Abstract:This paperproposes arecursively adapted radial basis function networkandprovides additional insights into several well-known techniques suchasradial basis function networks, resource allocating networks andstochastic gradient descent inreproducing kernel Hilbert spaces. Throughthis perspective, resource allocating networks areinvestigated ina moreprincipled way so thatissues ofconvergence and generalization canbemathematically analyzed intheleast meansquare framework. I.INTRODUCTION RESOURCE allocating networks (RAN)areanintuitive, veryappealing ideaintroduced byPlatt (1)butits convergence andgeneralization havenotbeenfully addressed mathematically. Recently, Kivinen, SmolaandWilliamson (2)introduced anonline kernel learning method(OKL)whichcanbe simply viewedasa stochastic gradient descent method applied ontheregularized costfunction ofaradial basis function network (RBFN). Interestingly, OKL bears much resemblance withRAN intermsoflearning strategy and network architecture. However, there isonefundamental difference. OKL employs aforgetting factor initstraining, whichisthelegacy oftheregularization termintheRBFN costfunction, toimpose stability inlearning andthus ensure generalization. Further statistical analysis concludes that this forgetting factor appears essential forits well-posedness (3). Therefore, this reasoning raises thequestion ifRAN iswell- posed. Generalization hasbeensubstantially studied inthe framework ofregularization theories (4), (5). Regularization asaremedy forill-posedness became widely knowndueto theworkofTikhonov (6). Insolving RBFN like problems, theTikhonov regularization isessentially a trade-off between fitting training dataandreducing solution norms. Consequently itreduces thesensitivity ofthesolution to small changes inthetraining dataandimposes stability on theill-posed problem. Ourworkimproves onrelated research inseveral ways: 1.We propose arecursively adapted RBFNandprovide the convergence conditions.

On Radial Basis Function Net and Kernel Regression: Approximation Ability, Convergence Rate and Receptive Field Size

Kernel Regression and Radial Basis Function Net: Some Theoretical Studies

On L/sub 2/ Convergence Rates of Radial Basis Function Networks and Kernel Regression Estimators

Some Results on L/sub 1/ Convergence Rate of RBF Networks and Kernel Regression Estimators

Almost optimal estimates for approximation and learning by radial basis function networks

A Novel Radial Basis Function Neural Network For Approximation

Universal Approximation Using Radial-Basis-Function Networks

On the universal approximation property of radial basis function neural networks

Approximation Capability to Functions of Several Variables, Nonlinear Functionals and Operators by Radial Basis Function Neural Networks

Critical Vector Learning for Rbf Networks

A Novel Rbf Neural Network With Fast Training And Accurate Generalization

Recursively AdaptedRadial Basis Function Networks Andits Relationship Toresource Allocating Networks Andonline Kernel Learning

Radial basis function neural networks: a topical state-of-the-art survey

A MATLAB-Based Study on the Realization and Approximation Performance of RBF Neural Networks

Frequency Domain Analysis Based Rbf Networks And Their Applications To Function Approximations

Learning Algorithms for RBF Functions and Subspace Based Functions

Face recognition based on kernel radial basis function networks

L P Approximation Capability of RBF Neural Networks

On the numerical rank of radial basis function kernels in high dimension

Approximation of Function by Adaptively Growing Radial Basis Function Neural Networks.

Classification and Prediction of Neural Network Based on Generalized Radial Basis Function