NUMERICAL SOLUTIONS OF FUZZY DIFFERENTIAL EQUATIONS BY TAYLOR METHOD

Transcription

1 COMPUTATIONAL METHODS IN APPLIED MATHEMATICS, Vol.2(2002), No.2, pp c Institute of Mathematics of the National Academy of Sciences of Belarus NUMERICAL SOLUTIONS OF FUZZY DIFFERENTIAL EQUATIONS BY TAYLOR METHOD SAEID ABBASBANDY Department of Mathematics, Imam Khomeini International University Qazvin, Iran abbasbandy@yahoo.com TOFIGH ALLAH VIRANLOO Department of Mathematics, Islamic Azad University Tehran, Iran alahviranlo@yahoo.com Abstract In this paper, numerical algorithms for solving fuzzy ordinary differential equations are considered. A scheme based on the Taylor method of order p is discussed in detail and this is followed by a complete error analysis. The algorithm is illustrated by solving some linear and nonlinear fuzzy Cauchy problems Mathematics Subject Classification: 34A12; 65L05. Keywords: fuzzy differential equation, p-order Taylor method, fuzzy Cauchy problem. 1. Introduction Knowledge about dynamical systems modeled by differential equations is often incomplete or vague. It concerns, for example, parameter values, functional relationships, or initial conditions. The well-known methods for solving analytically or numerically initial value problems can only be used for finding a selected system behavior, e.g., by fixing the unknown parameters to some plausible values. However, in this case, it is not possible to describe the whole set of system behaviors compatible with our partial knowledge. We may set that the fuzzy input is somehow transformed into the fuzzy output defined by the corresponding crisp systems. This reasons us to refer such systems to as Fuzzy Input Fuzzy Output (FIFO) systems. Here, we are going to operationalize our approach, i.e., to propose a method for computing the approximate solution for a fuzzy differential equation using numerical methods. Since finding this set of solutions analytically does only work with trivial examples, a numerical approach seems to be the only way to solve such problems. The topics of fuzzy differential equations, which attracted a growing interest for some time, in particular, in relation to the fuzzy control, have been rapidly developed recent years. The concept of a fuzzy derivative was first introduced by S. L. Chang, L. A. Zadeh in [4]. It was followed up by D. Dubois, H. Prade in [5], who defined and used the extension principle.

2 114 S. Abbasbandy and T. Allah Viranloo Other methods have been discussed by M. L. Puri, D. A. Ralescu in [10] and R. Goetschel, W. Voxman in [6]. Fuzzy differential equations and initial value problems were regularly treated by O. Kaleva in [7] and [8], S. Seikkala in [11]. A numerical method for solving fuzzy differential equations has been introduced by M. Ma, M. Friedman, A. Kandel in [9] via the standard Euler method. The structure of this paper is organized as follows. In section 2, some basic results on fuzzy numbers and definition of a fuzzy derivative, which have been discussed by S. Seikkala in [9], are given. In section 3, we define the problem that is a fuzzy Cauchy one. Its numerical solution is of the main interest of this work. Solving numerically the fuzzy differential equation by the Taylor method of order p is discussed in section 4. The proposed algorithm is illustrated by some examples in section 5 and the conclusion is in section Preliminaries Consider the initial value problem { x (t) = f(t, x(t)); a t b, x(a) = α. (1) Let Y (t) be an exact solution of (1) and Y (t i ) be approximated by y i = y(t i ), which in the p-order Taylor method is as follows: and y i+1 = y i + ht (t i, y i ), i = 0, 1,..., N 1, (2) T (t i, y i ) = p 1 h i (i + 1)! f (i) (t i, y i ), (3) where (b a) a = t 0 t 1... t N = b and h = N = t i+1 t i. (4) A triangular fuzzy number v is defined by three numbers a 1 < a 2 < a 3, where the graph of v(x), the member function of the fuzzy number v, is a triangle with the base on the interval [a 1, a 3 ] and the vertex at x = a 2. We specify v as (a 1 /a 2 /a 3 ). We will write: (1) v > 0 if a 1 > 0; (2) v 0 if a 1 0; (3) v < 0 if a 3 < 0; and (4) v 0 if a 3 0. Let E be a set of all the upper semicontinuous normal convex fuzzy numbers with bounded r-level sets. It means that if v E, then the r-level set [v] r = {s v(s) r}, 0 < r 1, is a closed bounded interval which is denoted by [v] r = [v 1 (r), v 2 (r)]. Let I be a real interval. The mapping x : I E is called a fuzzy process and its r-level set is denoted by [x(t)] r = [x 1 (t; r), x 2 (t; r)], t I, r (0, 1]. The derivative x (t) of the fuzzy process x is defined by [x (t)] r = [x 1(t; r), x 2(t; r)], t I, r (0, 1],

3 Numerical solutions of fuzzy differential equations by Taylor method 115 provided that this equation determines the fuzzy number, according to Seikkala [11]. Denote by κ the set of all nonempty compact subsets of R and by κ c the subset of κ consisting of nonempty convex compact sets. Recall that ρ(x, A) = min x a a A is a distance of the point x R from A κ and that the Hausdorff separation ρ(a, B) of A, B κ is defined as ρ(a, B) = max ρ(a, B). a A Note that the notation is consistent, since ρ(a, B) = ρ({a}, B). Now, ρ is not a metric. In fact, ρ(a, B) = 0 if and only if A B. The Hausdorff metric d H on κ is defined by The metric d H on E is as follows: 3. Fuzzy Cauchy problem d H (A, B) = max{ρ(a, B), ρ(b, A)}. d (u, v) = sup{d H ([u] r, [v] r ) : 0 r 1}, u, v E. Consider the fuzzy initial value problem { x (t) = f(t, x(t)), t I = [0, T ], x(0) = x 0, (5) where f is a continuous mapping from R + R into R and x 0 E with the r-level set [x 0 ] r = [x 1 (0; r), x 2 (0; r)], r (0, 1]. The extension principle of Zadeh leads to the following definition of f(t, x) when x = x(t) is a fuzzy number: f(t, x)(s) = sup{x(τ) s = f(t, τ)}, s R. From this it follows that where by [f(t, x)] r = [f 1 (t, x; r), f 2 (t, x; r)], r (0, 1], f 1 (t, x; r) = f 2 (t, x; r) = min{f(t, u) u [x 1 (t; r), x 2 (t; r)]}, max{f(t, u) u [x 1 (t; r), x 2 (t; r)]}. The mapping f(t, x) is a fuzzy process and the derivative f (i) (t, x), i = 1,..., p is defined [f (i) (t, x)] r = [f (i) 1 (t, x; r), f (i) 2 (t, x; r)], t I, r (0, 1], provided that this equation determines the fuzzy number f (i) (t, x) E, where (6) f (i) 1 (t, x; r) = min{f (i) (t, u) u [x 1 (t; r), x 2 (t; r)]}, f (i) 2 (t, x; r) = max{f (i) (t, u) u [x 1 (t; r), x 2 (t; r)]}. (7)

4 116 S. Abbasbandy and T. Allah Viranloo Theorem 3.1. Let f satisfy f(t, v) f(t, v) g(t, v v ); t 0; v, v R, where g : R + R + R + is a continuous mapping such that r g(t, r) is nondecreasing. Let the initial value problem u (t) = g(t, u(t)), u(0) = u 0, (8) have a solution on R + for u 0 > 0 and u(t) = 0 be the only solution of (8) for u 0 = 0. Then the fuzzy initial value problem (5) has a unique fuzzy solution. Proof. [11] Since x (i) (t) = f (i 1) (t, x(t)) = from (7) it follows that f (i 2) { f (i 1) f (i 2) 1 (t, x; r) = min (t, u) + t { f (i 1) f (i 2) 2 (t, x; r) = max (t, u) + t t f (i 2) u f (i 2) u (t, x(t)) + f (i 2) x (t, x(t))f (i 2) (t, x(t)), (9) } (t, u)f (i 2) (t, u) u [x 1 (t; r), x 2 (t; r)], } (t, u)f (i 2) (t, u) u [x 1 (t; r), x 2 (t; r)]. (10) 4. Taylor method of order p Let the exact solution [Y (t)] r = [Y 1 (t; r), Y 2 (t; r)] be approximated by [y(t)] r = [y 1 (t; r), y 2 (t; r)]. The Taylor method of order p is based on the expansion x(t + h; r) = p h i i! x(i) (t; r), (11) where x(t; r) is Y 1 or Y 2. We define F [t, x; r] = p 1 p 1 G[t, x; r] = h i (i + 1)! f (i) 1 (t, x; r), h i (i + 1)! f (i) 2 (t, x; r). (12) The exact and approximate solutions at t n, 0 n N are denoted by [Y (t n )] r = [Y 1 (t n ; r), Y 2 (t n ; r)] and [y(t n )] r = [y 1 (t n ; r), y 2 (t n ; r)],

5 Numerical solutions of fuzzy differential equations by Taylor method 117 respectively. The solution is calculated at the grid points of (4). Using the Taylor method of order p and substituting Y 1, Y 2 into (11) and considering (12), we have Hence, we get where Y 1 (t n+1 ; r) Y 1 (t n ; r) + hf [t n, Y (t n ); r], Y 2 (t n+1 ; r) Y 2 (t n ; r) + hg[t n, Y (t n ); r]. y 1 (t n+1 ; r) = y 1 (t n ; r) + hf [t n, y(t n ); r], y 2 (t n+1 ; r) = y 2 (t n ; r) + hg[t n, y(t n ); r], y 1 (0; r) = x 1 (0; r), y 2 (0; r) = x 2 (0; r). The following lemmas will be applied to show convergence of these approximations, i.e., lim y 1(t; r) = Y 1 (t; r), h 0 lim y 2(t; r) = Y 2 (t; r). h 0 Lemma 4.1. Let the sequence of numbers {W n } N n=0 satisfy W n+1 A W n + B, 0 n N 1, for the given positive constants A and B. Then Proof. See [9]. W n A n W 0 + B An 1, 0 n N. A 1 Lemma 4.2. Let the sequences of numbers {W n } N n=0, {V n } N n=0 satisfy W n+1 W n + A max{ W n, V n } + B, V n+1 V n + A max{ W n, V n } + B, for the given positive constants A and B. Then, denoting we have U n = W n + V n, 0 n N, U n Ān U 0 + B Ān 1, 0 n N, Ā 1 where Ā = 1 + 2A and B = 2B. Proof. See [9]. Let F (t, u, v) and G (t, u, v) be the functions F and G in (12), where u and v are constants and u v. In other words, F (t, u, v) = p 1 p 1 G (t, u, v) = h i { } (i + 1)! min f (i) 1 (t, τ) τ [u, v], h i { } (i + 1)! max f (i) 1 (t, τ) τ [u, v], (13) (14)

6 118 S. Abbasbandy and T. Allah Viranloo or F (t, u, v) and G (t, u, v) are obtained by substituting [x(t)] r = [u, v] into (12). The domain of F and G is K = {(t, u, v) 0 t T, < v <, < u v}. Theorem 4.1. Let F (t, u, v) and G (t, u, v) belong to C p 1 (K) and the partial derivatives of F and G be bounded over K. Then, for arbitrarily fixed 0 r 1, the numerical solutions of (14) converge to the exact solutions Y 1 (t; r) and Y 2 (t; r) uniformly in t. Proof. It is sufficient to show lim y 1(t N ; r) = Y 1 (t N ; r), h 0 lim y 2 (t N ; r) = Y 2 (t N ; r), h 0 where t N = T. For n = 0, 1,..., N 1, using the Taylor theorem, we get Y 1 (t n+1 ; r) = Y 1 (t n ; r) + hf [t n, Y 1 (t n ; r), Y 2 (t n ; r)] + Y 2 (t n+1 ; r) = Y 2 (t n ; r) + hg [t n, Y 1 (t n ; r), Y 2 (t n ; r)] + hp+1 (p + 1)! Y (p+1) 1 (ξ n,1 ), hp+1 (p + 1)! Y (p+1) 2 (ξ n,2 ), (15) where ξ n,1, ξ n,2 (t n, t n+1 ). Denoting W n = Y 1 (t n ; r) y 1 (t n ; r), V n = Y 2 (t n ; r) y 2 (t n ; r), from (14) and (15) it follows that W n+1 = W n + h{f [t n, Y 1 (t n ; r), Y 2 (t n ; r)] F [t n, y 1 (t n ; r), y 2 (t n ; r)]} + V n+1 = V n + h{g [t n, Y 1 (t n ; r), Y 2 (t n ; r)] G [t n, y 1 (t n ; r), y 2 (t n ; r)]} + Hence, where W n+1 W n + 2Lh max{ W n, V n } + hp+1 (p + 1)! M, V n+1 V n + 2Lh max{ W n, V n } + hp+1 (p + 1)! M, M 1 = max Y (p+1) 1 (t; r), M 2 = max Y (p+1) 2 (t; r) for t [0, T ], hp+1 (p + 1)! Y (p+1) 1 (ξ n,1 ), hp+1 (p + 1)! Y (p+1) 2 (ξ n,2 ). M = max{m 1, M 2 }, and L > 0 is a bound for the partial derivatives of F and G. Therefore, from Lemma 4.2, we obtain W n (1 + 4Lh) n U 0 + 2hp+1 (p + 1)! M (1 + 4Lh)n 1, 4Lh V n (1 + 4Lh) n U 0 + 2hp+1 (p + 1)! M (1 + 4Lh)n 1, 4Lh

7 Numerical solutions of fuzzy differential equations by Taylor method 119 where U 0 = W 0 + V 0. In particular, Since W 0 = V 0 = 0, we have W N (1 + 4Lh) N U 0 + hp+1 T (1 + 4Lh) h 1 M, (p + 1)! 2Lh V N (1 + 4Lh) N U 0 + hp+1 T (1 + 4Lh) h 1 M. (p + 1)! 2Lh W N M e4lt 1 2L(p + 1)! hp, V N M e4lt 1 2L(p + 1)! hp. Thus, if h 0, we get W N 0 and V N 0, which completes the proof. 5. Examples Example 5.1. Consider the fuzzy initial value problem [9] { y (t) = y(t), t I = [0, 1], y(0) = ( r, r), 0 < r 1. Using the Taylor method of order p, we have p h i p y 1 (t n+1 ; r) = y 1 (t n ; r) i!, y 2(t n+1 ; r) = y 2 (t n ; r) The exact solution is given by Y 1 (t; r) = y 1 (0; r)e t, Y 2 (t; r) = y 2 (0; r)e t, which at t = 1 is Y (1; r) = [( r)e, ( r)e], 0 < r 1. The exact and approximate solutions for p = 2 and p = 4 are compared and plotted at t = 1 in Figs. 1 and 2. h i i!. 1 Euler 2nd-order Taylor 0.8 Exact Figure 1. h = 0.2

8 120 S. Abbasbandy and T. Allah Viranloo 1 Euler 4th-order Taylor 0.8 Exact Figure 2. h = 0.2 The Hausdorff distances of the exact solution from the Euler solution, the 2nd- and 4th-order Taylor ones are d = , d = and d = e 005, respectively. Example 5.2. Consider the fuzzy initial value problem [9] { y (t) = ty(t), [a, b] = [ 1, 1], y( 1) = ( e.5(1 r), e +.5(1 r)), 0 < r 1. We separate the problem into two steps. (a) t < 0. The parametric form in this case is y 1(t; r) = ty 2 (t; r), y 2(t; r) = ty 1 (t; r), y 1(t; r) = (1 + t 2 )y 2 (t; r), y 2(t; r) = (1 + t 2 )y 1 (t; r), y (3) 1 (t; r) = min{(1 + 2t + 2t 2 + t 4 )u u [y 1 (t; r), y 2 (t; r)]}, y (3) 2 (t; r) = max{(1 + 2t + 2t 2 + t 4 )u u [y 1 (t; r), y 2 (t; r)]}, y (4) 1 (t; r) = min{(2 + 4t 2 + 4t(1 + t 2 ) + 4t(1 + t 2 ) 2 + (1 + t 2 ) 4 )u u [y 1 (t; r), y 2 (t; r)]}, y (4) 2 (t; r) = max{(2 + 4t 2 + 4t(1 + t 2 ) + 4t(1 + t 2 ) 2 + (1 + t 2 ) 4 )u u [y 1 (t; r), y 2 (t; r)]}, with the initial conditions given. The unique exact solution is where A = e (t2 a 2 ) 2, B = 1 A. Y 1 (t; r) = A B 2 Y 2 (t; r) = A + B 2 y 2 (0; r) + A + B y 1 (0; r), 2 y 2 (0; r) + A B y 1 (0; r), 2

9 Numerical solutions of fuzzy differential equations by Taylor method 121 (b) t 0. The parametric equations are y 1(t; r) = ty 1 (t; r), y 2(t; r) = ty 2 (t; r), y 1(t; r) = (1 + t 2 )y 1 (t; r), y 2(t; r) = (1 + t 2 )y 2 (t; r), and [y (3) ] r, [y (4) ] r are calculated as in part (a) with the initial conditions given. The unique exact solution at t > 0 is Y 1 (t; r) = y 1 (0; r)e t2 2, Y2 (t; r) = y 2 (0; r)e t2 2. The exact and approximate solutions for p = 2 and p = 4 are compared and plotted in Figs. 3 and 4. r 1 Euler 2nd-order Taylor 0.8 Exact t Figure 3. h = Euler 4th-order Taylor 0.8 Exact Figure 4. h = 0.05

10 122 S. Abbasbandy and T. Allah Viranloo The Hausdorff distances of the exact solution from the Euler solution, the 2nd- and 4th-order Taylor ones are d = , d = and d = e 008, respectively. Example 5.3. Consider the fuzzy initial value problem y (t) = k 1 y 2 (t) + k 2, y(0) = 0, where k j > 0 (j = 1, 2) are triangular fuzzy numbers [2]. The exact solution is given by with where Y 1 (t; r) = l 1 (r) tan(w 1 (r)t), Y 2 (t; r) = l 2 (r) tan(w 2 (r)t), l 1 (r) = k 2,1 (r)/k 1,1 (r), l 2 (r) = k 2,2 (r)/k 1,2 (r), w 1 (r) = k 1,1 (r)k 2,1 (r), w 2 (r) = k 1,2 (r)k 2,2 (r), [k 1 ] r = [k 1,1 (r), k 1,2 (r)] and [k 2 ] r = [k 2,1 (r), k 2,2 (r)], k 1,1 (r) =.5 +.5r, k 1,2 (r) = 1.5.5r, k 2,1 (r) = r k 2,2 (r) = r. The r-level sets of y (t) are which defines the fuzzy number. We have Y 1(t; r) = k 2,1 (r) sec 2 (w 1 (r)t), Y 2(t; r) = k 2,2 (r) sec 2 (w 2 (r)t), f 1 (t, y; r) = min{k 1 u 2 + k 2 u [y 1 (t; r), y 2 (t; r)], k j [k j,1 (r), k j,2 (r)], j = 1, 2}, f 2 (t, y; r) = max{k 1 u 2 + k 2 u [y 1 (t; r), y 2 (t; r)], k j [k j,1 (r), k j,2 (r)], j = 1, 2}, f 1(t, y; r) = min{2k 2 1u 3 + 2uk 1 k 2 u [y 1 (t; r), y 2 (t; r)], k j [k j,1 (r), k j,2 (r)], j = 1, 2}, f 2(t, y; r) = max{2k 2 1u 3 + 2uk 1 k 2 u [y 1 (t; r), y 2 (t; r)], k j [k j,1 (r), k j,2 (r)], j = 1, 2}, 1 (t, y; r) = min{ f (i 1) (t, u) f (i 1) (t, u) u u [y 1 (t; r), y 2 (t; r)], k j [k j,1 (r), k j,2 (r)], j = 1, 2}, f (i) f (i) 2 (t, y; r) = max{ f (i 1) (t, u) f (i 1) (t, u) u u [y 1 (t; r), y 2 (t; r)], k j [k j,1 (r), k j,2 (r)], j = 1, 2}, for i = 1, 2. There are two nonlinear programming problems. They can be solved by the GAMS (the acronym stands for General Algebraic Modeling Systems) software. GAMS is designed to make the construction and solution of large and complex mathematical programming models more straightforward for programmers and more comprehensible to users of models from other disciplines, e.g., economists [1]. Thus, the Taylor method suggested in this paper can be used. The exact and approximate solutions are shown in Figs. 5 and 6 at t = 1.

11 Numerical solutions of fuzzy differential equations by Taylor method 123 r 1 Euler 2nd-order Taylor 0.8 Exact t Figure 5. h = Euler 4th-order Taylor 0.8 Exact Figure 6. h = 0.1 The Hausdorff distances of the exact solution from the Euler solution, the 2nd- and 4th-order Taylor ones are d = , d = and d = , respectively. 6. Conclusion It is shown that the convergence order of the method proposed is O(h p ), while the Euler method from [9] converges with the rate O(h) only. In comparison with [9] the solutions of Examples 5.1, 5.2 have the higher accuracy in this paper.

12 124 S. Abbasbandy and T. Allah Viranloo References [1] A. Brooke, D. Kendrick, and A. Meeraus, GAMS: A User s Guide, 2nd edn., The Scientific Press, Redwood city, CA, [2] J. J. Buckley and T. Feuring, Fuzzy differential equations, Fuzzy Sets and Systems, 110 (2000), pp [3] R. L. Burden and J. D. Faires, Numerical Analysis, seventh edn., Brooks-Cole Publishing, [4] S. L. Chang and L. A. Zadeh, On fuzzy mapping and control, IEEE Trans. Systems Man Cybernet., 2 (1972), pp [5] D. Dubois and H. Prade, Towards fuzzy differential calculus: Part 3, Differentiation, Fuzzy Sets and Systems, 8 (1982), pp [6] R. Goetschel and W. Voxman, Elementary fuzzy calculus, Fuzzy Sets and Systems, 24 (1987), pp [7] O. Kaleva, Fuzzy differential equations, Fuzzy Sets and Systems, 24 (1987), pp [8] O. Kaleva, The Cuachy problem for fuzzy differential equations, Fuzzy Sets and Systems, 35 (1990), pp [9] M. Ma, M. Friedman, and A. Kandel, Numerical solutions of fuzzy differential equations, Fuzzy Sets and Systems, 105 (1999), pp [10] M. L. Puri and D. A. Ralescu, Differentials of fuzzy functions, J. Math. Anal. Appl., 91 (1983), pp [11] S. Seikkala, On the fuzzy initial value problem, Fuzzy Sets and Systems, 24 (1987), pp Received 5 Dec Revised 3 Aug. 2002