You are on page 1of 12

NEURO APPROACH

FOR SOLVING MATRIX RICCATI DIFFERENTIAL EQUATION

P. Balasubramaniam§, J. Abdul Samatht, N. Kumaresan§ and A. Vincent Antony Kumar§l §Department of Mathematics Gandhigram Rural University Gandhigram - 624 302, Tamilnadu, India Email: pbalgri@rediffmail.com tDepartment of Computer Science and Applications Gandhigram Rural University Gandhigram - 624 302, Tamilnadu, India Email: abduLsamath@yahoo.com Abstract: In this paper, neuro approach is proposed for solving Matrix Riccati Differential Equation(MRDE). Experiments show that the proposed Levenberg-Marquardt training method can be efficient for obtaining solutions in shorter computation time. The goal is to provide optimal control with reduced calculus effort by comparing the solutions of the MRDE in the well known RK-method and RK-Butcher method. A numerical example is presented to illustrate the implementation of artificial neural network (ANN) for the proposed method is given. Key words: Matrix Riccati Differential Equation, Levenberg-Marquardt RK-Butcher method, Neural networks. algorithm, RK-method,

1. Introduction Neural networks or simply neural nets are computing systems, which can be trained to learn a complex relationship between two or many variables or data sets. Having the structures similar to their biological counterparts, neural networks are representational and computational models processing information in a parallel distributed fashion composed of interconnecting simple processing nodes [17]. Neural net techniques have been successfully applied in various fields such as function approximation, signal processing and adaptive (or) learning control for nonlinear systems. Using neural networks a variety of off-line learning control algorithms have been developed for nonlinear systems [9, 11]. A variety of numerical algorithms have been discussed for solving the algebraic Riccati equation[3]. Neural networks problem have attracted considerable attentions in recent years of many researchers for numerical aspects for algebraic Riccati equations (see [4, 6, 18, 19]). Solving the Matrix Riccati Differential Equation(MRDE) is a central issue in optimal control theory. The needs for solving such equations often arise in analysis and synthesis such as linear-quadratic optimal control systems, robust control systems with H2 and Hoo-control [20] performance criteria, stochastic filtering and control systems, model reduction, differential games etc., and many others. In particular, the solution of algebraic Riccati matrix equations(ARME) in continuous-time (or) discrete time is an important task for nonlinear matrix equations arising in mathematics and engineering. This equation, in one form or another, has an important role in optimal control problems, multivariable and large scale systems, scattering theory estimation, detection, transportation and radiative transfer [7]. The solution of this equation is difficult to obtain from two points of view. One is nonlinear and
lThe author also working in the Department of Mathematics, 624622. Tamilnadu, INDIA, Email: vincypsna@rediffmail.com PSNA College of Engineering and Technogly, Dindigul-

x(O) = XQ where x(t) E jRn is the state vector. XQ is a given initial state vector and m :S n.the other is matrix form. R E jRmxm is a symmetric and positive definite weighting matrix for u(t). In section 2 the statement of the problem is given. An example is given which illustrates the advantage of the fast and accurate solutions compared with RK-method and RK-Butcher method solutions. If all state variables are measurable. Most general methods to solve MRDE with a terminal boundary condition are obtained on transforming MRDE into an equivalent linear differential Hamiltonian system [8]. It is well known in the control literature that to minimize J is equivalent to minimize the Hamiltonian equation 1 H(x. solution of the MRDE is presented. 15]). where f(t) and p(t) are solutions of certain first order ordinary linear differential equations. Although parallel algorithms can compute the solutions faster than sequential algorithms. then a linear state feedback control law u(t) = -W1BTA(t) can be obtained to the system described by equation (1) where K(t) E jRnxn matrix such that K(tf) = S. In [10]. t) = "2xT(t)Qx(t) along the optimal trajectory. The solution K(t) of the MRDE is obtained by using K(t) = p(t)/ f(t). a numerical example is discussed. In section 4. 16]). A. Recently Balasubramaniam et al [1. an analytic procedure of solving the MRDE of the linear quadratic control problem for homing missile systems is presented. u(t) E jRm is the control vector. However the given technique is restricted to single input. The structured neural network architecture is trained using Levenberg-Marquardt algorithm [5J to prove the efficiency for solutions in shorter computation time. there is no report on neural network solutions for MRDE. The final conclusion section demonstrates the efficiency of the method. This paper is organized as follows. A E jRnxn and B E jRnxm are known coefficient matrices associated with x(t) and u(t) respectively. u. Another class of method is based on transforming MRDE into a linear matrix differential equation and then solving MRDE analytically (or) computationally (see [10. Q E jRnxn is a symmetric and positive definite (or semidefinite) weighting matrix for x(t). Consider a completely state controllable linear system as follows: x(t) = Ax(t) + Bu(t). In section 3. 2] focused upon to find the optimal control for singular systems using neural networks. In order to minimize both the state and control signals of the feedback control system a quadratic performance index is usually minimized : 1 T (tf )Sx(tf) min J = _x 2 +- 2 I1 Q t ! [xT (t)Qx(t) + uT (t)Ru(t)]dt where the superscript T denotes the transpose operator. This paper focuses upon the implementation of neurocomputing approach for solving MRDE so for the optimal solution. 14. + "2uTRu(t) + AT(t)[Ax(t) + Bu(t)] 1 . By using this approach the solution of MRDE is obtained by partitioning the transition matrix of the associated Hamiltonian system (see [13.

aH au (x.. RK-Butcher method and Neural networks approaches respectively.() t ~(t) ] = [A [ .(t) aH a. Hence the equation (9) can be written as In the following subsections the above system will be solved by RK-method.>. .>.(t) -Q -BR-1B _AT T ] [ x(t) ] '>'(t) Since equation (8) holds for all non-zero x(t). the MRDE is transformed into a system of nonlinear differential equations.u.. In the process of solving the equation (9). we obtain the following (MRDE) for the linear system (1) x(t) must be zero. =x .t) =0 aH ax = ).>. then the term pre-multiplying Therefore.

kll(i k12(i k22(i where k1 11 m1 k2 12 m2 k3 l3 m3 k4 14 m4 + 1) + 1) + 1) kll(i) k12(i) k22(i) + ~(k1 + 2k2 + 2k3 + k4) + ~(11 + 212 + 213 + 14) + ~(m1 + 2m2 + 2m3 + m4) k12. Since the system (10) contains n2 first order ODEs with n2 variables. k22 + m3) h * iJ>13(kll + k3. k12 + 13. Runge-Kutta Method. k22 + m3).k22+ 11 12 12 12 h* iJ>12(kll h h h h h 2) m1 * iJ>13(kll + 2.k12+ k1 11 2. h * iJ>13(kll. RK-algorithms have always been considered as the best tool for the numerical integration of ordinary differential equations(ODEs). k22) k12. RK method is explained for a system of three first order ODEs with three variables. h h * iJ>ll(kll + 2' k12 + 2' k22 + 2) + 2. 3.k12 + 2. ODEs with three variables. h * iJ>12(kll. RK-Butcher algorithm order since it requires six functions evaluation. k22) k12.k22 + 2) * iJ>1l(kll + k3. is nominally considered as a sixth this algorithm is better than other a fifth order method only. k12 + 13.k12 + 2. k12 + 13.k22 + 2) * iJ>ll(kll + 2.2. Runge-Kutta Butcher Method. it is RK-Butcher method is explained for a system of three first order kll (i) k12(i) k22(i) h + 90 (7k1 + 32k3 + 12k4 + 32k5 + 7k6) + :0 +~ (711 + 3213 + 1214 + 3215 + 716) (7m1 + 32m3 + 12m4 + 32m5 + 7m6) . The accuracy of algorithms. k22) k1 k1 11 m1 m1 * iJ>ll(kll. k22 + m3) h * iJ>12(kll + k3.1.3. Even though it seems to be sixth order method.k12 k2 k2 k2 + 2.k22 + 2) m2 m2 m2 * iJ>12(kll + 2' k12 + 2' k22 + 2) * iJ>13(kll + 2.

h * 12 k12. k12. m3 + 12 * h * m4 * h * m5) .8 .2*h .k22 + h * 4 + h *""""4 11 k22 + h 11 m1 <P12 ( k11 + h * 4' k12 + h * 4' k1 * """"4 m1) <P (k 12 11 + + h h * k1 + h * k2 8 ' ' k 12 + + h h * 11 + h * 12 8 ' ' k 22 + + h h * m1 + h * m2) 8 <P (k 13 11 * k1 + h * k2 8 k 12 * 11 + h * 12 8 h * m2 -2- k 22 * m1 + h * m2) 8 <P11 ( k11.-2+ (h h * k2 . <P11 ( k11. k22 + h * 13. <P13 (k11.-2k k) 12.12 * h.-2h * k2 . k22) k12. * h * 11) + (9 * h * 14) 16 ' k 12 * h * m1) + (9 * h * m4) 16 16 ' <P12 ( k11 + + + + (3 * h * k1) + (9 * h * k4) 16 + (3 * h * 11) + (9 * h * 14) k 22 (3*h*m1)+(9*h*m4)) 16 (3 (3 <P13 ( k 11 * h * k1) + (9 * h * k4) 16 k ' 12 + (3 * h * 11) + (9 * h * 14) 16 ' * h*m1) + (9 *h*m4)) 16 k22 <P11 ( k11. 22 * 13) - + (h * m3) h * m2 -2- <P12 ( k11 + h * k3. k22) ) <P12 (k11. k22) k12.-2- <P (k 11 11.<P11 (k11.-2h * k2 .k12 + (3 h * m2 .k12. k22) h h _C * _C * * 11 * m1 2 * h * 12 * m2 12 * ~ * 13 + 12 * h * 14 - 8 * h * 15) .k22 + (3 + h * m3 ) + h * m3 ) <P13 ( k11 + h * k3. k22) .-2- + h * 13. h * k2 k12 . k12.

The network consists of integrators and adders with associated connecting weights. The architecture consists of 3 layers of artificial neurons which are connected in a feed forward node.8 ' k22 _ . The learning supervised rule that modifies the weights of the connections according to the input patterns that are presented in the Levenberg-Marquardt algorithm [12].3 and its appropriate weight vector and possibly by passing the result through the nonlinear tansigmoid functions.2 * 12 - 12 * h * 13 + 12 * h 7 -12 7 * 14 - 8 * h * 15) .<I> (k 12 11 - 3*h * k1 - 2*h 2 * k2 - 12 * h 7 * k3 + 12 * h * k4 * 14 8 8*h * k5 ' k 12 _ (3 (3 * h * 11 * h * m1 * h * 12 * m2 12 * h * 13 + 12 * h 7 -12 * h * 15) . In this approach.8 ' . Neural Network Solution. (kij)))2 =0 .2. Each neuron produces its output by computing the inner product of its input kij. During the training the weights and biases of the network are iteratively adjusted till the following error quantity becomes E = L i.j = 1.j=l 3 ((kij) - c/>ij(t. The connection weights denoted by aij (= aij) and Ji-ij are the coefficients of the third order matrices A and Ji-. A neural network architecture is shown in figure 1.3. k ~ * h * m2 * h * m3 + 12 * h * m4 * h * m5)) 3. m3 + 12 * h * m4 * h * m5)) k 12 _ (3 _ (3 * h * 11 * h * m1 2*h . a new feedforward neural network is used to solve the system of differential equations (10). i.2*h * h.

Initialize randomized weight matrix and bias in each layer. Pass the output of the third layer through integrator with initial values.8 1.6 0. The computation neural network solution is 1.509861 -0.4 0.007613 -1.624486 -2.259361 -12.3 and 4.767664 -5.2 0.2.kij)) =0 The numerical solution of MRDE were calculated and displayed respectively in the following 2 and 3 using the RK-method.242057 -0.j = 1. solution time for solution .237297 -0. Repeat the neural network training until the following error function becomes zero.083993 -6.4 1.293067 -9. tables 1.780564 1. The learning rule that was implemented with a PC.1 x 10-4 in MATLAB for the neuro computing approach to solve MRDE (9) for the linear system (1).583961 -4.0 1. The of MRDE and the deviation of RK solution is displayed in Fig.0 -0. whereas the RK method is 2.532972 -1.j=l (kij-<I>ij(t.6330sec. 2. CPU 1.0.833062 3.419200 -4. i.572630 -1.059739 0. Computing the inner product in each neuron and pass through tansigmoidal function.7 GHz which has unit roundoff == 1.336832 -2. RK-Butcher methd and the Neural Networks approach. n 2 E= L i.8 2.428650 0. Hence the neural is faster than RK method.698837 -0.981502 -0.684054 -0.076698 -3.267480 -3. Step Step Step Step Step 1: 2: 3: 4: 5: Feed the random input vector kij.530706 -0.1930sec.808259 -16.495683 This algorithm is fastest method for training moderate-sized feedforward neural networks.070754 -4.133368 -5.724925 2.2 1.3.6 1.656955 -4.

023458 -5.6 1.4832 2.0612 -1.5491 -0.331792 -0.9843 -8.6554 -0.2818 -0.716645 -0.0525 0.427505 -4.466076 0.487425 -1.6 0.0 0.258775 -0.643066 3.2 0.112443 -1.1257 -5.2291 -2.6431 3.721680 -0.9159 -3.3191 -0.4 1.4655 .238316 -0.774734 -5.6045 -1.6 0.4640 -4.8157 -15.750488 -2.622707 2.1458 -2.4 1.5424 -4.4 0.517358 -0.0628 -5.124072 -5.982123 -8.6632 -0.220125 -2.953969 -3.057855 -4.740360 1.6 1.596852 -4.7166 -0.8107 -5.0 -0.604144 -11.8 1.8 2.2 1.716365 -0.0 1.663129 -4.2927 -0.2 1.4 0.8 2.2 0.0 1.8098 1. 7266 -2.6888 -4.453504 0.627679 -1.8 1.041253 0.112111 -2.4961 -1.4907 0.807094 -15.5722 -11.0501 -4.0.

1 . .•. the computation of transformation matrix is numerically unreliable in a fixed precession of arithmetic. High accuracy of numerical solutions compared with traditional methods like Rung-Kutta.15 ·10 ~. ~. there is a difficulty to apply for higher order Riccati Differential Equation (ii) In negative exponential method. RK-Butcher method have been analysed. Computation time is very minimum when compared with existing methods stated in the survey paper [3]. ~.' 1..' 0.. . Once network is trained... "0 l:!iiii!i5] 0..3 0 O~ 0. .' •• Solving MRDE by neural network is a naval approach.05 ~ .. 0. it allows instantaneous evaluation of solutions at any desired number of points with less memory occupation. 1.25 .1 .. The neural network approach for solving MRDE is different from the previous methods discussed in the survey paper [3] in various directions stated sequentially as follows: (i) In direct integration method.." ·1 • .' 0..2 1A 1.•.' 0...

E. (1990).. M.1 x 10-4. Pad approximation is used to compute the solution. 5. Pure and Appl. E. 167-192. Math. [9] A. A. Abdul Samath. Math. 619-621. Dynamics. E. Kumaresan. Ham. M. Guidance. (viii) A matrix valued approach the accuracy of solution depends upon the solution of Sylvester equation. S. Comput. 1 (2000). J. K. Balasubramaniam. J. the accuracy of solution depends upon accuracy of solution of Algebraic Riccati equation. An overview on the solutions of the algebraic matrix Riccati equation and related problems. Lovren. 182 (2006). [11] K. Rosu. 617 . 696-700. Acknowledgements The authors would like to thank the referee for their valuable comments and suggestions to improve the quality of the paper in the presented form. 185-187. Appl. (2006) (in press). IEEE Conference on Power System Technology. [2] P. 23 (1992). Appl. Solution of matrix Riccati differential equation for the linear quadratic singular system using neural networks. New Delhi under the grant No. Large scale Systems. Narendra. Acta Numer. compared with other well-known RK-method solutions. 919-930. N. Gaiceanu. [3] C. Conclusion The numerical results of the MRDE (9) in the above tables indicate that the neural networks solutions are much more efficient. 4 (1993). The long calculus time of the MRDE (9) was avoided by using a neuro-optimal controller based LenvenbergMarquardt training algorithm. G. The efficient approximations of the optimal solution were done with PC. (v) In a method using an algebric Riccati solution the computation of matrix exponential of size n x n is also time consuming. Identification and decentralized adaptive control using neural networks with application to robotic manipulators. Sudharsanan. A. J. [6] F. Tataru. K. IEEE TI-ansactions on Neuml Networks. [8] L. [5] M. Closed analytical solution of Riccati type matrix differential equations. Ellacott. 145-202. IEEE TI-ansactions on Neuml Networks.(iii) In Davison Maki metho. Collins. 1 (1996). Kumaresan. A survey of numerical methods for solving matrix Riccati differential equation. 1832-1839. Indian J. Proceedings IEEE International Conference on Neuml Networks.. 1 (1980). [1] P. Jamshidi. M. N. (vii) In analytic approximation method. 17 (1994). L. . Parathasarathy. The work of the first author was supported by the DST. Comput. of India. Identification and control of dynamical systems using neural networks. SR/FTP /MA-05/2002. CPU 1. Cont. Analytic solution of the Riccati equation for the homing missile linear quadratic control problem. [4] S. Abdul Samath. Jodar. [10] N. Navarro.. 5 (1994). Vincent Antony Kumar. 1 (1990). Karakasoglu. Neuro-optimal controller for three-phase induction motor based on LevenbergMarquardt training algorithm Proceedings PowerCon. Aspects of the numerical analysis of neural networks. Accuracy of Pade's solution is not better than this neural network approach (iv) In ASP(Automatic Synthesis program) matrix iteration procedure for computation of matrix exponential of size 2n x 2n is time consuming.622 [7] M. Sundareshan. Choi. M.7 GHz in MATLAB which has unit roundoff ~ 1. Tomic. W. Proceedings of Southeastcon. 97-102. Math. Balasubramaniam.. H.4-27. S. A neurocomputing approach for solving the algebraic matrix Riccati equation. Optimal control for nonlinear singular systems with quadratic performance using neural networks. Govt.

Razzaghi. London(1997). Technical University of Denmark. Comput. Neural Parallel Sci. 11 (1998). Systems and Control Letters. 61-73.. Neural Networks. A Schur method for the solution of the matrix Riccati equation. Kannan. Wang. 8(34)(2000). A numerical analysis of stochastic neural networks. 14(1969). Norgaard. A multilayer recurrent neural network for solving continuous time algebraic Riccati equations. D. [20J K. Information Sci. Int. [19] B. 209--241. 16 (1978).[12] M. [14] M. 85--91. 11 (1998).. Math. Sci. 335-338. Department of Automation. Zhang. Neural Network Based System Indentification Toolbox. 20 (1997). Numerische Math. 271-279. Solution of the matrix Riccati equation in optimal control. Vaughu. Control. Math.. G. Technical Report.. . 97-&851. Wu. [16] D. Neural Network Models. 1997. J. Wilde. [15] M. [18] J. Second ed. De. [17] P. IEEE 1ra1l8 Automat. Zhou. Razzaghi. Springer-Verlag.R. An algebraic Riccati equation approach to Hoc optimization. 72-75. 32 (1979). [13] M. A computational solution for a matrix Riccati differential equation.. 939--950. Razzaghi. A negative exponential solution for the matrix Riccati equation. Khargonekar.