AN EVOLUTIONARY APPROACH FOR SOLVING DIFFERENTIAL EQUATIONS

Similar documents
On Control Problem Described by Infinite System of First-Order Differential Equations

7 Wave Equation in Higher Dimensions

General Non-Arbitrage Model. I. Partial Differential Equation for Pricing A. Traded Underlying Security

Combinatorial Approach to M/M/1 Queues. Using Hypergeometric Functions

The sudden release of a large amount of energy E into a background fluid of density

Reinforcement learning

On The Estimation of Two Missing Values in Randomized Complete Block Designs

Low-complexity Algorithms for MIMO Multiplexing Systems

A Weighted Moving Average Process for Forecasting. Shou Hsing Shih Chris P. Tsokos

Chapter Finite Difference Method for Ordinary Differential Equations

Sections 3.1 and 3.4 Exponential Functions (Growth and Decay)

MEEN 617 Handout #11 MODAL ANALYSIS OF MDOF Systems with VISCOUS DAMPING

Representing Knowledge. CS 188: Artificial Intelligence Fall Properties of BNs. Independence? Reachability (the Bayes Ball) Example

Lecture 18: Kinetics of Phase Growth in a Two-component System: general kinetics analysis based on the dilute-solution approximation

An Automatic Door Sensor Using Image Processing

Orthotropic Materials

Lecture-V Stochastic Processes and the Basic Term-Structure Equation 1 Stochastic Processes Any variable whose value changes over time in an uncertain

Lecture 17: Kinetics of Phase Growth in a Two-component System:

Molecular Evolution and Phylogeny. Based on: Durbin et al Chapter 8

Monochromatic Wave over One and Two Bars

Stress Analysis of Infinite Plate with Elliptical Hole

Application of Bernoulli wavelet method for numerical solution of fuzzy linear Volterra-Fredholm integral equations Abstract Keywords

Research on the Algorithm of Evaluating and Analyzing Stationary Operational Availability Based on Mission Requirement

FINITE DIFFERENCE APPROACH TO WAVE GUIDE MODES COMPUTATION

Quantum Algorithms for Matrix Products over Semirings

Lecture 22 Electromagnetic Waves

Computer Propagation Analysis Tools

On the Semi-Discrete Davey-Stewartson System with Self-Consistent Sources

STUDY OF THE STRESS-STRENGTH RELIABILITY AMONG THE PARAMETERS OF GENERALIZED INVERSE WEIBULL DISTRIBUTION

Discretization of Fractional Order Differentiator and Integrator with Different Fractional Orders

The shortest path between two truths in the real domain passes through the complex domain. J. Hadamard

A Negative Log Likelihood Function-Based Nonlinear Neural Network Approach

ON 3-DIMENSIONAL CONTACT METRIC MANIFOLDS

, on the power of the transmitter P t fed to it, and on the distance R between the antenna and the observation point as. r r t

Exponential and Logarithmic Equations and Properties of Logarithms. Properties. Properties. log. Exponential. Logarithmic.

CS 188: Artificial Intelligence Fall Probabilistic Models

MATHEMATICAL FOUNDATIONS FOR APPROXIMATING PARTICLE BEHAVIOUR AT RADIUS OF THE PLANCK LENGTH

Online Completion of Ill-conditioned Low-Rank Matrices

Risk tolerance and optimal portfolio choice

A Study on Non-Binary Turbo Codes

Numerical solution of fuzzy differential equations by Milne s predictor-corrector method and the dependency problem

Variance and Covariance Processes

Extremal problems for t-partite and t-colorable hypergraphs

International Journal of Pure and Applied Sciences and Technology

NUMERICAL SIMULATION FOR NONLINEAR STATIC & DYNAMIC STRUCTURAL ANALYSIS

Servomechanism Design

Fig. 1S. The antenna construction: (a) main geometrical parameters, (b) the wire support pillar and (c) the console link between wire and coaxial

Lecture 20: Riccati Equations and Least Squares Feedback Control

BMOA estimates and radial growth of B φ functions

336 ERIDANI kfk Lp = sup jf(y) ; f () jj j p p whee he supemum is aken ove all open balls = (a ) inr n, jj is he Lebesgue measue of in R n, () =(), f

Distribution Free Evolvability of Polynomial Functions over all Convex Loss Functions

Unsupervised Segmentation of Moving MPEG Blocks Based on Classification of Temporal Information

Synchronization of Fractional Chaotic Systems via Fractional-Order Adaptive Controller

Energy dispersion relation for negative refraction (NR) materials

EFFECT OF PERMISSIBLE DELAY ON TWO-WAREHOUSE INVENTORY MODEL FOR DETERIORATING ITEMS WITH SHORTAGES

Dual Hierarchies of a Multi-Component Camassa Holm System

[ ] 0. = (2) = a q dimensional vector of observable instrumental variables that are in the information set m constituents of u

The Production of Polarization

MIMO Cognitive Radio Capacity in. Flat Fading Channel. Mohan Premkumar, Muthappa Perumal Chitra. 1. Introduction

Measures the linear dependence or the correlation between r t and r t-p. (summarizes serial dependence)

Vehicle Arrival Models : Headway

IMPROVED DESIGN EQUATIONS FOR ASYMMETRIC COPLANAR STRIP FOLDED DIPOLES ON A DIELECTRIC SLAB

Dynamic Estimation of OD Matrices for Freeways and Arterials

Topic 4a Introduction to Root Finding & Bracketing Methods

( ) ( ) if t = t. It must satisfy the identity. So, bulkiness of the unit impulse (hyper)function is equal to 1. The defining characteristic is

Today - Lecture 13. Today s lecture continue with rotations, torque, Note that chapters 11, 12, 13 all involve rotations

A Shooting Method for A Node Generation Algorithm

A STOCHASTIC MODELING FOR THE UNSTABLE FINANCIAL MARKETS

On Energy-Efficient Node Deployment in Wireless Sesnor Networks

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients

Final Exam. Tuesday, December hours, 30 minutes

Support Vector Machines

Tracking Control for Hybrid Systems via Embedding of Known Reference Trajectories

Chapter 7. Interference

An Open cycle and Closed cycle Gas Turbine Engines. Methods to improve the performance of simple gas turbine plants

4.6 One Dimensional Kinematics and Integration

Chapter 2. First Order Scalar Equations

Chapter 3 Boundary Value Problem

On the Design of Optimal Zoning for Pattern Classification

AN EFFICIENT INTEGRAL METHOD FOR THE COMPUTATION OF THE BODIES MOTION IN ELECTROMAGNETIC FIELD

d 1 = c 1 b 2 - b 1 c 2 d 2 = c 1 b 3 - b 1 c 3

envionmen ha implemens all of he common algoihmic deails of all nodal mehods u pemis he specic mehod o e used in any concee insance o e specied y he u

156 There are 9 books stacked on a shelf. The thickness of each book is either 1 inch or 2

( ) exp i ω b ( ) [ III-1 ] exp( i ω ab. exp( i ω ba

LAPLACE TRANSFORM AND TRANSFER FUNCTION

Turbo-Like Beamforming Based on Tabu

CHAPTER 10 VALIDATION OF TEST WITH ARTIFICAL NEURAL NETWORK

arxiv: v2 [stat.me] 13 Jul 2015

Q & Particle-Gas Multiphase Flow. Particle-Gas Interaction. Particle-Particle Interaction. Two-way coupling fluid particle. Mass. Momentum.

Numerical Integration

ENGI 4430 Advanced Calculus for Engineering Faculty of Engineering and Applied Science Problem Set 9 Solutions [Theorems of Gauss and Stokes]

POSITIVE SOLUTIONS WITH SPECIFIC ASYMPTOTIC BEHAVIOR FOR A POLYHARMONIC PROBLEM ON R n. Abdelwaheb Dhifli

Sections 2.2 & 2.3 Limit of a Function and Limit Laws

Chapter 7: Solving Trig Equations

12: AUTOREGRESSIVE AND MOVING AVERAGE PROCESSES IN DISCRETE TIME. Σ j =

MATH 128A, SUMMER 2009, FINAL EXAM SOLUTION

Kalman Filter: an instance of Bayes Filter. Kalman Filter: an instance of Bayes Filter. Kalman Filter. Linear dynamics with Gaussian noise

Engineering Accreditation. Heat Transfer Basics. Assessment Results II. Assessment Results. Review Definitions. Outline

System of Linear Differential Equations

Pressure Vessels Thin and Thick-Walled Stress Analysis

Transcription:

AN EVOLUTIONARY APPROACH FOR SOLVING DIFFERENTIAL EQUATIONS M. KAMESWAR RAO AND K.P. RAVINDRAN Depamen of Mechanical Engineeing, Calicu Regional Engineeing College, Keala-67 6, INDIA. Absac:- We eploe he possibiliy of solving diffeenial equaions by using evoluionay algoihm. The ial soluion is consuced by incopoaing a Neual Newok componen and a non-neual pa so as o saisfy he iniial/bounday condiions. The effecive minimisaion of he eo funcion gives he closed analyic fom of he soluion fo he given diffeenial equaion wih specified bounday/iniial condiions. We y o minimise he eo funcion by using he Evoluionay algoihm. Compaison of he soluions hus obained and he soluion obained analyically is done. The pesen wok eveals he abiliy of Evoluionay algoihms fo solving Diffeenial equaions. The pesence of he Neual Newok componen makes he soluion hus obained amenable o design on hadwae. Key wods:- Neual Newoks, Diffeenial Equaions, Evoluionay algoihm, iniial/bounday condiions.. Inoducion The mahemaical analysis of any physical sysem is done by fomulaing he Diffeenial equaion govening he sysem wih he appopiae iniial/bounday condiions and solving i. As sysems become comple, i is highly difficul and impossible o obain a closed fom soluion fo he govening equaion [GE]. Thus he compuaional scheme employed becomes he conesone in he analysis of he physical sysem. We can a he mos obain he soluion a discee poins by appoimaing he ems in he GE by Taylo seies, o obain a sysem of equaions. By solving hese equaions, he soluion a discee poins of he domain is obained. Also, by he disceizaion of he domain ino elemens and using appoimaion funcions fo he soluion in each elemen, he soluion a he nodal values can be found, fom which he values a any poin inside he elemens can be found by using inepolaion mehods. All hese mehods have he shofall of he lack of closed fom soluion and limied diffeeniabiliy. In he pevious wok employing Neual Newoks fo solving diffeenial equaions, soluions wee obained by consucing he ial soluion using neual newoks o saisfy he iniial/bounday condiions. The eo funcion ove a numbe of poins in he domain was obained by he collocaion mehod and his eo funcion was minimised employing gadien-based mehods []. Also, Diffeenial equaions wee solved by disceizaion of he poblem domain o obain a sysem of equaions. The soluion of his sysem of equaions is mapped ono a Hopfield neual newok. The effecive minimisaion of he enegy funcion being he opimisaion poblem ha gives he equied soluion [-]. Anohe appoach fo solving diffeenial equaions using Neual Newoks was o employ B -splines as basis funcions and mapping he soluion fom of he coefficiens of hese ono a feed fowad Neual Newok. Hee, each spline is eplaced wih he sum of piecewise linea acivaion funcions ha coespond o hidden unis [4-5]. This mehod consides local basis funcions and in geneal equies many splines in ode o yield accuae soluions. Moeove, i is no easy o eend his echnique o muli-dimensional domains. In he pesen wok he minimisaion of he eo funcion poposed in [] is minimised using Evoluionay Algoihm. The minimisaion of he eo funcion is iself he aining of he Neual Newok. As he gadien descen mehods have he endency o ge apped in he local minima, he evoluionay algoihm is used in he opimisaion of he eo funcion and hence he aining of he Neual Newok. The pobabiliy of he Evoluionay Algoihm o convege o global minima is moe han convenional calculus based echniques. The usage of evoluionay algoihm may enhance he effeciveness of he mehod of solving diffeenial equaions by using Neual Newoks. In addiion o he

feaues menioned in [], he poposed appoach will ensue a bee convegence o global opima. Descipion of he mehod The mehod of consucion of ial soluion is eplained wih he following geneal diffeenial equaion poposed in []. G(, Ψ(), Ψ(), Ψ()) =, D () subjec o he bounday condiions (BC s), eihe Diichle o Neumann, whee n n = (,,,...n ) R,D R denoes he definiion of he domain and Ψ ( ) is he soluion o be compued. Applicaion of he collocaion mehod educes he poblem o G(i, Ψ(i ), Ψ(i ), Ψ(i )) =, i ^ D () subjec o he consains imposed by he BC s. Le Ψ (,p) denoes a ial soluion wih adjusable paamees p, and hen he poblem is ansfomed o Min (G(i, Ψ (i,p), Ψ (i,p), Ψ (i,p))) ^ i D () consained by he BC s. The ial soluion Ψ (,p) employs a feed fowad Neual Newok and he paamees p coespond o he weighs and biases of he neual achiecue [6],[7]. The ial fom of he soluion by consucion saisfies he bounday condiions, which is achieved by wiing i as sum of wo ems Ψ () = A() + F(, N(,p)) (4) whee N(,p ) is single oupu feed fowad neual newok wih paamees p and n inpu unis fed wih he inpu veco. The em A( ) conains no adjusable paamees and saisfies he bounday condiions. The second em is consuced so as no o conibue o he BC s, since Ψ () mus also saisfy hem. The weighs and biases of he Neual Newok in his em ae o be adjused and his consiues he opimisaion poblem. As he ial soluion is consuced so as o saisfy he BC s, he opimisaion poblem becomes an unconsained opimisaion poblem. The employmen of Gadien mehods [] may make he soluion o convege a a local minimum insead of he global minimum. Also he minimisaion using Gadien mehod equies he compuaion of he Gadien of he eo funcion wih espec o he newok paamees and an efficien opimisaion package as well. On he ohe hand, he employmen of he Evoluionay algoihm doesn equie Gadien compuaion and is vey easy o implemen. I has demonsaed good ageemen wih he analyic soluion.. Evoluionay Algoihm The Evoluionay Algoihm elies on he pinciples of naual selecion and suvival of he fies. I sas he seach pocess fom a collecion of poins ahe han a single poin. I doesn equie any auiliay knowledge like Gadien of he eo funcion fo is seach, bu insead, needs he value of eo funcion alone [8]. Evoluionay Algoihms ae usually applied o assign values ha allow a paicula pefomance o finess o he paamees of a sysem. An Evoluionay Algoihm begins wih a populaion of seveal samples of he same sysem wih vaiables assigned andomly o abiaily [9]. The diffeen samples undego a se of ials in ode o compae hei pefomances wih he equied one. In his way, each sample eceives an evaluaion. Samples ha have obained he bes finess ae hen seleced and epoduced wih andom cossove and muaion in ode o compose a new geneaion. The cossove pocess consiss of wo seps. Fis, selecing he newly epoduced sings andomly and hen geneaing new sings by swapping all chaaces fom andomly seleced cossove sies. The muaion opeaion consiss of copying he values of vaiables by adding lile andom changes. The new geneaion undegoes he same se of ials. The pocess is coninued unil desied accuacy of he soluion is obained. In he pesen wok we use Evoluionay Algoihm fo he minimisaion of he unconsained eo funcion. The vaiables fo he unconsained opimisaion poblem ae he weighs and biases of he Neual Newok. In he pesen mehod we used eal numbe coding, as he numbe of vaiables is lage and i is difficul o do binay coding. Fom epeimens i was found ha he values of he weighs ange fom o +, he iniial populaion was geneaed accodingly. An appopiae value fo he eo funcion was chosen as he cieion fo eminaion.

The paamees of he Evoluionay Algoihm used ae shown in Table. Ψ a / e = + + + (8) Populaion size Sing lengh Numbe of cossove sies 6 The Neual Newok achiecue fo OD s is shown in Fig. and he paamees used in he Newok ae shown in Table. Type of coding Real numbe coding Objecive funcion Minimising he eo funcion Vaiables Weighs of he inpu o hidden laye, hidden laye o he oupu laye and biases in he hidden laye I/P O/P Table Evoluionay algoihm paamees Illusaion wih eamples The following eamples illusae he mehod and compaison of he soluion wih he analyic soluion.. Odinay Diffeenial Equaions In his secion he Odinay Diffeenial Equaions ae deal... Poblem The Diffeenial Equaion and Bounday condiion ae: dψ + + + d + + = Ψ + + + + + BC s: Ψ ()=, [, ] (5) The ial soluion can be wien as, Ψ () = A() + F(, N(,p)) (6) Fo his poblem his educes o Ψ () = +.N(,p) (7) The ue soluion (analyic) is, Fig. Neual Newok Achiecue fo ODE s Numbe of layes Numbe of inpu nodes Numbe of hidden nodes Type of squashing funcion Numbe of oupu nodes Sigmoid funcion /(+e - ) Table Paamees used in Neual Newok The numbe of aining poins aken was en, wih unifom spacing in he poblem domain. Fig. shows accuacy of he soluion obained by he Evoluionay Algoihm as compaed wih he analyic soluion, Hidden laye

while he plo of analyical soluion is given in Fig.. The soluion accuacy obained hee is compaable wih he esuls in []. soluion, while he plo of analyical soluion is given in Fig.5. soluion accuacy.7.6.5.4... -...4.6.8..4 -. f() 6 5 4.5.5 Fig. Soluion accuacy fo Poblem Fig.4 Plo of Analyical soluion fo Poblem II f().5.5.5.5 Fig. Plo of analyical soluion fo Poblem... Poblem-II This is anohe Odinay Diffeenial Equaion and he Bounday condiion used in his poblem ae: dy = d y soluion accuacy.5.4....5. -..45.6.75 Fig.5 Soluion accuacy fo Poblem II. Paial Diffeenial Equaions In his secion he soluion of paial diffeenial equaions (PDE s) ae discussed. Fig.6 shows he Neual Newok achiecue used fo PDE,s..9 B.C. a y () = ; (9) The ue soluion fo his poblem is, ( ) 4e + ya = ( ); () The ial fom of he soluion used in his eample is: y = + ( )N(,p); () Fig.4 shows accuacy of he soluion obained by Evoluionay Algoihm as compaed wih he analyic I/P weighs biases O/P weighs Fig.6 Neual Newok Achiecue fo PDE s

.. Poblem III u u + = y [,] y [,] B.c s ae u (,) = ; u (,y) = ; u (,y) = ; u(,) = sin( π) () Analyic soluion fo his poblem is u a = sin( π).sin( πy) / sinh( π) () The ial soluion o his poblem can be wien as u = y sin( π) + ( )( y)yn (4) This is he govening equaion fo a hea ansfe poblem fo a fla plae wih fou Diichle BC s. The soluion accuacy of he poblem is ploed fo poins in he poblem domain. The accuacy of he soluion is checked wih he analyic soluion of he poblem. Fig.7 shows he plo of soluion accuacy a vaious poins in he domain. The plo is made by aking poins in he domain. The X ais in he plo epesens he poins on X dimension of he poblem domain and Y ais epesens he poins on Y dimension of he poblem domain. The Z ais epesens he soluion accuacy, which is he diffeence beween he analyic soluion and he soluion obained by he Neual Newok mehod.. Poblem IV ϕ(,y) ϕ(,y) + ϕ(,y) = y [,] sin( π)( π y [,] y + y sin( π); B.C. s ae ϕ (,y) =, ϕ (,y) =, ϕ (,) =, ϕ(,)/ y = sin( π) (5) The ial soluion is, ϕ = B(, y) + ( )y[(n(, y,p) The analyical soluion o his poblem is, N(,, p) N(,, p) (6) ] y ϕ = sin( π) (7) a y This is a non-linea paial diffeenial equaion wih boh Diichle and Neumann bounday condiions. The plo of soluion accuacy vesus vaious poins in he domain is shown in Fig. 8. The plo of he soluion accuacy a vaious poins in he domain is shown simila o ha of he pevious poblems. The Z ais epesen he soluion accuacy, which is he diffeence beween he ue soluion and he soluion obained by he Neual Newok mehod. Fom he plo i is eviden ha he soluion accuacy a he boundaies is zeo because of he choice of he ial soluion. The accuacy a ohe poins is posiive as well as negaive. Fig.7 Soluion accuacy fo Poblem III Fig.8 Soluion accuacy fo Poblem IV

4 CONCLUSIONS AND SCOPE The applicabiliy of Evoluionay Algoihm fo solving Diffeenial Equaions wih ial soluions having a Neual Newok is eploed. The compaison of he soluions hus obained, wih he analyic soluion showed good ageemen. I should be noed ha ou epeimens wee caied on an odinay PC wih Penium III pocesso wih 56MB RAM. The mehod is o be fuhe modified o apply o Paial Diffeenial Equaions wih iegula domain []. Fuue wok involves compaison of soluions hus obained, wih his mehod and he one s employing Gadien Algoihm as well as Finie Elemen Mehod. Refeences: [] I.E.Lagais, A.Likas and D.I.Foiadis, Aificial Neual Newoks fo Solving Odinay and Paial Diffeenial Equaions, IEEE Tansacions on Neual Newoks, Vol.9, No.5, 998, pp 987-. [] H.Lee and I.Kang, Neual Algoihms fo solving Diffeenial Equaions, J.Compu.Phys., Vol.9, 99, pp. -7. [] L.Wang and J.M. Mendel, Sucued ainable newoks fo mai algeba, IEEE In. Join Conf. Neual Newoks, Vol., 99, pp.5-8. [4] A.J.Meade, J., and A.A.Fenandez, The numeical soluion of linea odinay diffeenial equaions by feedfowad neual newoks, Mah.Compu. Modelling, vol.9, no., 994, pp -5. [5] A.J.Meade, J., and A.A.Fenandez, Soluion of nonlinea odinay diffeenial equaions by feedfowad neual newoks, Mah.Compu.Modelling, vol., no.9, 994, pp. 9-44. [6] David.E.Goldbeg, Geneic Algoihms in Seach, Opimizaion & Machine Leaning,, Addison Wesley, 999. [7] William. H.Pess e al, Numeical Recipes in C, nd ediion, Cambidge Univesiy Pess, 99. [8] Simon Haykin, Neual Newoks A compehensive foundaion, nd ed., Penice Hall Inenaional, Inc.,999. [9] N.K.Bose, P.Liang, Neual Newok Fundamenals wih Gaphs, Algoihms and Applicaions, Taa McGaw Hill, 998 [] I.E.Lagais, A.Likas and D.I.Foiadis, Neual Newok Mehods fo Bounday Value Poblems wih Iegula Boundaies, IEEE Tansacions on Neual Newoks, Vol., No.5,, pp. 4-49.