2008 American Control Conference FrB17.

2
Westin Seattle Hotel, Seattle, Washington, USA
June 11-13, 2008

An SOS-based Stable Control of Polynomial Discrete Fuzzy Systems

Kazuo Tanaka, Hiroshi Ohtake and Hua O. Wang

Abstract— This paper presents stability and stabilization nomial Lyapunov functions, we derive a stability condition
conditions of polynomial discrete fuzzy systems. First, based on of polynomial discrete fuzzy systems without inputs. Sec-
polynomial Lyapunov functions, we derive a stability condition ondly, we derive a stabilization condition to design a stable
of polynomial discrete fuzzy systems without inputs. Secondly,
we derive a stabilization condition to design a stable polynomial polynomial fuzzy controller. Both of them are represented in
fuzzy controller. Both of them are represented in terms of terms of SOS and are numerically (partially symbolically)
SOS and are numerically (partially symbolically) solved via the solved via the recent developed SOSTOOLS [7]. In addition,
recent developed SOSTOOLS. In addition, polynomial discrete polynomial discrete fuzzy systems and controllers contain
fuzzy systems and controllers contain discrete Takagi-Sugeno discrete Takagi-Sugeno (T-S) fuzzy systems and controllers
(T-S) fuzzy systems and controllers as a special case. Hence,
the approach discussed in this paper is more general than that as a special case. Hence, the approach discussed in this
based on the existing LMI approaches to discrete T-S fuzzy paper is more general than that based on the existing LMI
control system designs. To illustrate the validity of the design approaches to discrete T-S fuzzy control system designs.
approach, a design example is provided. The example shows To illustrate the validity of the design approach, a design
the utility of our approach. example is provided. The example shows the utility of our
approach.
I. INTRODUCTION
The Takagi-Sugeno (T-S) fuzzy model-based control II. D ISCRETE TAKAGI -S UGENO FUZZY CONTROL
methodology [1] provides a natural, simple and effective SYSTEM AND STABILITY CONDITION BASED ON
design approach to complement other nonlinear control tech- QUADRATIC LYAPUNOV FUNCTIONS
niques (e.g., [2]) that require special and rather involved
knowledge. Moreover, there is no loss of generality in adopt- Section II summarizes discrete Takagi-Sugeno fuzzy con-
ing the T-S fuzzy model based control design framework as trol system and stability condition based on quadratic Lya-
it has been established that any smooth nonlinear control punov functions. The discrete Takagi-Sugeno fuzzy model
systems can be approximated by the T-S fuzzy models (with [8] is described by fuzzy IF-THEN rules which represent
liner model consequence) [3]. Within the general framework local linear input-output relations of a nonlinear system. The
of T-S fuzzy model-based control systems, there has been, main feature of this model is to express the local dynamics
in particular, a flurry of research activities in the analysis of each fuzzy implication (rule) by a linear system model.
and design of fuzzy control systems based on linear matrix The overall fuzzy model of the system is achieved by fuzzy
inequalities (LMIs) (e.g., [1]). Hence, this approach have blending of the linear system models.
received a great deal of attention over the last decade. Consider the following nonlinear system:
Recently, we presented a sum of squares (SOS) approach
to stability analysis [4] and stable controller design [5] x(t + 1) = f (x(t), u(t)), (1)
for polynomial continuous fuzzy systems. In [6], we dealt
with guaranteed cost control of polynomial continuous fuzzy where f is a nonlinear function. x(t) =
systems. These are completely different approaches from the [x1 (t) x2 (t) · · · xn (t)]T is the state vector and
existing LMI approaches. To the best of our knowledge, u(t) = [u1(t) u2 (t) · · · um (t)]T is the input vector.
the paper [4] presented the first attempt at applying an Based on the sector nonlinearity concept [1], we can exactly
SOS to fuzzy systems. Our SOS approaches provide more represent (1) with the discrete Takagi-Sugeno fuzzy model
extensive results for the existing LMI approaches to T-S (2) (globally or at least semi-globally).
fuzzy systems. However, SOS-based design approaches of Model Rule i:
polynomial discrete fuzzy systems have not been reported in
the literature. If z1 (t) is Mi1 and · · · and zp (t) is Mip
This paper presents stability and stabilization conditions then x(t + 1) = Ai x(t) + B i u(t) i = 1, 2, · · · , r, (2)
of polynomial discrete fuzzy systems. First, based on poly-
where zj (t) (j = 1, 2, · · · , p) is the premise variable. The
This work was supported in part by a Grant-in-Aid for Scientific Research
(C) 18560244 from the Ministry of Education, Science and Culture of Japan. membership function associated with the ith Model Rule
K. Tanaka and H. Ohtake are with the Department of Mechanical Systems and jth premise variable component is denoted by Mij . r
and Intelligent Systems, The University of Electro-Communications, Chofu, denotes the number of Model Rules. Each zj (t) is a mea-
Tokyo 182-8585 Japan [email protected]
H. O. Wang is with the Department of Aerospace and Mechanical Engi- surable time-varying quantity that may be states, measurable
neering, Boston University, Boston, MA 02215 USA [email protected] external variables and/or time. The defuzzification process

978-1-4244-2079-7/08/$25.00 ©2008 AACC. 4875

Authorized licensed use limited to: National Kaohsiung Univ. of Science & Technology. Downloaded on May 01,2025 at 09:31:34 UTC from IEEE Xplore. Restrictions apply.
 of the model (2) can be represented as III. P OLYNOMIAL D ISCRETE F UZZY M ODEL AND
r P OLYNOMIAL LYAPUNOV FUNCTION


wi (z(t)){Ai x(t) + B i u(t)} Section II summarized discrete Takagi-Sugeno fuzzy
i=1 model and stability analysis based on quadratic Lyapunov
x(t + 1) = r

functions. In this section, we propose a new type of discrete
wi (z(t)) fuzzy model with polynomial model consequence, i.e., fuzzy
i=1
r
model whose consequent parts are represented by polynomi-


= hi (z(t)){Ai x(t) + B i u(t)}, (3) als.
i=1
As shown in Section II, the stability conditions (6) -
(8) for the T-S fuzzy system and the quadratic Lyapunov
where
p z(t) = [z1 (t) · · · zp (t)] and wi (z(t)) = function reduce to LMIs. Hence, the stability conditions
j=1 M ij (zj (t)). It should be noted from the properties of can be solved numerically and efficiently by interior point
membership functions that the following relations hold. algorithms such as the LMI toolbox of MATLAB1 . On


r the other hand, stability conditions for polynomial fuzzy
wi (z(t)) > 0, wi (z(t)) ≥ 0 i = 1, 2, · · · , r systems and polynomial Lyapunov functions reduce to an
i=1 SOS problem. Clearly, the problem is never solved by the
Hence, LMI toolbox and can be solved via SOSTOOLS [7].
SOSTOOLS [7] is a free, third party MATLAB toolbox
wi (z(t)) for solving sum of squares problems. The techniques behind
hi (z(t)) = r ≥ 0,

it are based on the sum of squares decomposition for
wi (z(t))
multivariate polynomials, which can be efficiently computed
i=1
using semidefinite programming. SOSTOOLS is developed
and as a consequence of the recent interest in sum of squares
r

polynomials, partly due to the fact that these techniques
hi (z(t)) = 1. provide convex relaxations for many hard problems such
i=1 as global, constrained, and boolean optimization. For more
The parallel distributed compensation (PDC) offers a details, see the manual of SOSTOOLS [7].
procedure to design a fuzzy controller from the given Takagi- A. Polynomial discrete fuzzy model
Sugeno fuzzy model (2):
A polynomial continuous fuzzy model has been proposed
Control Rule i:
in [4]. In this section, we propose a polynomial discrete fuzzy
If z1 (t) is Mi1 and · · · and zp (t) is Mip model. Using the sector nonlinearity concept, we exactly
then u(t) = −F i x(t) i = 1, 2, · · · , r (4) represent (1) with the following polynomial fuzzy model (9).
The main difference between (2) and (9) is consequent part
The overall fuzzy controller can be calculated by representation. The fuzzy models of (2) and (9) have linear
r model consequence and polynomial model consequence,


u(t) = − hi (z(t))F i x(t). (5) respectively.
i=1
Model Rule i:
By employing the quadratic Lyapunov function
xT (t)X −1 x(t), the well-known LMI condition [1] for If z1 (t) is Mi1 and · · · and zp (t) is Mip
the stability of the feedback system consisting of (3) and then x(t + 1) = Ai (x(t))x̂(x(t)) + B i (x(t))u(t), (9)
(5) is given as follows;
where i = 1, 2, · · · , r. Ai (x(t)) and Bi (x(t)) are polyno-
X >0 (6) mial matrices in x(t). x̂(x(t)) is a column vector whose
entries are all monomials in x(t). That is, x̂(x(t)) ∈ RN is
  an N × 1 vector of monomials in x(t). A monomial in x(t)
X XATi − M Ti B Ti
> 0, (7) is a function of the form xα 1 α2 αn
1 x2 · · · xn , where α1 , α2 , · · · ,
Ai X − B i M i X
αn are nonnegative integers. Therefore, Ai (x(t))x̂(x(t)) +
⎡ ⎤ B i (x(t))u(t) is a polynomial vector. Thus, the polynomial
 X  ∗ fuzzy model (9) has a polynomial in each consequent part. In
⎣ Ai X + Aj X − B i M j − B j M i ⎦ ≥ 0,
X Sections IV and V, we will consider T (x̃(t)) ∈ RN ×m that
2 is a polynomial matrix defined by x̂(x(t)) = T (x̃(t))x(t).
i < j (8) The details of x̂(x(t)) will be given in Proposition 1. The
definition of x̃(t) will be also presented later.
where M i = F i X. The conditions (6)-(8) are represented
We assume that
in terms of LMIs. Thus, the stability conditions can be solved
numerically and efficiently by interior point algorithms. 1A registered trademark of MathWorks, Inc.

4876
Authorized licensed use limited to: National Kaohsiung Univ. of Science & Technology. Downloaded on May 01,2025 at 09:31:34 UTC from IEEE Xplore. Restrictions apply.
 x̂(x(t)) = 0 iff x(t) = 0
 
throughout this paper. 1 d
A1 = ,
The defuzzification process of the model (9) can be −4 2
 π 
represented as 1 d
A2 = ,
−1 2
r

 
x(t + 1) = hi (z(t)){Ai (x(t))x̂(x(t)) 1 −d
A3 = ,
i=1 − π4 2
 
+B i (x(t))u(t)}. (10) 1 −d
A4 = ,
−1 2
 
Thus, the overall fuzzy model is achieved by fuzzy blending 1
of the polynomial system models. B1 = ,
0
If x̂(x(t)) = x(t) and Ai (x(t)) and Bi (x(t)) are con-  
1
stant matrices for all i, then Ai (x(t))x̂(x(t))+B i (x(t))u(t) B2 = .
0
reduces to Ai x(t)+B i u(t), that is, then (10) reduces to (3).
Therefore, (10) is a more general representation. The membership functions are obtained as
x2 (t) + d tan x1 (t) − x1 (t)
h1 (z(t)) = ,
Remark 1: As we can see in Section III-C, the number of 2d ( π4 − 1)x1 (t)
rules in polynomial fuzzy model generally becomes fewer
x2 (t) + d π4 x1 (t) − tan x1 (t)
than that in T-S fuzzy model, and our SOS approach to h2 (z(t)) = ,
polynomial fuzzy models provides much more relaxed sta-
2d ( π4 − 1)x1 (t)
bility results than the existing LMI approaches to T-S fuzzy d − x2 (t) tan x1 (t) − x1 (t)
h3 (z(t)) = ,
models. 2d ( π4 − 1)x1 (t)
d − x2 (t) π4 x1 (t) − tan x1 (t)
h4 (z(t)) = .
B. Polynomial Lyapunov function 2d ( π4 − 1)x1 (t)
We can see that four rules are needed to represent the
To obtain more relaxed stability results, we employ a
dynamics.
polynomial Lyapunov function [4] represented by
On the other hand, we have the following polynomial
fuzzy model that can exactly represent the dynamics under
x̂T (x(t))P (x(t))x̂(x(t)), (11) −π/4 ≤ x1 (t) ≤ π/4.
2


where P (x(t)) is a polynomial matrix in x(t). If x̂(t) = x(t + 1) = hi (z(t)){Ai (x(t))x(t) + Bi (x(t))u(t)}
x(t) and P (x(t)) is a constant matrix, then (11) reduces i=1
to the quadratic Lyapunov function xT (t)P x(t). Therefore, (14)
(11) is a more general representation. T
where x̂(t) = x(t) = x1 (t) x2 (t) , z(t) = x1 (t) and
 
C. Fuzzy Modeling Example 1 x2 (t)
A1 (x(t)) = ,
− π4 2
Consider the following nonlinear system:  
1 x2 (t)
A2 (x(t)) = ,
−1 2
x1 (t + 1) = x1 (t) + x22 (t) + u(t),  
(12) 1
x2 (t + 1) = − tan x1 (t) + 2x2 (t), B1 (x(t)) = B1 = ,
0
 
1
where we assume that −π/4 ≤ x1 (t) ≤ π/4. B2 (x(t)) = B2 = .
0
To obtain a T-S fuzzy model, we also assume that −d ≤
The membership functions are obtained as
x2 (t) ≤ d. Using the concept of sector nonlinearity, we
have the T-S fuzzy model (13) that can exactly represent the tan x1 (t) − x1 (t)
h1 (z(t)) = ,
dynamics under −π/4 ≤ x1 (t) ≤ π/4 and −d ≤ x2 (t) ≤ d. ( π4 − 1)x1 (t)
4
x1 (t) − tan x1 (t)


4 h2 (z(t)) = π 4 .
x(t + 1) = hi (z(t)){Ai x(t) + Bi u(t)} (13) ( π − 1)x1 (t)
i=1 Note that the assumption that −d ≤ x2 (t) ≤ d is not
needed in the construction of the polynomial fuzzy models.
T T
where x(t) = x1 (t) x2 (t) , z(t) = x1 (t) x2 (t) , In addition, the number of rules in the polynomial fuzzy
and model is fewer than that in the T-S fuzzy model.

4877
Authorized licensed use limited to: National Kaohsiung Univ. of Science & Technology. Downloaded on May 01,2025 at 09:31:34 UTC from IEEE Xplore. Restrictions apply.
 IV. SOS S TABILITY C ONDITION (xk1 , xk2 , · · · xkm ). From the definition, a partial system of
(10) can be represented as
A. Sum of Squares
The computational method used in this paper relies on x̃(t + 1) = Ã(x)x, (16)
the sum of squares decomposition of multivariate polyno-
mials. A multivariate polynomial f(x(t)) (where x(t) ∈ where Ã(x) is a polynomial matrix. The equation (16) will
Rn ) is a sum of squares (SOS, for brevity) if there exist play an important role in Theorems 1 and 2.
polynomials
m 2 f1 (x(t)), · · · , fm (x(t)) such that f(x(t)) =
f (x(t)). It is clear that f(x(t)) being an SOS nat- Theorem 1: The zero equilibrium of the system (10) with
i=1 i
urally implies f(x(t)) > 0 for all x(t) ∈ Rn . This can be u(t) = 0 is stable if there exists a symmetric polynomial
shown equivalent to the existence of a special quadric form matrix P (x̃) ∈ RN ×N such that (17) and (18) are satisfied,
stated in the following proposition [10]. where 1 (x) and 2i (x) are non negative polynomials such
that 1 (x) > 0 (x = 0) and 2i (x) ≥ 0 for all x.
Proposition 1: [11] Let f(x(t)) be a polynomial in
x̂T (x)(P (x̃) − 1 (x)I)x̂(x) is SOS (17)
x(t) ∈ Rn of degree 2d. In addition, let x̂(x(t)) be a column
vector whose entries are all monomials in x(t) with degree 
no greater than d. Then f(x(t)) is a sum of squares iff there −x̂T (x) ATi (x)T T (Ã(x)x)P (x̃)T (Ã(x)x)Ai (x)
exists a positive semidefinite matrix P such that 
−P (x̃) + 2i (x)I x̂(x)is SOS ∀i, (18)
f(x(t)) = x̂T (x(t))P x̂(x(t)). (15)

Expressing an SOS polynomial using a quadratic form as in where T (Ã(x)x) ∈ RN ×n is a polynomial matrix given by
(15) has also been referred to as the Gram matrix method.
T (Ã(x)x) = T (x̃(t + 1)). (19)
As mentioned before, a monomial in x(t) is a function
In addition, if (18) holds with 2i (x) > 0 for x = 0, then
of the form xα 1 α2 αn
1 x2 · · · xn , where α1 , α2 , · · · , αn are the zero equilibrium is asymptotically stable. If P (x̃) is a
nonnegative integers. In this case, the degree of the monomial
constant matrix, then the stability holds globally.
is given by α1 + α2 + · · · + αn .
Proof: The proof is omitted due to lack of space.
A sum of squares decomposition for f(x(t)) can be
computed using semidefinite programming, since it amounts
Remark 2: Selection of the non-negative polynomials
to searching for an element P in the intersection of the cone
1 (x) and 2i (x) such that 1 (x) > 0 (x = 0) and 2i (x) ≥
of positive semidefinite matrices and a set defined by some
0 for all x influence the feasibility of the SOS problem.
affine constraints that arise from (15). Note in particular that
Hence, the polynomial structure of 1 (x) and 2i (x) is
the polynomial f(x(t)) is globally nonnegative if it can be
needed to select carefully.
decomposed as a sum of squares. Hence the sum of squares
decomposition in conjunction with semidefinite program-
Remark 3: When Ai (x) and P (x̃) are constant matrices
ming provides a polynomial-time computational relaxation
and x̂(x) = x, the system representation is the same as the
for proving global nonnegativity of multivariate polynomials
Takagi-Sugeno fuzzy model and stability analysis based on
[11], [12], which belongs to the class of NP-hard problems.
quadratic Lyapunov function used in many of the references,
Even though the sum of squares condition is not necessary
e.g., [1]. Thus, our SOS approach to polynomial fuzzy
for nonnegativity, numerical experiments seem to indicate
models contains the existing LMI approaches to Takagi-
that the gap between sum of squares and nonnegativity is
Sugeno fuzzy models as a special case. Therefore, our SOS
small [10].
approach to polynomial fuzzy models provides much more
relaxed stability results than the existing LMI approaches to
B. SOS Stability Condition
Takagi-Sugeno fuzzy models.
To lighten the notation, only this subsection will drop the
notation with respect to time t. For instance, we will employ V. S TABLE C ONTROLLER D ESIGN VIA SOS
x, x̂ instead of x(t), x̂(t), respectively. Thus, we drop the
notation with respect to time t, but it should be kept in mind A. Polynomial Fuzzy Controller
that x means x(t). However, to distinguish between x(t) Since the PDC mirrors the structure of the fuzzy model
and x(t + 1), we will remain the notation with respect to of a system, a fuzzy controller with polynomial rule conse-
time t + 1. Hence, we will employ x(t + 1), x̃(t + 1), etc. quence can be constructed from the given fuzzy model (9).
K = {k1 , k2 , · · · , km } denote the row indices of Bi (x)
whose corresponding row is equal to zero and the row Control Rule i:
indices of Ai (x) whose corresponding row does not contain
non-polynomial nonlinear terms (e.g. trigonometric func- If z1 (t) is Mi1 and · · · and zp (t) is Mip
tions). Using k1 , k2 , · · · , km−1 and km , we define x̃ = then u(t) = −F i (x(t))x̂(x(t)) i = 1, 2, · · · , r (20)

4878
Authorized licensed use limited to: National Kaohsiung Univ. of Science & Technology. Downloaded on May 01,2025 at 09:31:34 UTC from IEEE Xplore. Restrictions apply.
 The overall fuzzy controller can be calculated by
v1T (X(x̃) − 1 (x)I)v1 is SOS (23)
r

 
u(t) = − hi (z(t))F i (x(t))x̂(x(t)). X( x̃) − 2ii (x)I ∗
(21) v2T v2
i=1 T (Ã(x)x)Ωii (x) X(Ã(x)x)
is SOS, (24)

From (10) and (21), the closed-loop system can be repre- X( x̃) − 2ij (x)I
v3T 1
sented as 2 T (Ã(x)x) (Ωij (x) + Ωji (x))

∗
r


r v3 is SOS, i < j, (25)
X(Ã(x)x)
x(t + 1) = hi (z(t))hj (z(t)) where Ωij (x) = Ai (x)X(x̃) − Bi (x)Mj (x). * denotes
i=1 j=1
the transposed elements (matrices) for symmetric positions.
×{Ai (x(t)) − Bi (x(t))Fj (x(t))}x̂(x(t)). (22) v1 ∈ RN , v2 , v3 ∈ R2N are vectors that are independent of
x. In addition, if (24) and (25) hold with 2ij (x) > 0 for
If x̂(x(t)) = x(t) and Ai (x(t)), Bi (x(t)) and Fj (x(t)) are x = 0, then the zero equilibrium is asymptotically stable. If
constant matrices for all i, then (10) and (21) reduce to (3) X(x̃) is a constant matrix, then the stability holds globally.
and (5), respectively. Therefore, (10) and (21) are a more A stabilizing feedback gain Fi (x) can be obtained from
general representation. X(x̃) and Mi (x) as
We provide another important proposition with respect to Fi (x) = Mi (x)X −1 (x̃). (26)
the relaxation.
Proof: The proof is omitted due to lack of space.
Proposition 2: [10] Let L(x(t)) be an N × N symmetric
polynomial matrix of degree 2d in x(t) ∈ Rn . Furthermore, Remark 4: Note that v1 , v2 and v3 are vectors that are
let x̂(x(t)) be a column vector whose entries are all mono- independent of x, because L(x) is not always a positive
mials in x(t) with degree no greater than d, and consider semi-definite matrix for all x even if x̂T (x)L(x)x̂(x) is
the following conditions. an SOS, where L(x) is a symmetric polynomial matrix in
x(t). However, it is guaranteed from Proposition 2 that if
(1) L(x(t)) ≥ 0 for all x(t) ∈ Rn . v T L(x)v is an SOS, then L(x) ≥ 0 for all x.
(2) v T (t)L(x(t))v(t) is a sum of squares, where v(t) ∈
RN . Remark 5: Selection of the non-negative polynomials
(3) There exists a positive semidefinite matrix Q such 1 (x) and 2ij (x) such that 1 (x) > 0 (x = 0) and
that v T (t)L(x(t))v(t) = (v(t) ⊗ x̂(x(t)))T Q(v ⊗ 2ij (x) ≥ 0 for all x influence the feasibility of the SOS
x̂(x(t))), where ⊗ denotes the Kronecker product. problem. Hence, the polynomial structure of 1 (x) and
2ij (x) is needed to select carefully.
Then (1) ⇐ (2) and (2) ⇐⇒ (3).
Remark 6: To avoid introducing non-convex condition,
we assume that X(x̃) only depends on states x̃ whose cor-
responding row in Ai (x) does not contain non-polynomial
B. SOS Stabilization Condition nonlinear terms, and states whose dynamics is not directly
affected by the control input, namely states whose corre-
This subsection gives a stable control design condition rep- sponding rows in Bi (x) are zero.
resented in terms of SOS. Hence the stable fuzzy controller
design with polynomial rule consequence is numerically a Remark 7: When Ai (x), Bi (x), Fi (x) and X(x̃) are
feasibility problem via SOSTOOLS. constant matrices and x̂(x) = x, the system representation
To lighten the notation, this subsection will also drop the is basically the same as the Takagi-Sugeno fuzzy model and
notation with respect to time t. For instance, we will employ control used in many of the references, e.g., [1]. Thus, our
x, x̂(x) instead of x(t), x̂(x(t)), respectively. Thus, we drop SOS approach to fuzzy model and control with polynomial
the notation with respect to time t, but it should be kept in rule consequence contains the existing LMI approaches to
mind that x means x(t). Takagi-Sugeno fuzzy model and control as a special case.
Therefore, our SOS approach provides much more relaxed
Theorem 2: The control system consisting of (10) and results than the existing approaches to Takagi-Sugeno fuzzy
(21) is stable if there exist a symmetric polynomial matrix model and control.
X(x̃) ∈ RN ×N and a polynomial matrix Mi (x) ∈ Rm×N
such that (23), (24) and (25) are satisfied, where 1 (x) and VI. D ESIGN E XAMPLE
2ij (x) are non negative polynomials such that 1 (x) > 0 To illustrate the validity of the design approach, this
(x = 0) and 2ij (x) ≥ 0 for all x. section provides a design example.

4879
Authorized licensed use limited to: National Kaohsiung Univ. of Science & Technology. Downloaded on May 01,2025 at 09:31:34 UTC from IEEE Xplore. Restrictions apply.
 −4
Recall the following nonlinear system as shown in Section x 10

III-C:
x1 (t + 1) = x1 (t) + x22 (t) + u(t) 3
(27)
x2 (t + 1) = − tan x1 (t) + 2x2 (t),

V(x)
where we assume that −π/4 ≤ x1 (t) ≤ π/4. 2

As mentioned in Sectiob III-C, we have the following

polynomial fuzzy model that can exactly represent the dy- 1
namics under −π/4 ≤ x1 (t) ≤ π/4.
2

0
x(t + 1) = hi (z(t)){Ai (x(t))x(t) + Bi (x(t))u(t)}. 0 10 20
t
30 40

i=1

The SOS conditions are infeasible if X(x̃) and Mi (x) are Fig. 2. Time Transient of Lyapunov function.
constant marices in Theorem 2. However, the SOS conditions
are feasible if we select Mi (x) as a polynomial matrix. Thus,
the polynomial fuzzy controller (21) with the polynomial Guaranteed cost control for polynomial discrete fuzzy
feedback vectors Fi (x) is more useful than the T-S fuzzy systems will be presented in [13].
controller (5) with the constant feedback vectors Fi . This
shows the utility of our SOS-based design approach. VIII. ACKNOWLEDGMENTS
Fig. 1 shows the control result of the polynomial fuzzy The authors would like to thank Mr. K. Yamauchi, UEC,
controller (21) for the initial states [0.5 0]T . Fig. 2 shows Japan, for his contribution for this research.
time transient of Lyapunov function during the control. The
R EFERENCES
designed polynomial fuzzy controller stabilizes the system.
[1] K. Tanaka and H. O. Wang: Fuzzy Control Systems Design and
Analysis: An Linear Matrix Inequality Approach, JOHN WILEY &
0.5 SONS, INC, 2001
[2] R. Sepulcher, M. Jankovic and P. Kokotovic: Constructive Nonlinear
0 Control, Springer, 1997
[3] H. O. Wang, J. Li, D. Niemann and K. Tanaka, “T-S fuzzy Model
x1

−0.5 with Linear Rule Consequence and PDC Controller: A Universal

Framework for Nonlinear Control Systems”, 9th IEEE International
−1 Conference on Fuzzy Systems, San Antonio, May, 2000, pp.549-554.
0 10 20 30 40 [4] K. Tanaka, H. Yoshida, H. Ohtake and H. O. Wang “A Sum of Squares
t Approach to Stability Analysis of Polynomial Fuzzy Systems”, 2007
0 American Control Conference, pp.4071-4076, New York, July, 2007.
[5] K. Tanaka, H. Yoshida, H. Ohtake and H. O. Wang, “ Stabilization of
Polynomial Fuzzy Systems via a Sum of Squares Approach”, 2007
x2

−0.5 IEEE International Sympoisum on Intelligent Control, pp.160-165,

Singapore, October, 2007.
[6] K. Tanaka, K. Yamauchi, H. Ohtake and H. O. Wang, “ Guaranteed
−1 Cost Control of Polynomial Fuzzy Systems via a Sum of Squares
0 10 20 30 40
t Approach, 2007 IEEE Conference on Decision and Control, pp.5954-
5959, New Orleans, Dec., 2007.
[7] S. Prajna, A. Papachristodoulou, P. Seiler and P. A. Parrilo: SOS-
Fig. 1. Control result. TOOLS:Sum of Squares Optimization Toolbox for MATLAB, Version
2.00, 2004.
[8] T. Takagi and M. Sugeno, “Fuzzy Identification of Systems and Its
Applications to Modeling and Control”, IEEE Trans. on SMC 15, no.
VII. C ONCLUSIONS 1, pp.116-132, 1985.
[9] L. Xie, S. Shishkin and M. Fu, “Piecewise Lyapunov Functions for
This paper has presented stability and stabilization condi- Robust Stability of Linear Time-Varying Systems”, Systems & Control
tions of polynomial discrete fuzzy systems. First, based on Letters 31 pp.165-171, 1997.
[10] S. Prajna, A. Papachristodoulou and F. Wu: “Nonlinear Control
polynomial Lyapunov functions, we have derived a stability Synthesis by Sum of Squares Optimization: A Lyapunov-based Ap-
condition of polynomial discrete fuzzy systems without in- proach”, Proceedings of the Asian Control Conference (ASCC), Mel-
puts. Secondly, we have derived a stabilization condition to bourne, Australia, Feb. 2004, pp.157-165.
[11] P. A. Parrilo: “Structured Semidefinite Programs and Semialgebraic
design a stable polynomial fuzzy controller. Both of them are Geometly Methods in Robustness and Optimization”, PhD thesis,
represented in terms of SOS and are numerically (partially California Institute of Technology, Pasadena, CA, 2000.
symbolically) solved via the recent developed SOSTOOLS. [12] N. Z. Skor, “Class of global minimum bounds of polynomial func-
tions”, Cybernetics, No.23, Vol.6, pp.731-734, 1987.
The approach discussed in this paper is more general than [13] K. Tanaka, H. Ohtake and H. O. Wang: “A Sum of Squares Approach
that based on the existing LMI approaches to discrete T-S to Guaranteed Cost Control of Polynomial Discrete Fuzzy Systems”,
fuzzy control system designs. To illustrate the validity of the 17th IFAC World Congress, accepted, 2008.
design approach, a design example has been provided. The
example has shown the utility of our approach.

4880
Authorized licensed use limited to: National Kaohsiung Univ. of Science & Technology. Downloaded on May 01,2025 at 09:31:34 UTC from IEEE Xplore. Restrictions apply.