (x) are nonlinear functions of the state vector x. The dynamics of the closed loop autonomous system is again linear only for points in the fuzzy state space that belong to the center of an arbitrary fuzzy region.

Conclusion

This article was just a brief introduction to the fuzzy control systems and in particular Tagaki Sugeno formulation. A lot of research work has been done in this eld , from designing Parallel Distribution Compensation(PDC) controllers to elaborate studies on the stability of the such controllers. Our aim was to formulate the basics and describe the system dynamics of a Takagi Sugeno Fuzzy controller.

References
[1] L. Zadeh. Fuzzy Sets, Information and Control, Vol. 8, pp. 338353, 1965. [2] T. Takagi and M. Sugeno. Fuzzy identication of systems and its application to modeling and control. IEEE Trans. Systems, Man, and Cybernetics, 15:116132, 1985. [3] Mamdani, E. H. and S. Assilian, An experiment in linguistic synthesis with a fuzzy logic controller, Int. J. Man-machine Studies, Vol. 7, 113, 1975. [4] J. J. Buckley, Universal Fuzzy Controllers, Automatica, 1992, Vol. 28, pp. 1245-1248. [5] K. Tanaka and M. Sugeno. Stability analysis and design of fuzzy control systems. Fuzzy Sets and Systems, 45:135156, 1992.

10

You might also like