Mall, Susmita (2015) Connectionist Learning Based Numerical Solution of Differential Equations. PhD thesis.
It is well known that the differential equations are back bone of different physical systems. Many real world problems of science and engineering may be modeled by various ordinary or partial differential equations. These differential equations may be solved by different approximate methods such as Euler, Runge-Kutta, predictor-corrector, finite difference, finite element, boundary element and other numerical techniques when the problems cannot be solved by exact/analytical methods. Although these methods provide good approximations to the solution, they require a discretization of the domain via meshing, which may be challenging in two or higher dimension problems. These procedures provide solutions at the pre-defined points and computational complexity increases with the number of sampling points.In recent decades, various machine intelligence methods in particular connectionist learning or Artificial Neural Network (ANN) models are being used to solve a variety of real-world problems because of its excellent learning capacity. Recently, a lot of attention has been given to use ANN for solving differential equations. The approximate solution of differential equations by ANN is found to be advantageous but it depends upon the ANN model that one considers. Here our target is to solve ordinary as well as partial differential equations using ANN. The approximate solution of differential equations by ANN method has various inherent benefits in comparison with other numerical methods such as (i) the approximate solution is differentiable in the given domain, (ii) computational complexity does not increase considerably with the increase in number of sampling points and dimension of the problem, (iii) it can be applied to solve linear as well as non linear Ordinary Differential Equations (ODEs) and Partial Differential Equations (PDEs). Moreover, the traditional numerical methods are usually iterative in nature, where we fix the step size before the start of the computation. After the solution is obtained, if we want to know the solution in between steps then again the procedure is to be repeated from initial stage. ANN may be one of the ways where we may overcome this repetition of iterations. Also, we may use it as a black box to get numerical results at any arbitrary point in the domain after training of the model.Few authors have solved ordinary and partial differential equations by combining the feed forward neural network and optimization technique. As said above that the objective of this thesis is to solve various types of ODEs and PDEs using efficient neural network. Algorithms are developed where no desired values are known and the output of the model can be generated by training only. The architectures of the existing neural models are usually problem dependent and the number of nodes etc. are taken by trial and error method. Also, the training depends upon the weights of the connecting nodes. In general, these weights are taken as random number which dictates the training. In this investigation, firstly a new method viz. Regression Based Neural Network (RBNN) has been developed to handle differential equations. In RBNN model, the number of nodes in hidden layer may be fixed by using the regression method. For this, the input and output data are fitted first with various degree polynomials using regression analysis and the coefficients involved are taken as initial weights to start with the neural training. Fixing of the hidden nodes depends upon the degree of the polynomial.We have considered RBNN model for solving different ODEs with initial/boundary conditions. Feed forward neural model and unsupervised error back propagation algorithm have been used for minimizing the error function and modification of the parameters (weights and biases) without use of any optimization technique. Next, single layer Functional Link Artificial Neural Network (FLANN) architecture has been developed for solving differential equations for the first time. In FLANN, the hidden layer is replaced by a functional expansion block for enhancement of the input patterns using orthogonal polynomials such as Chebyshev, Legendre, Hermite, etc. The computations become efficient because the procedure does not need to have hidden layer. Thus, the numbers of network parameters are less than the traditional ANN model. Varieties of differential equations are solved here using the above mentioned methods to show the reliability, powerfulness, and easy computer implementation of the methods. As such singular nonlinear initial value problems such as Lane-Emden and Emden-Fowler type equations have been solved using Chebyshev Neural Network (ChNN) model. Single layer Legendre Neural Network (LeNN) model has also been developed to handle Lane-Emden equation, Boundary Value Problem (BVP) and system of coupled ordinary differential equations. Unforced Duffing oscillator and unforced Van der Pol-Duffing oscillator equations are solved by developing Simple Orthogonal Polynomial based Neural Network (SOPNN) model. Further, Hermite Neural Network (HeNN) model is proposed to handle the Van der Pol-Duffing oscillator equation. Finally, a single layer Chebyshev Neural Network (ChNN) model has also been implemented to solve partial differential equations.
|Item Type:||Thesis (PhD)|
|Uncontrolled Keywords:||Artificial neural network; Differential equation; Regression based neural network; Lane-Emden equation; Functional link artificial neural network; Duffing oscillator; Orthogonal polynomial.|
|Subjects:||Mathematics and Statistics > Applied Mathematics|
|Divisions:||Sciences > Department of Mathematics|
|Deposited By:||Mr. Sanat Kumar Behera|
|Deposited On:||18 Jul 2016 18:50|
|Last Modified:||18 Jul 2016 18:50|
Repository Staff Only: item control page