Processing math: 100%
Research article

Periodic solutions of Cohen-Grossberg-type Bi-directional associative memory neural networks with neutral delays and impulses

  • Received: 11 October 2020 Accepted: 16 December 2020 Published: 23 December 2020
  • MSC : 34C25, 92B20

  • This paper considers a class of delayed Cohen-Grossberg-type bi-directonal associative memory neural networks with impulses. By using Mawhin continuation theorem and constructing a new Lyapunov function, some sufficient conditions are presented to guarantee the existence and stability of periodic solutions for the impulsive neural network systems. A simulation example is carried out to illustrate the efficiency of the theoretical results.

    Citation: Shuting Chen, Ke Wang, Jiang Liu, Xiaojie Lin. Periodic solutions of Cohen-Grossberg-type Bi-directional associative memory neural networks with neutral delays and impulses[J]. AIMS Mathematics, 2021, 6(3): 2539-2558. doi: 10.3934/math.2021154

    Related Papers:

    [1] Yang Chen, Jinxia Wu, Jie Lan . Study on reasonable initialization enhanced Karnik-Mendel algorithms for centroid type-reduction of interval type-2 fuzzy logic systems. AIMS Mathematics, 2020, 5(6): 6149-6168. doi: 10.3934/math.2020395
    [2] Yang Chen, Jiaxiu Yang, Chenxi Li . Design of reasonable initialization weighted enhanced Karnik-Mendel algorithms for centroid type-reduction of interval type-2 fuzzy logic systems. AIMS Mathematics, 2022, 7(6): 9846-9870. doi: 10.3934/math.2022549
    [3] Yang Chen, Chenxi Li . Study on sensible beginning divided-search enhanced Karnik-Mendel algorithms for centroid type-reduction of general type-2 fuzzy logic systems. AIMS Mathematics, 2024, 9(5): 11851-11876. doi: 10.3934/math.2024580
    [4] Mengmeng Wang, Xiangzhi Kong . Multi-attribute group decision making algorithm based on (p, q)-rung interval-valued orthopair fuzzy set and weight optimization model. AIMS Mathematics, 2023, 8(10): 23997-24024. doi: 10.3934/math.20231224
    [5] Wenyi Zeng, Rong Ma, Deqing Li, Qian Yin, Zeshui Xu, Ahmed Mostafa Khalil . Novel operations of weighted hesitant fuzzy sets and their group decision making application. AIMS Mathematics, 2022, 7(8): 14117-14138. doi: 10.3934/math.2022778
    [6] Murugan Palanikumar, Nasreen Kausar, Harish Garg, Shams Forruque Ahmed, Cuauhtemoc Samaniego . Robot sensors process based on generalized Fermatean normal different aggregation operators framework. AIMS Mathematics, 2023, 8(7): 16252-16277. doi: 10.3934/math.2023832
    [7] Rana Muhammad Zulqarnain, Xiao Long Xin, Muhammad Saeed . Extension of TOPSIS method under intuitionistic fuzzy hypersoft environment based on correlation coefficient and aggregation operators to solve decision making problem. AIMS Mathematics, 2021, 6(3): 2732-2755. doi: 10.3934/math.2021167
    [8] Miaoxia Chen, Abdul Samad Shibghatullah, Kasthuri Subramaniam, Xiaopeng Yang . Weighted minimax programming subject to the max-min fuzzy relation inequalities. AIMS Mathematics, 2024, 9(6): 13624-13641. doi: 10.3934/math.2024665
    [9] Kanjanaporn Tansri, Pattrawut Chansangiam . Gradient-descent iterative algorithm for solving exact and weighted least-squares solutions of rectangular linear systems. AIMS Mathematics, 2023, 8(5): 11781-11798. doi: 10.3934/math.2023596
    [10] Jun Hu, Jie Wu, Mengzhe Wang . Research on VIKOR group decision making using WOWA operator based on interval Pythagorean triangular fuzzy numbers. AIMS Mathematics, 2023, 8(11): 26237-26259. doi: 10.3934/math.20231338
  • This paper considers a class of delayed Cohen-Grossberg-type bi-directonal associative memory neural networks with impulses. By using Mawhin continuation theorem and constructing a new Lyapunov function, some sufficient conditions are presented to guarantee the existence and stability of periodic solutions for the impulsive neural network systems. A simulation example is carried out to illustrate the efficiency of the theoretical results.



    The computationally relative simple interval type-2 fuzzy sets (IT2 FSs) are currently the most commonly used T2 FSs. Therefore, interval type-2 fuzzy logic systems (IT2 FLSs [1,2,57]) based on IT2 FSs have superior ability to cope with uncertainties environments like power systems [3], permanent magnetic drive [4,5,6], intelligent controllers [7], medical systems [8], pattern recognition systems [9], database and information systems [10] and so on. The footprint of uncertainty (FOU) of an IT2 FS [11] make it own more design degrees of freedom compared with a T1 FS. Generally speaking, a T2 FLS (see Figure 1) is composed of fuzzifier, rules, inference, type-reducer and defuzzifier. Among which, the block of type-reduction plays the central role of transforming the T2 to the T1 FS. Finally the defuzzification changes the T1 FS to the crisp output.

    Figure 1.  A T2 FLS [12].

    In the past decades, many types of type-reduction (TR) were proposed gradually. Among which, the most famous one is the Karnik-Mendel (KM) algorithms [13]. This type of algorithms has the advantages of preserving the uncertainties flow between the upper and lower membership functions (MFs) of T2 FLSs. Then the continuous KM (CKM) algorithms [14] were put forward, in addition, the monotoncity and super convergence property of them were proved. For the sake of improving the calculation efficiency, Wu and Mendel proposed the enhanced KM (EKM) algorithms [15]. Extensive simulation experiments show that the EKM algorithms can save two iterations on average compared with the KM algorithms. Then Liu et al. gave the theoretical explanations for the initialization of EKM algorithms and extended the EKM algorithms to three different forms of weighted-based EKM (WEKM) algorithms [16,17] to calculate the more accurate centroids of type-2 fuzzy sets.

    However, the iterative natures of KM types of algorithms make the applications of corresponding IT2 FLSs more challenge. Therefore, other types of noniterative algorithms [18] are proposed gradually, and they are Nagar-Bardini (NB) algorithms [19], Nie-Tan (NT) algorithms [20], Begian-Melek-Mendel (BMM) algorithms [21,22] and so on. Among which, IT2 FLSs based on the NB algorithms are proved to have superior performances to respond to the affect of uncertainties in systems’ parameters to the IT2 FLSs according to other algorithms like EKM, BMM, Greenfield-Chiclana Collapsing Defuzzifier (GCCD [24]) and Wu-Mendel Uncertainty Bound (UB [24]). Moreover, Nie-Tan (NT) algorithms have the simple closed form. Recent studies prove that the continuous NT (CNT) algorithms [20] to be accurate algorithms to calculate the centroid of IT2 FSs. In addition, BMM algorithms [18] are proved to be more generalized forms of NB and NT algorithms. All these works have laid theoretical foundations for studying the TR of T2 FLSs.

    In order to obtain more accurate centroid TR of IT2 FLSs, this paper extends the NB, NT, and BMM algorithms to the corresponding WNB, WNT, and WBMM algorithms according to the Newton-Cotes quadrature formulas. The rest of this paper is organized as follows. Section 2 introduces the background of IT2 FLSs. Section 3 provides the Newton-Cotes formulas, the weighted-based noniterative algorithms, and how to adopt them to perform the centroid TR of IT2 FLSs. Section 4 gives four computer simulation examples to illustrate the performances of weighted-based noniterative algorithms. Finally Section 5 is the conclusions and expectations.

    Definition 1. A T2 FS ˜A can be characterized by its T2 MF μ˜A(x,u), i.e.,

    ˜A={(x,u),μ˜A(x,u)|xX,u[0,1]} (1)

    where the primary variable xX, the secondary variable u[0,1], here Eq (1) is called as the point-value expression, whose compact form is as:

    ˜A=xXu[0,1]μ˜A(x,u)/(x,u) (2)

    Definition 2. The secondary MF of ˜A is a vertical slice of μ˜A(x,u), i.e.,

    μ˜A(x=x,u)μ˜A(x)=u[0,1]fx(u)/u. (3)

    Definition 3. The two dimensional support of μ˜A(x,u) is referred to as the footprint of uncertainty (FOU) of ˜A, i.e.,

    FOU(˜A)=xXJx={(x,u)X×[0,1]|μ˜A(x,u)>0} (4)

    The upper and lower bounds of FOU(˜A) are called as the upper MF (UMF) and lower MF (LMF), respectively, i.e.,

    UMF(˜A)=¯μ˜A(x)=¯FOU(˜A),LMF(˜A)=μ_˜A(x)=FOU(˜A)_. (5)

    Because the secondary membership grades of IT2 FSs are all uniformly equal to 1, i.e., fx(u)1, every IT2 FS can be completely described by its UMF and LMF. Then the vertical slices representation of ˜A can be as:

    ˜A=xXμ˜A(x)/x=xX˜A(x)/x (6)

    Here μ˜A(x) can be denoted as ˜A(x) for simplicity.

    Definition 4. An embedded T1 FS Ae depends on μ˜A(x,u), i.e.,

    Ae={(x,u(x)|xX,uJx}. (7)

    Definition 5. An IT2 FS can be considered as the union of all its embedded T2 FSs ˜Aje, i.e.,

    ˜A=mj=1˜Aje (8)

    where m denotes the number of embedded T2 FSs, ˜Aje=1/Aje, and Eq (8) is referred to as the wavy slices representation [28] of ˜A.

    From the aspect of inference structure, IT2 FLSs can usually be divided into two categories: Mamdani type [3,6,12,18,25] and Takagi-Sugeno-Kang type [4,25,26,27]. Without loss of generality, here we consider a Mamdani IT2 FLS with n inputs x1X1,xnXn and one output yY. The IT2 FLS can be characterized by M fuzzy rules, where the sth fuzzy rule is of the form:

    ˜Rs:Ifx1is˜Fs1andandxnis˜Fsnthenyis ˜Gs(s=1,2,,M) (9)

    in which ˜Fsi(i=1,,n;s=1,,M) is the antecedent IT2 FS, and ˜Gs(s=1,,M) is the consequent IT2 FS.

    For simplicity, here we use the singleton fuzzifier, i.e., the input measurements are modeled as crisp sets (type-0 FSs). As x=x, the firing interval of each fuzzy rule is computed as:

    Fs:{Fs(x)[f_s(x),¯fs(x)],f_s(x)Tni=1μ_˜Fsi(xi),¯fs(x)Tni=1¯μ˜Fsi(xi) (10)

    where T denotes the minimum or product t-norm, and f_s(x) and ¯fs(x) are the left and right end points of the firing interval, respectively.

    As for the centroid TR, we combine the firing interval of each rule with its consequent IT2 FS to obtain the fired-rule output FS ˜Bl(which can be described by its FOU):

    ˜Bs:{FOU(˜Bs)=[μ_˜Bs(y|x),¯μ˜Bs(y|x)],μ_˜Bs(y|x)=f_s(x)μ_˜Gs(y),¯μ˜Bs(y|x)=¯fs(x)¯μ˜Gs(y) (11)

    where represents the minimum or product t-norm.

    Then the output IT2 FS ˜B can be obtained by agammaegating all the fired-rule output FSs as :

    ˜B:{FOU(˜B)=[μ_˜B(y|x),¯μ˜B(y|x)],μ_˜B(y|x)=μ_˜B1(y|x)μ_˜B2(y|x)μ_˜BM(y|x),¯μ˜B(y|x)=¯μ˜B1(y|x)¯μ˜B2(y|x)¯μ˜BM(y|x) (12)

    where denotes the maximum operation.

    Finally the type-reduced set YC(x) can be obtained by computing the centroid of ˜B, i.e.,

    YC(x)=1/[l˜B(x),r˜B(x)] (13)

    where the two end points l˜B(x) and r˜B(x) can be calculated by different types of TR algorithms [13,14,15,16,17,18,19,20,21,22,23,24,29].

    Before introducing the weighted-based noniterative algorithms, we first give the preliminary knowledge: Newton-Cotes quadrature formulas [16,30].

    Generally speaking, the numerical integration is an approach that approximates the definite integral baf(x)dx according to the functional value f(xi) on some certain discrete points. Therefore, the calculation of definite integral can be ascribed to computing functional values.

    Definition 6. (Quadrature formula [16,17,46,47]) Let the discrete points satisfy that a=x0<x1<x2<<xN=b, and the definite integral baf(x)dx=Q(f)+E(f)with the property:

    Q(f)=Ni=0wif(xi)=w0f(x0)++wNf(xN) (14)

    then Eq (14) is referred to as the numerical integration or quadrature formula, where E(f) is called as the remainder or truncation error of integration, {wi}Ni=0 is called as the weight coefficient, and {xi}Ni=0 is the integration node.

    Next, the composite trapezoidal rule, composite Simpson rule, and composite Simpson 3/8 rule to adopted to approximate f(x) as the straight line, quadratic polynomial function, and cubic polynomial function, respectively.

    Theorem 1. (Composite trapezoidal rule [16,17,46,47]) Let y=f(x) be a function defined on [a,b]. Divide the interval [a,b] into N subintervals {xi=1,xi}Ni=1 with the equidistance h=baN, where the equidistance node is as xi=x0+ih(i=0,1,,N), then the numerical approximation of definite integral with the composite trapezoidal rule is as

    baf(x)dx=h2[f(a)+f(b)+2N1i=1f(xi)]+ET(f,h) (15)

    Suppose that f be second order continuous differentiable on [a,b], then the remainder ET(f,h)=(ba)f(ζ)12h2, where ζ(a,b).

    Theorem 2. (Composite Simpson rule [16,17,46,47]) Let y=f(x) be a function defined on [a,b]. Divide the interval [a,b] into 2N subintervals {xi=1,xi}2Ni=1 with the equidistance h=ba2N, where the equidistance node is as xi=x0+ih(i=0,1,,2N), then the numerical approximation of definite integral with the composite Simpson rule is as

    baf(x)dx=h3[f(a)+f(b)+2N1i=1f(x2i)+4N1i=0f(x2i+1)]+ES(f,h) (16)

    Suppose that f be fourth order continuous differentiable on [a,b], then the remainder ES(f,h)=(ba)f(4)(ζ)180h4, where ζ(a,b).

    Theorem 3. (Composite Simpson 3/8 rule [16,17,46,47]) Let y=f(x) be a function defined on [a,b]. Divide the interval [a,b] into 3N subintervals {xi=1,xi}3Ni=1 with the equidistance h=ba3N, where the equidistance node is as xi=x0+ih(i=0,1,,3N), then the numerical approximation of definite integral with the composite Simpson 3/8 rule is as

    baf(x)dx=3h8[f(a)+f(b)+Ni=12f(x3i)+Ni=13f(x3i2)+Ni=13f(x3i1)]+ESC(f,h) (17)

    Suppose that f be fourth order continuous differentiable on [a,b], then the remainder ESC(f,h)=(ba)f(4)(ζ)80h4, where ζ(a,b).

    Here all the integrals are measured in the Lebesgue sense.

    IT2 FLSs based on the closed form of Nagar-Bardini (NB) algorithms [19,46,47] can make distinctly improvement on coping with uncertainties. For the centroid output IT2 FS ˜B, suppose that the primary variable y be equally discretized into N points, i.e., y1<y2<<yN, then the left and right centroid interval can be computed as:

    l˜B=Ni=1yiμ_˜B(yi)Ni=1μ_˜B(yi), and r˜B=Ni=1yi¯μ˜B(yi)Ni=1¯μ˜B(yi). (18)

    Then the defuzzified output can be obtained as:

    yNB=l˜B+r˜B2. (19)

    Similar to the continuous KM types of algorithms [13,14,15,16,17,31], the continuous NB (CNB) algorithms can be adopted for studying the theoretical property of centroid TR and defuzzification of IT2 FLSs.

    Let a=y1<y2<<yN=b, where a and b are the left and right end point of y, respectively, then the CNB algorithms calculate the centroids as:

    l˜B=bayμ_˜B(y)dybaμ_˜B(y)dy, and r˜B=bay¯μ˜B(y)dyba¯μ˜B(y)dy. (20)

    Here the output centroid two end points can be computed without iterations. Furthermore, the output is a linear combination of the output of two T1 FLSs: one constructed from the LMFs, and the other constructed from the UMFs.

    In this section, we propose a type of weighted NB (WNB) algorithms, i.e.,

    l˜B=Ni=1wiyiμ_˜B(yi)Ni=1wiμ_˜B(yi), and r˜B=Ni=1wiyi¯μ˜B(yi)Ni=1wi¯μ˜B(yi). (21)

    WNB algorithms can be considered as the numerical implementation of CNB algorithms. Comparing Eqs (18) and (20), it is found that the CNB and NB algorithms are very similar, i.e., the sum operations in the discrete version are transformed into the definite integral operations in the continuous version, i.e., the sum operations for the sampling points yi play the role of integrating for the integrand.

    According to the quadrature formula (see Eq (14)), the corresponding weights wi for every MF of sampling points yi can be assigned appropriately, then the more accurate computational results may be obtained. Actually, many types of weights assignment method can be used. However, this paper only considers the numerical integration methods based on the Composite trapezoidal rule, Composite Simpson rule, and Composite Simpson 3/8 rule in Newton-Cotes quadrature formulas. And the corresponding WNB algorithms are referred to as the TWNB, SWNB, and S3/8WNB algorithms, respectively. Table 1 provides the weights assignment approach for the three types of weighted-based noniterative algorithms.

    Table 1.  Weights assignment for the weighted-based noniterative algorithms.
    Algorithms Integration rule Weights
    NB, NT, BMM ______________ wi=1(i=1,,N)
    TWNB, TWNT, TWBMM Composite Trapezoidal rule wi={1/2,i=1,N,1,i1,N.
    SWNB, SWNT, SWBMM Composite Simpson rule wi={1/2,i=1,N,1,i=1mod(2),i1,N,2,i=0mod(2),iN.
    S3/8WNB, S3/8WNT, S3/8WBMM Composite Simpson 3/8 rule wi={1/3,i=1,N,2/3,i=1mod(3),i1,N,1,i=2mod(3),iN,1,i=0mod(2),iN.

     | Show Table
    DownLoad: CSV

    The closed form of discrete Nie-Tan (NT) algorithms can compute the centroid output of ˜B straightly as:

    yNT=Ni=1yi[μ_˜B(yi)+¯μ˜B(yi)]Ni=1[μ_˜B(yi)+¯μ˜B(yi)]. (22)

    Most recent studies prove the continuous NT (CNT) algorithms [20] to be an accurate method for computing the centroids of IT2 FSs, i.e.,

    yCNT=bay[μ_˜B(y)+¯μ˜B(y)]dyba[μ_˜B(y)+¯μ˜B(y)]dy. (23)

    This section proposes a type of weighted NT (WNT) algorithms, i.e.,

    yWNT=Ni=1wiyi[μ_˜B(yi)+¯μ˜B(yi)]Ni=1wi[μ_˜B(yi)+¯μ˜B(yi)]. (24)

    WNT algorithms can be viewed as the numerical implementation of CNT algorithms. The sum operations in the discrete NT algorithms are transformed into the definite integral operations in the continuous NT (CNT) algorithms, i.e., the sum operations for the sampling points yi act as integrating for the integrand. Therefore, we can assign weights for every MF of sampling points yi to try to obtain more accurate computational results. Here the corresponding WNT algorithms are referred to as the TWNT, SWNT, and S3/8WNT algorithms, respectively.

    Begian-Melek-Mendel (BMM) algorithms can also obtain the output of IT2 FLSs directly, i.e.,

    yBMM=αNi=1yiμ_˜B(yi)Ni=1μ_˜B(yi)+βNi=1yi¯μ˜B(yi)Ni=1¯μ˜B(yi) (25)

    where α and β are two adjustable coefficients.

    IT2 FLSs based on BMM algorithms [13,21,22] are superior to T1 FLSs counterparts on both robustness and stability. In addition, the BMM algorithms are more generalized form of NB and NT algorithms. Observing the Eqs (19) and (25), it can be found that BMM and NB algorithms are exactly the same while α=αNB=12, and β=βNB=12. For the NT algorithms, the Eq (22) can be simply transformed to:

    yNT=Ni=1μ_˜B(yi)Ni=1[μ_˜B(yi)+¯μ˜B(yi)]×Ni=1yiμ_˜B(yi)Ni=1μ_˜B(yi)+Ni=1¯μ˜B(yi)Ni=1[μ_˜B(yi)+¯μ˜B(yi)]×Ni=1yi¯μ˜B(yi)Ni=1¯μ˜B(yi) =αNTNi=1yiμ_˜B(yi)Ni=1μ_˜B(yi)+βNTNi=1yi¯μ˜B(yi)Ni=1¯μ˜B(yi) (26)

    in which

    αNT=Ni=1μ_˜B(yi)Ni=1[μ_˜B(yi)+¯μ˜B(yi)], and βNT=Ni=1¯μ˜B(yi)Ni=1[μ_˜B(yi)+¯μ˜B(yi)]. (27)

    Comparing the Eqs (25) and (26), it can be found that, as α=αNT and β=βNT, BMM and NT algorithms become the same.

    Continuous BMM (CBMM) algorithms can also be used for studying the theoretical properties of TR of IT2 FLSs, i.e.,

    yCBMM=αbayμ_˜B(y)dybaμ_˜B(y)dy+βbay¯μ˜B(y)dyba¯μ˜B(y)dy. (28)

    Based on the quadrature formula, this section gives a type of weighted BMM (WBMM) algorithms, i.e.,

    yWBMM=αNi=1wiyiμ_˜B(yi)Ni=1wiμ_˜B(yi)+βNi=1wiyi¯μ˜B(yi)Ni=1wi¯μ˜B(yi). (29)

    Here WBMM algorithms can be considered as the numerical implementation of CBMM algorithms. Comparing Eqs (25) and (28), it is found that the CBMM and BMM algorithms are very similar, i.e., the sum operations in the discrete version are transformed into the definite integral operations in the continuous version, i.e., the sum operations for the sampling points yi play the role of integrating for the integrand. This paper considers the numerical integration methods based on the Composite trapezoidal rule, Composite Simpson rule, and Composite Simpson 3/8 rule in Newton-Cotes quadrature formulas. And the corresponding WBMM algorithms are referred to as the TWBMM, SWBMM, and S3/8WBMM algorithms, respectively.

    For the above three types of weighed-based noniterative algorithms (see Table 1), suppose that the primary variable be the letter x, and x is uniformly sampled on[a,b], i.e., xi=a+i1N1(ba), i=1,,N, and mod represents the modular arithmetic.

    Next we can make conclusions about assigning weights for three types of weighted-based noniterative algorithms as:

    1) Substitute xi(i=0,1,,N), x0=a, xN=b in Eq (15), xi(i=0,1,,2N), x0=a, x2N=b in Eq (16), and xi(i=0,1,,3N), x0=a, x3N=b in Eq (17) all by xi(i=1,,N), x1=a, xN=b.

    2) The coefficients h/2h/33h/8 in Eqs (15)-(17) can be cancelled straightly by the quotient of two definite integrals.

    3) In Tables 13, for the TWNB, TWNT, TWBMM, and SWNB, SWNT, SWBMM algorithms, the weights are assigned as 1/2 of the right of Eqs (15) and (16); while for the S3/8WNB, S3/8WNT, and S3/8WBMM algorithms, the weighted are assigned as 1/3 of the right of Eq (17).

    Table 2.  MF expressions for FOUs.
    Num Expression
    1 μ_˜A1(x)=max{[x16, 1x47x6, 4<x70, otherwise],[x36, 3x58x9, 5<x80, otherwise]},
    ¯μ˜A1(x)=max{[x12, 1x37x4, 3<x70, otherwise],[x25, 2x6162x5, 6<x80, otherwise]}
    2 μ_˜A2(x)={0.6(x+5)19,5x2.60.4(14x)19,2.6<x14, ¯μ˜A2(x)={exp[12(x25],5x7.185exp[12(x91.75],7.185<x14
    3 μ_˜A3(x)=max{0.5exp[(x3)22],0.4exp[(x6)22]},
    ¯μ˜A3(x)=max{exp[0.5(x3)24],0.8exp[0.5(x6)24]},
    4 μ_˜A4(x)=exp[12(x30.25)2], ¯μ˜A4(x)=exp[12(x31.75)2],

     | Show Table
    DownLoad: CSV
    Table 3.  Defuzzified values obtained by three types of continuous noniterative algorithms.
    Num Approach
    CNB CNT CBMM
    1 4.3050 4.3208 4.3041
    2 3.7774 3.7141 3.7747
    3 4.3702 4.3953 4.3690
    4 5.0000 5.0000 5.0000

     | Show Table
    DownLoad: CSV

    4) In Tables 13, for the SWNB, SWNT, and SWBMM, and S3/8WNB, S3/8WNT, and S3/8WBMM algorithms, the number of sampling N is not only restricted to N=2n+1 and N=3n+1(as the requirements of N=1mod(2)and N=1mod(3) in Eqs (16) and (17), where n is an integer).

    Finally the inner relations between weighted-based noniterative algorithms and continuous noniterative algorithms for performing the centroid TR of IT2 FLSs can be made as:

    1) Weighted-based noniterative algorithms calculate the type-reduced set YWNontierative based on the linear combinations of functional values on sampling points. While the benchmark continuous noniterative algorithms compute according to the integral operations. In theory, the solutions of weighted-based noniterative algorithms will approach to the continuous noniterative algorithms as the number of sampling N.

    2) As the number of sampling increases, weighted-based noniterative algorithms may obtain more accurate computational results.

    3) Weighted-based noniterative algorithms perform the numerical calculations according to the sum operation, whereas the continuous noniterative algorithms perform the calculations symbolically by means of the integral operations. On the whole, weighted-based noniterative algorithms can be viewed as the numerical implementation of continuous noniterative algorithms according to the numerical integration approaches.

    Four computer simulation examples are provided in this section. Here we suppose that the centroid output IT2 FS [16,18,29,32] has been obtained by merging or weighting fuzzy rules under the guidance the inference before the TR and defuzzification. For the first example, the FOU is bounded by the piece-wise linear functions. For the second example, the FOU is bounded by both the piece-wise linear functions and Gaussian functions. For the third example, the FOU is bounded by the Gaussian functions. For the last example, the FOU is defined as the symmetric Gaussian primary MF with uncertainty derivations. Then Figure 2 and Table 2 show the defined FOUs for four examples. In examples 1, 3 and 4, here we choose the primary variable x[0,10] for tests. In example 2, the primary variable is selected as x[5,15] for test.

    Figure 2.  Graphs of FOUs, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.

    Firstly, the CNB, CNT, and CBMM algorithms are considered the benchmarks to compute the centroid defuzzified values for four examples. And they are provided in Table 3. Here the adjustable coefficient α for CBMM algorithms is chosen as the average of 1000 random numbers distributed on [0, 1], and the other adjustable coefficient is defined asβ=1α.

    Next, we study the performances of proposed three types of weighted-based noniteraitve algorithms, respectively. Here the number of sampling is chosen as N=50:50:4000 for tests, then the graphs of centroid defuzzified values computed by weighted-based noniterative algorithms are shown in Figures 35, respectively. Furthermore, the functional graphs of absolute errors between the continuous noniteraive algorithms and weighted-based noniteraive algorithms are shown in Figures 68.

    Figure 3.  Defuzzified values computed by the WNB algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 4.  Defuzzified values computed by the WNT algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 5.  Defuzzified values computed by the WBMM algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 6.  Absolute errors of defuzzified values between the CNB and WNB algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 7.  Absolute errors of defuzzified values between the CNT and WNT algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 8.  Absolute errors of defuzzified values between the CBMM and WBMM algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.

    Next, let’s quantitatively measure the calculation accuracies of proposed weighted-based noniterative algorithms. For the number of sampling N=50:50:4000, we compute the defined average relative errors |yWNoniteriveiyCNoniterativei|/|yCNoniterativei|(i=1,,4). Then Tables 46 provide the values of average of relative errors with respect to N, in which the last line represents the total averages for four examples.

    Table 4.  Averages of relative errors for (|yWNBiyCNBi|/|yCNBi|(i=1,,4).
    Algorithm NB TWNB SWNB S3/8WNB
    Example 1 0.000029 0.000029 0.000572 0.000067
    Example 2 0.038400 0.001690 0.001450 0.001390
    Example 3 0.009800 0.000094 0.000122 0.001908
    Example 4 0.000110 0.000110 0.000110 0.000731
    Total average 0.016110 0.000640 0.000750 0.001370

     | Show Table
    DownLoad: CSV
    Table 5.  Averages of relative errors for |yWNTiyCNTi|/|yCNTi|(i=1,,4).
    Algorithm NT TWNT SWNT S3/8WNT
    Example 1 0.000043 0.000043 0.000519 0.00096
    Example 2 0.067610 0.002170 0.002430 0.002280
    Example 3 0.013730 0.000050 0.000050 0.002770
    Example 4 0.000028 0.000028 0.000028 0.000458
    Total average 0.027140 0.000760 0.001010 0.001870

     | Show Table
    DownLoad: CSV
    Table 6.  Averages of relative errors for |yWBMMiyCBMMi|/|yCBMMi|(i=1,,4).
    Algorithm BMM TWBMM SWBMM S3/8WBMM
    Example 1 0.000028 0.000028 0.000576 0.000056
    Example 2 0.039650 0.001590 0.001320 0.001290
    Example 3 0.009598 0.000096 0.000125 0.001866
    Example 4 0.000112 0.000112 0.000112 0.000738
    Total average 0.016460 0.000610 0.000710 0.001320

     | Show Table
    DownLoad: CSV

    Then we study the specific computation times of weighted-based noniterative algorithms for better applications. The number of sampling is still chosen as N=50:50:4000. Here the unrepeatable computation times depend on the hardware and software environments. The simulations are performed by a dual-core CPU dell desktop with E5300@2.60GHz and 2.00 GB memory. The programs are operated by Matlab 2013a on Windows XP. Figures 911 provide the comparisons of computation times of weighted-based noniteraitve algorithms for these four examples.

    Figure 9.  Comparsions of computation times for WNB algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 10.  Comparsions of computation times for WNT algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.
    Figure 11.  Comparsions of computation times for WBMM algorithms, (a) example 1; (b) example 2; (c) example 3; and (d) example 4.

    Considering the Figures 38, Tables 46, and Figures 911 comprehensively, the following conclusions can be obtained:

    1) In these four examples, the absolute errors of three types of weighted-based noniterative algorithms all converge as the number of sampling increases. In example 1, NB, TWNB, NT, TWNT, and BMM, TWBMM algorithms can obtain the smallest absolute errors and errors amplitudes of variation, while SNB, SNT, and SBMM algorithms get the largest absolute errors and errors amplitude of variation. In both examples 2 and 3, the proposed weighted-based noniterative can obtain the values of absolute errors and errors amplitudes of variation that are obviously less than their corresponding original noniterative algorithms. In the last example, the first three types of weighted-based noniterative algorithms can get almost the same absolute errors and errors amplitudes of variation, while the last type of weighted-based noniterative algorithms obtain the larger ones.

    2) For the NB algorithms, the largest average of relative error is 3.84%. While the largest average of relative error of proposed WNB algorithms is only 0.169%. For the NT algorithms, the largest average of relative error is 6.761%. The largest average of relative error of proposed WNT algorithms is only 0.243%. For the BMM algorithms, the largest average of relative error is 3.965%. The largest average of relative error of proposed WBMM algorithms is only 0.159%.

    3) The total mean of average of relative error of NB algorithms is 1.611%. While the largest total mean of average of relative error of proposed WNB algorithms is only 0.137%. The total mean of average of relative error of NT algorithms is 2.714%. While the largest total mean of average of relative error of proposed WNT algorithms is only 0.187%. The total mean of average of relative error of BMM algorithms is 1.646%. While the largest total mean of average of relative error of proposed WBMM algorithms is only 0.132%.

    4) In general, see Figures 911, the computational speeds of original noniteraive algorithms are faster than their corresponding weighted-based noniterative algorithms. However, the computational speeds first two types on weighted-based noniterative algorithms are almost completely the same. As the number of sampling is fixed, the size relation of computation times is as: S3/8WNoiteraive > SWNoiteraive > TWNoniterative > Noniterative. It may just because the weights of proposed weighted-based noniterative algorithms are more complex than the noniterative algorithms. In other words, the convergence speeds of proposed weighted-based noniteraive algorithms are faster than the original noniterative algorithms.

    5) From the above analysis, it can be found that, by choosing the proposed weighted-based noniteraive algorithms appropriately, which can improve both the calculation accuracies and convergence speeds.

    The proposed weighted-based noniteraive algorithms can be used to investigate the TR and defuzzification of IT2 FLSs. If only the computational accuracy were considered, the proposed three types of weighted-based noniteraive algorithms outperformed the original noniteraive algorithms, in which the second type of weighed-based noniteraive algorithms were the best. Moreover, the computation time of proposed weighted-based noniteraive algorithms were not much different from the original noniteraive algorithms. Considering the above analysis comprehensively, we advise to use the second or third types of weighted-based noniterative algorithms for the TR and defuzzification of IT2 FLSs with the combination of linear functions and nonlinear functions as in examples 1 and 4, and adopt the first or second types of weighted-based noniteraive algorithms for the TR and defuzzification of IT2 FLSs with the combination of linear and nonlinear functions and nonlinear functions as in examples 2 and 3.

    Finally, it is important to point out that, we only focus on the experimental performances of weighted-based noniteraive algorithms. It can be obtained from the simulation examples that, compare with the noniteraive algorithms, the proposed weighted-based noniteraive algorithms can improve the computational accuracies. However, if the requirements of computational accuracy were not high, the weighted-based noniteraive algorithms can not show their advantages, as the simplest noniteraive algorithms could attain well results.

    This paper compares the operations between three types of discrete noniterative algorithms with their corresponding continuous versions. According to the Newton-Cotes quadrature formulas in the numerical integration technique, three types of noniterative algorithms are extended to the weighted-based noniterative algorithms. The continuous noniterative algorithms are considered as the benchmarks for performing the centroid TR and defuzzification of IT2 FLSs. Four simulation examples illustrate and analyze the computational accuracies and computation times of the proposed algorithms. Compared with the original noniterative algorithms, the proposed weighted-based noniteraive algorithms can obtain both higher calculation accuracies and faster convergence speeds.

    In the future work, we will concentrate on designing the centroid TR of T2 FLSs [13,14,15,16,17,18,19,20,21,22,23,24,29,31,33,56] with weighted-based reasonable initialization enhanced Karnik-Mendel algorithms, the center-of-sets TR [12,13,34,45,46,47,48,49] of T2 FLSs, and seeking for global optimization algorithms [3,4,5,6,25,26,27,35,36,37,38,39,40,41,42,43,44,52,53] for designing and applying IT2 or GT2 FLSs in real world problems like forecasting, control [50,51,54,55] and so on.

    The paper is supported by the National Natural Science Foundation of China (No. 61973146, No. 61773188, No. 61903167, No. 61803189), the Liaoning Province Natural Science Foundation Guidance Project (No. 20180550056), and Talent Fund Project of Liaoning University of Technology (No. xr2020002). The author is very thankful to Professor Jerry Mendel, who has given the author some important advices.

    The authors declare that they have no conflict of interest.



    [1] G. Carpenter, Neural network models for pattern recognition and associative memory, Neural Networks, 4 (1989), 43-57.
    [2] A. Miller, B. Blott, T. Hames, Review of neural network applications in medical imaging and signal processing, Med. Biol. Eng. Comput., 30 (1992), 49-64.
    [3] B. Hu, Z. Guan, G. Chen, F. L. Lewis, Multistability of delayed hybrid impulsive neural networks with application to associative memories, IEEE Trans. Neural Networks Learn. Syst., 30 (2018), 1-15.
    [4] M. Cohen, S. Grossberg, Absolute stability and global pattern formation and parallel memory storage by competitive neural networks, IEEE Trans. Man Cybernet. SMC, 13 (1983), 815-826.
    [5] B. Kosko, Bidirectional associative memories, IEEE Trans. Syst. Man Cybern. SMC, 18 (1988), 49-60. doi: 10.1109/21.87054
    [6] S. Guo, L. Huang, Stability analysis of a delayed Hopfield neural network, Phys. Rev. E, 67 (2003), 061902. doi: 10.1103/PhysRevE.67.061902
    [7] J. Cao, Global stability analysis in delayed cellular neural networks, Phys. Rev. E, 59 (1999), 5940-5944. doi: 10.1103/PhysRevE.59.5940
    [8] W. Yao, C. Wang, Y. Sun, C. Zhou, H. Lin, Exponential multistability of memristive Cohen-Grossberg neural networks with stochastic parameter perturbations, Appl. Math. Comput., 386 (2020), 1-18.
    [9] Z. Dong, X. Wang, X. Zhang, A nonsingular M-matrix-based global exponential stability analysis of higher-order delayed discrete-time Cohen-Grossberg neural networks, Appl. Math. Comput., 385 (2020), 125401.
    [10] L. Ke, W. Li, Exponential synchronization in inertial Cohen-Grossberg neural networks with time delays, J. Franklin Inst., 356 (2019), 11285-11304. doi: 10.1016/j.jfranklin.2019.07.027
    [11] S. Gao, R. Shen, T. Chen, Periodic solutions for discrete-time Cohen-Grossberg neural networks with delays, Phys. Lett. A, 383 (2019), 414-420. doi: 10.1016/j.physleta.2018.11.016
    [12] C. Bai, Global exponential stability and existence of periodic solution of Cohen-Grossberg type neural networks with delays and impulses, Nonlinear Anal.: Real World Appl., 9 (2008), 747-761. doi: 10.1016/j.nonrwa.2006.12.007
    [13] C. Aouiti, I. B. Gharbiaa, J. Cao, A. Alsaedi, Dynamics of impulsive neutral-type BAM neural networks, J. Franklin Inst., 356 (2019), 2294-2324. doi: 10.1016/j.jfranklin.2019.01.028
    [14] F. Kong, Q. Zhu, K. Wang, J. J. Nieto, Stability analysis of almost periodic solutions of discontinuous BAM neural networks with hybrid time-varying delays and D operator, J. Franklin Inst., 356 (2019), 11605-11637. doi: 10.1016/j.jfranklin.2019.09.030
    [15] J. Cao, Global asymptotic stability of delayed bi-directional associative memory neural networks, Appl. Math. Comput., 142 (2003), 333-339.
    [16] Z. Pu, R. Rao, Exponential stability criterion of higher order BAM neural networks with delays and impulse via fixed point approach, Neurocomputing, 292 (2018), 63-71. doi: 10.1016/j.neucom.2018.02.081
    [17] S. Cai, Q. Zhang, Existence and stability of periodic solutions for impulsive fuzzy BAM Cohen-Grossberg neural networks on time scales, Adv. Difference Equ., 64 (2016), 944-967.
    [18] C. Aouiti, F. Dridi, New results on interval general Cohen-Grossberg BAM neural networks, J. Syst. Sci. Complex., 33 (2020), 944-967. doi: 10.1007/s11424-020-8048-9
    [19] F. Yang, C. Zhang, D. Wu, Global stability analysis of impulsive BAM type Cohen-Grossberg neural networks with delays, Appl. Math. Comput., 186 (2007), 932-940.
    [20] C. Bai, Periodic oscillation for Cohen-Grossberg-type bidirectional associative memory neural networks with neutral time-varying delays, In: 2009 Fifth International Conference on Natural Computation, 2 (2009), 18-23.
    [21] S. Chen, X. Liu, Stability analysis of discrete-time coupled systems with delays, J. Franklin Inst., 357 (2020), 9942-9959. doi: 10.1016/j.jfranklin.2020.07.035
    [22] Q. Feng, S. Nguang, A. Seuret, Stability analysis of linear coupled differential-difference systems with general distributed delays, IEEE Trans. Automat. Control, 65 (2020), 1356-1363. doi: 10.1109/TAC.2019.2928145
    [23] S. Arik, New criteria for stability of neutral-type neural networks with multiple time delays, IEEE Trans. Neural Networks Learn. Syst., 31 (2020), 1504-1513. doi: 10.1109/TNNLS.2019.2920672
    [24] C. Cheng, T. Liao, J. Yuan, C. Hwang, Globally asymptotic stability of a class of neutral type neural networks, IEEE Trans. Syst. Man Cybern. SMC, 36 (2006), 1191-1194. doi: 10.1109/TSMCB.2006.874677
    [25] J. Liu, G. Zong, New delay-dependent asymptotic stability conditions concerning BAM neural networks of neutral-type, Neurocomputing, 72 (2009), 2549-2555. doi: 10.1016/j.neucom.2008.11.006
    [26] Z. He, C. Li, H. Li, Q. Zhang, Global exponential stability of high-order Hopfield neural networks with state-dependent impulses, Phys. A, 542 (2020), 1-21.
    [27] R. Kumar, S. Das, Exponential stability of inertial BAM neural network with time-varying impulses and mixed time-varying delays via matrix measure approach, Commun. Nonlinear Sci. Numer. Simul., 81 (2020), 1-13.
    [28] J. Chen, C. Li, X. Yang, Asymptotic stability of delayed fractional-order fuzzy neural networks with impulse effects, J. Franklin Inst., 355 (2018), 7595-7608. doi: 10.1016/j.jfranklin.2018.07.039
    [29] J. Wang, H. Jiang, T. Ma, C. Hu, A new approach based on discrete-time high-order neural networks with delays and impulses, J. Franklin Inst., 355 (2018), 4708-4726. doi: 10.1016/j.jfranklin.2018.04.032
    [30] X. Wang, X. Yu, S. Zhong, et al., Existence and globally exponential stability of periodic solution of BAM neural networks with mixed delays and impulses, Dyn. Contin. Discrete Impuls. Syst. Ser. B Appl. Algorithms, 24 (2017), 447-459.
    [31] X. Zhang, C. Li, T. Huang, Impacts of state-dependent impulses on the stability of switching Cohen-Grossberg neural networks, Adv. Difference Equ., 316 (2017), 1-21.
    [32] H. Gu, H. Jiang, Z. Teng, BAM-type impulsive neural networks with time-varying delays, Nonlinear Anal. Real World Appl., 10 (2009), 3059-3072. doi: 10.1016/j.nonrwa.2008.10.039
    [33] H. Liao, Z. Zhang, L. Ren, W. Peng, BAM neural networks via novel computing method of degree and inequality techniques, Chaos, Solitons Fractals, 104 (2017), 785-797. doi: 10.1016/j.chaos.2017.09.035
    [34] R. E. Gains, J. L. Mawhin, Coincidence degree and nonlinear differential equations, Springer Verlag, Berlin., 1977.
  • This article has been cited by:

    1. Chuang Liu, Jinxia Wu, Weidong Yang, Robust H output feedback finite-time control for interval type-2 fuzzy systems with actuator saturation, 2022, 7, 2473-6988, 4614, 10.3934/math.2022257
    2. Xiaoyu Peng, Xiaodong Pan, Interval type-2 fuzzy systems on the basis of vague partitions and their approximation properties, 2024, 43, 2238-3603, 10.1007/s40314-024-02629-2
  • Reader Comments
  • © 2021 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2241) PDF downloads(48) Cited by(2)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog