Identification of MIMO systems using radial basis function networks with hybrid learning algorithm

When a radial basis function network (RBFN) is used for identification of a nonlinear multi-input multi-output (MIMO) system, the number of hidden layer nodes, the initial parameters of the kernel, and the initial weights of the network must be determined first. For this purpose, a systematic way that integrates the support vector regression (SVR) and the least squares regression (LSR) is proposed to construct the initial structure of the RBFN. The first step of the proposed method is to determine the number of hidden layer nodes and the initial parameters of the kernel by the SVR method. Then the weights of the RBFN are determined by solving a simple minimization problem based on the concept of LSR. After initialization, an annealing robust learning algorithm (ARLA) is then applied to train the RBFN. With the proposed initialization approach, one can find that the designed RBFN has few hidden layer nodes while maintaining good performance. To show the feasibility and superiority of the annealing robust radial basis function networks (ARRBFNs) for identification of MIMO systems, several illustrative examples are included.     

Asymptotic coefficients for Gaussian radial basis function interpolants

Gaussian radial basis functions (RBFs) on an infinite interval with uniform grid pacing h are defined by ?(x;α,h)≡exp(-[α²/h²]x²). The only significant numerical parameter is α, the inverse width of the RBF functions relative to h. In the limit α→0, we demonstrate that the coefficients of the interpolant of a typical function f(x) grow proportionally to exp(π²/[4α²]). However, we also show that the approximation to the constant f(x)≡1 is a Jacobian theta function whose coefficients do not blow up as α→0. The subtle interplay between the complex-plane singularities of f(x) (the function being approximated) and the RBF inverse width parameter α are analyzed. For α≈1/2, the size of the RBF coefficients and the condition number of the interpolation matrix are both no larger than O(10⁴) and the error saturation is smaller than machine epsilon, so this α is the center of a “safe operating range” for Gaussian RBFs.     

Numerical solution of the two-dimensional Fredholm integral equations using Gaussian radial basis function

In this paper, we introduce a numerical method for the solution of two-dimensional Fredholm integral equations. The method is based on interpolation by Gaussian radial basis function based on Legendre-Gauss-Lobatto nodes and weights. Numerical examples are presented and results are compared with the analytical solution to demonstrate the validity and applicability of the method.     

Rotorcraft parameter estimation using radial basis function neural network

Increased emphasis on rotorcraft performance and operational capabilities has resulted in accurate computation of aerodynamic stability and control parameters. System identification is one such tool in which the model structure and parameters such as aerodynamic stability and control derivatives are derived. In the present work, the rotorcraft aerodynamic parameters are computed using radial basis function neural networks (RBFN) in the presence of both state and measurement noise. The effect of presence of outliers in the data is also considered. RBFN is found to give superior results compared to finite difference derivatives for noisy data.     

Boundary integral equations on the sphere with radial basis functions: error analysis

Radial basis functions are used to define approximate solutions to boundary integral equations on the unit sphere. These equations arise from the integral reformulation of the Laplace equation in the exterior of the sphere, with given Dirichlet or Neumann data, and a vanishing condition at infinity. Error estimates are proved. Numerical results supporting the theoretical results are presented.     

On the solution of the non-local parabolic partial differential equations via radial basis functions

In this paper, the problem of solving the one-dimensional parabolic partial differential equation subject to given initial and non-local boundary conditions is considered. The approximate solution is found using the radial basis functions collocation method. There are some difficulties in computing the solution of the time dependent partial differential equations using radial basis functions. If time and space are discretized using radial basis functions, the resulted coefficient matrix will be very ill-conditioned and so the corresponding linear system cannot be solved easily. As an alternative method for solution, we can use finite-difference methods for discretization of time and radial basis functions for discretization of space. Although this method is easy to use but an accurate solution cannot be provided. In this work an efficient collocation method is proposed for solving non-local parabolic partial differential equations using radial basis functions. Numerical results are presented and are compared with some existing methods.     

Parallel radial basis function methods for the global optimization of expensive functions

We introduce a master–worker framework for parallel global optimization of computationally expensive functions using response surface models. In particular, we parallelize two radial basis function (RBF) methods for global optimization, namely, the RBF method by Gutmann [Gutmann, H.M., 2001a. A radial basis function method for global optimization. Journal of Global Optimization 19(3), 201–227] (Gutmann-RBF) and the RBF method by Regis and Shoemaker [Regis, R.G., Shoemaker, C.A., 2005. Constrained global optimization of expensive black box functions using radial basis functions, Journal of Global Optimization 31, 153–171] (CORS-RBF). We modify these algorithms so that they can generate multiple points for simultaneous evaluation in parallel. We compare the performance of the two parallel RBF methods with a parallel multistart derivative-based algorithm, a parallel multistart derivative-free trust-region algorithm, and a parallel evolutionary algorithm on eleven test problems and on a 6-dimensional groundwater bioremediation application. The results indicate that the two parallel RBF algorithms are generally better than the other three alternatives on most of the test problems. Moreover, the two parallel RBF algorithms have comparable performances on the test problems considered. Finally, we report good speedups for both parallel RBF algorithms when using a small number of processors.     

Rational radial basis function interpolation with applications to antenna design

Functions with poles occur in many branches of applied mathematics which involve resonance phenomena. Such functions are challenging to interpolate, in particular in higher dimensions. In this paper we develop a technique for interpolation with quotients of two radial basis function (RBF) expansions to approximate such functions as an alternative to rational approximation. Since the quotient is not uniquely determined we introduce an additional constraint, the sum of the RBF-norms of the numerator and denominator squared should be minimal subjected to a norm condition on the function values. The method was designed for antenna design applications and we show by examples that the scattering matrix for a patch antenna as a function of some design parameters can be approximated accurately with the new method. In many cases, e.g. in antenna optimization, the function evaluations are time consuming, and therefore it is important to reduce the number of evaluations but still obtain a good approximation. A sensitivity analysis of the new interpolation technique is carried out and it gives indications how efficient adaptation methods could be devised. A family of such methods are evaluated on antenna data and the results show that much performance can be gained by choosing the right method.     

Particle swarm with radial basis function surrogates for expensive black-box optimization

This paper develops the OPUS (Optimization by Particle swarm Using Surrogates) framework for expensive black-box optimization. In each iteration, OPUS considers multiple trial positions for each particle in the swarm and uses a surrogate model to identify the most promising trial position. Moreover, the current overall best position is refined by finding the global minimum of the surrogate in the neighborhood of that position. OPUS is implemented using an RBF surrogate and the resulting OPUS-RBF algorithm is applied to a 36-D groundwater bioremediation problem, a 14-D watershed calibration problem, and ten mostly 30-D test problems. OPUS-RBF is compared with a standard PSO, CMA-ES, two other surrogate-assisted PSO algorithms, and an RBF-assisted evolution strategy. The numerical results suggest that OPUS-RBF is promising for expensive black-box optimization.     

Preconditioning for radial basis functions with domain decomposition methods

In our previous work, an effective preconditioning scheme that is based upon constructing least-squares approximation cardinal basis functions (ACBFs) from linear combinations of the RBF-PDE matrix elements has shown very attractive numerical results. The preconditioner costs O(N²) flops to set up and O(N) storage. The preconditioning technique is sufficiently general that it can be applied to different types of different operators. This was applied to the 2D multiquadric method, with c~1/√N on the Poisson test problem, the preconditioned GMRES converges in tens of iterations. In this paper, we combine the RBF methods and the ACBF preconditioning technique with the domain decomposition method (DDM). We studied different implementations of the ACBF-DDM scheme and provide numerical results for N > 10,000 nodes. We shall demonstrate that the efficiency of the ACBF-DDM scheme improves dramatically as successively finer partitions of the domain are considered.     

Error estimates for matrix-valued radial basis function interpolation

We introduce a class of matrix-valued radial basis functions (RBFs) of compact support that can be customized, e.g. chosen to be divergence-free. We then derive and discuss error estimates for interpolants and derivatives based on these matrix-valued RBFs.     

Using radial basis functions to construct local volatility surfaces

This article presents a new method for constructing a volatility surface for use in local volatility option pricing models. It builds on previous work focussing on non-parametric regression approaches using a set of radial basis functions, specifically thin plate splines. Optimal parameters are found using a trust region optimisation approach. While there is still much work to be done, the results are encouraging and show that the method is relatively tractable, stable and accurate.     

Some issues on interpolation matrices of locally scaled radial basis functions

Radial basis function interpolation on a set of scattered data is constructed from the corresponding translates of a basis function, which is conditionally positive definite of order m ? 0, with the possible addition of a polynomial term. In many applications, the translates of a basis function are scaled differently, in order to match the local features of the data such as the flat region and the data density. Then, a fundamental question is the non-singularity of the perturbed interpolation (N × N) matrix. In this paper, we provide some counter examples of the matrices which become singular for N ? 3, although the matrix is always non-singular when N = 2. One interesting feature is that a perturbed matrix can be singular with rather small perturbation of the scaling parameter.     

Analysis of stationary subdivision schemes for curve design based on radial basis function interpolation

This paper provides a large family of interpolatory stationary subdivision schemes based on radial basis functions (RBFs) which are positive definite or conditionally positive definite. A radial basis function considered in this study has a tension parameter λ>0 such that it provides design flexibility. We prove that for a sufficiently large , the proposed 2L-point (L∈N) scheme has the same smoothness as the well-known 2L-point Deslauriers-Dubuc scheme, which is based on 2L-1 degree polynomial interpolation. Some numerical examples are presented to illustrate the performance of the new schemes, adapting subdivision rules on bounded intervals in a way of keeping the same smoothness and accuracy of the pre-existing schemes on R. We observe that, with proper tension parameters, the new scheme can alleviate undesirable artifacts near boundaries, which usually appear to interpolatory schemes with irregularly distributed control points.     

A mended hybrid learning algorithm for radial basis function neural networks to improve generalization capability

A two-step learning scheme for radial basis function neural networks, which combines the genetic algorithm (GA) with the hybrid learning algorithm (HLA), is proposed in this paper. It is compared with the methods of the GA, the recursive orthogonal least square algorithm (ROLSA) and another two-step learning scheme for RBF neural networks, which combines the K-means clustering with the HLA (K-means + HLA). Our proposed method has the best generalization performance.     

Numerical solutions of KdV equation using radial basis functions

Numerical solution of the Korteweg-de Vries equation is obtained by using the meshless method based on the collocation with radial basis functions. Five standard radial basis functions are used in the method of the collocation. The results are compared for the numerical experiments of the propagation of solitons, interaction of two solitary waves and breakdown of initial conditions into a train of solitons.     

A low-cost-memory CUDA implementation of the conjugate gradient method applied to globally supported radial basis functions implicits

Hermitian radial basis functions implicits is a method capable of reconstructing implicit surfaces from first-order Hermitian data. When globally supported radial functions are used, a dense symmetric linear system must be solved. In this work, we aim at exploring and computing a matrix-free implementation of the Conjugate Gradients Method on the GPU in order to solve such linear system. The proposed method parallelly rebuilds the matrix on demand for each iteration. As a result, it is able to compute the Hermitian-based interpolant for datasets that otherwise could not be handled due to the high memory demanded by their linear systems.     

Trigonometric variable shape parameter and exponent strategy for generalized multiquadric radial basis function approximation

The generalized multiquadric radial basis function (φ_j=[(x-x_j)²+c²]^β) has the exponent β and shape parameter c that play an important role in the accuracy of the approximation. In this study, we present a trigonometric variable shape parameter and exponent strategy and apply it to function interpolations and linear boundary value problems. Several numerical experiments with the uniformly spaced nodes show that the inverse multiquadric radial basis function (β = −0.5) with the trigonometric variable shape parameter c strategy results in the best accuracy for the one-dimensional interpolations; the trigonometric variable shape parameters and exponent strategy produces the best accuracy for the two-dimensional interpolations and linear boundary value problems. For the non-uniformly spaced nodes, the random variable shape parameter c and exponent β strategy produces the best accuracy for the two-dimensional boundary value problem.     

On lower bounds in radial basis approximation

We investigate the approximation by manifolds _n() generated by linear combinations of n radial basis functions on R^d of the form (|–a|), where is the thin-plate spline type function. We obtain exact asymptotic estimates for the approximation of Sobolev classes W^r(B^d) in the space L(B^d) on the unit ball B^d. AMS subject classification 41A25, 41A63, 65D07, 41A15