Content deleted Content added
Complexica (talk | contribs) Undid revision 206073177 by Complexica (talk) |
Link suggestions feature: 3 links added. |
||
| (149 intermediate revisions by 95 users not shown) | |||
Line 1:
In the field of [[mathematical modeling]], a '''radial basis function network''' is an [[artificial neural network]] that uses [[radial basis function]]s as [[activation function]]s. The output of the network is a [[linear combination]] of radial basis functions of the inputs and neuron parameters. Radial basis function networks have many uses, including [[function approximation]], [[time series prediction]], [[Statistical classification|classification]], and system [[Control theory|control]]. They were first formulated in a 1988 paper by Broomhead and Lowe, both researchers at the [[Royal Signals and Radar Establishment]].<ref>{{cite tech report
|last1 = Broomhead
|first1 = D. S.
|last2 = Lowe
|first2 = David
|year = 1988
|title = Radial basis functions, multi-variable functional interpolation and adaptive networks
|institution = [[Royal Signals and Radar Establishment|RSRE]]
|number = 4148
|url = http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA196234
|archive-url = https://web.archive.org/web/20130409223044/http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA196234
|url-status = dead
|archive-date = April 9, 2013
}}</ref><ref>{{cite journal
|last1 = Broomhead
|first1 = D. S.
|last2 = Lowe
|first2 = David
|year = 1988
|title = Multivariable functional interpolation and adaptive networks
|journal = Complex Systems
|volume = 2
|pages = 321–355
|url = https://sci2s.ugr.es/keel/pdf/algorithm/articulo/1988-Broomhead-CS.pdf
|access-date = 2019-01-29
|archive-date = 2020-12-01
|archive-url = https://web.archive.org/web/20201201121028/https://sci2s.ugr.es/keel/pdf/algorithm/articulo/1988-Broomhead-CS.pdf
|url-status = live
}}</ref><ref name="schwenker"/>
==Network architecture==
[[File:Rbf-network.svg|thumb|252x252px|Architecture of a radial basis function network. An input vector <math>x</math> is used as input to all radial basis functions, each with different parameters. The output of the network is a linear combination of the outputs from radial basis functions.]]
Radial basis function (RBF) networks typically have three layers: an input layer, a hidden layer with a non-linear RBF activation function and a linear output layer. The input can be modeled as a vector of real numbers <math>\mathbf{x} \in \mathbb{R}^n</math>. The output of the network is then a scalar function of the input vector, <math> \varphi : \mathbb{R}^n \to \mathbb{R} </math>, and is given by
:<math>\varphi(\mathbf{x}) = \sum_{i=1}^N a_i \rho(||\mathbf{x}-\mathbf{c}_i||)</math>
where
|last1=Beheim|first1=Larbi
|last2=Zitouni|first2=Adel
|last3=Belloir|first3=Fabien
|date=January 2004
|title=New RBF neural network classifier with optimized hidden neurons number
|url=https://www.researchgate.net/publication/254467552
}}</ref><ref>{{cite conference
|conference=Proceedings of the Second Joint 24th Annual Conference and the Annual Fall Meeting of the Biomedical Engineering Society
|conference-url=https://ieeexplore.ieee.org/xpl/conhome/8844528/proceeding
|___location=Houston, TX, USA
|last1=Ibrikci|first1=Turgay
|last2=Brandt|first2=M.E.
|last3=Wang|first3=Guanyu
|last4=Acikkar|first4=Mustafa
|date=23–26 October 2002
|publication-date=6 January 2003
|volume=3
|pages=2184–5
|doi=10.1109/IEMBS.2002.1053230
|title=Mahalanobis distance with radial basis function network on protein secondary structures
|isbn=0-7803-7612-9
|issn=1094-687X
}}</ref>{{Editorializing|date=May 2020}}<!-- Was previously marked with a missing-citation tag asking in what sense using Mahalanobis distance is better and why the Euclidean distance is still normally used, but I found sources to support the first part, so it's likely salvageable. -->) and the radial basis function is commonly taken to be [[Normal distribution|Gaussian]]
:<math> \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) = \exp \left[ -\
The Gaussian basis functions are local to the center vector in the sense that
:<math>\lim_{||x|| \to \infty}\rho(\left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert) = 0</math>
Line 19 ⟶ 69:
i.e. changing parameters of one neuron has only a small effect for input values that are far away from the center of that neuron.
Given certain mild conditions on the shape of the activation function, RBF networks are [[universal approximator]]s on a [[Compact space|compact]] subset of <math>\mathbb{R}^n</math>.<ref name="Park">{{cite journal|last=Park|first=J.|author2=I. W. Sandberg|s2cid=34868087|date=Summer 1991|title=Universal Approximation Using Radial-Basis-Function Networks|journal=Neural Computation|volume=3|issue=2|pages=246–257|doi=10.1162/neco.1991.3.2.246|pmid=31167308}}</ref> This means that
The
[[Image:Unnormalized radial basis functions.svg|thumb|
]]
===
{{multiple images
| align = right
| direction = vertical
| width = 250
| image1 = Normalized radial basis functions.svg
| caption1 = Two normalized radial basis functions in one input dimension ([[logistic function|sigmoids]]). The basis function centers are located at <math> c_1=0.75 </math> and <math> c_2=3.25 </math>.
| image2 = 3 Normalized radial basis functions.svg
| caption2 = Three normalized radial basis functions in one input dimension. The additional basis function has center at <math> c_3=2.75 </math>.
| image3 = 4 Normalized radial basis functions.svg
| caption3 = Four normalized radial basis functions in one input dimension. The fourth basis function has center at <math> c_4=0 </math>. Note that the first basis function (dark blue) has become localized.
}}
====Normalized architecture====
In addition to the above ''unnormalized'' architecture, RBF networks can be ''normalized''. In this case the mapping is
Line 33 ⟶ 95:
where
:<math> u \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) \ \stackrel{\mathrm{def}}{=}\ \frac { \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) } { \sum_{
is known as a
====Theoretical motivation for normalization====
There is theoretical justification for this architecture in the case of stochastic data flow. Assume a [[stochastic kernel]] approximation for the joint probability density
:<math> P\left ( \mathbf{x} \land y \right ) = {1 \over N} \sum_{i=1}^N \, \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) \, \sigma \big ( \left \vert y - e_i \right \vert \big )</math>
where the weights <math> \mathbf{c}_i </math> and <math> e_i </math> are exemplars from the data and we require the kernels to be normalized
Line 48 ⟶ 109:
:<math> \int \sigma \big ( \left \vert y - e_i \right \vert \big ) \, dy =1</math>.
The probability densities in the input and output spaces are
:<math> P \left ( \mathbf{x} \right ) = \int P \left ( \mathbf{x} \land y \right ) \, dy = {1 \over N} \sum_{i=1}^N \, \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big )</math>
and
:
The expectation of y given an input <math> \mathbf{x} </math> is
Line 63 ⟶ 123:
:<math> P\left ( y \mid \mathbf{x} \right ) </math>
is the conditional probability of y given <math> \mathbf{x} </math>.
The conditional probability is related to the joint probability through [[Bayes' theorem]]
:<math> P\left ( y \mid \mathbf{x} \right ) = \frac {P \left ( \mathbf{x} \land y \right )} {P \left ( \mathbf{x} \right )} </math>
Line 73 ⟶ 133:
This becomes
:<math> \varphi \left ( \mathbf{x} \right ) = \frac { \sum_{i=1}^N
when the integrations are performed.
Line 82 ⟶ 142:
:<math> \varphi \left ( \mathbf{x} \right ) = \sum_{i=1}^N \left ( a_i + \mathbf{b}_i \cdot \left ( \mathbf{x} - \mathbf{c}_i \right ) \right )\rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) </math>
and
:<math> \varphi \left ( \mathbf{x} \right ) = \sum_{i=1}^N \left ( a_i + \mathbf{b}_i \cdot \left ( \mathbf{x} - \mathbf{c}_i \right ) \right )u \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) </math>
Line 99 ⟶ 159:
:<math> v_{ij}\big ( \mathbf{x} - \mathbf{c}_i \big ) \ \stackrel{\mathrm{def}}{=}\ \begin{cases} \delta_{ij} \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) , & \mbox{if } i \in [1,N] \\ \left ( x_{ij} - c_{ij} \right ) \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) , & \mbox{if }i \in [N+1,2N] \end{cases} </math>
in the unnormalized case and in the normalized case.
Here <math> \delta_{ij} </math> is a [[Kronecker delta function]] defined as
:<math> \delta_{ij} = \begin{cases} 1, & \mbox{if }i = j \\ 0, & \mbox{if }i \ne j \end{cases} </math>.
==Training==
RBF networks are typically trained from pairs of input and target values <math>\mathbf{x}(t), y(t)</math>, <math>t = 1, \dots, T</math> by a two-step algorithm.
In the first step, the center vectors <math>\mathbf c_i</math> of the RBF functions in the hidden layer are chosen. This step can be performed in several ways; centers can be randomly sampled from some set of examples, or they can be determined using [[k-means clustering]]. Note that this step is [[unsupervised learning|unsupervised]].
The second step simply fits a linear model with coefficients <math>w_i</math> to the hidden layer's outputs with respect to some objective function. A common objective function, at least for regression/function estimation, is the least squares function:
:<math> K( \mathbf{w} ) \ \stackrel{\mathrm{def}}{=}\ \sum_{t=1}^T K_t( \mathbf{w} ) </math>
where
:<math> K_t( \mathbf{w} ) \ \stackrel{\mathrm{def}}{=}\ \big [ y(t) - \varphi \big ( \mathbf{x}(t), \mathbf{w} \big ) \big ]^2 </math>.
We have explicitly included the dependence on the weights. Minimization of the least squares objective function by optimal choice of weights optimizes accuracy of fit.
There are occasions in which multiple objectives, such as smoothness as well as accuracy, must be optimized. In that case it is useful to optimize a regularized objective function such as
:<math> H( \mathbf{w} ) \ \stackrel{\mathrm{def}}{=}\ K( \mathbf{w} ) + \lambda S( \mathbf{w} ) \ \stackrel{\mathrm{def}}{=}\ \sum_{t=1}^
where
:<math> S( \mathbf{w} ) \ \stackrel{\mathrm{def}}{=}\ \sum_{t=1}^
and
Line 134 ⟶ 189:
where optimization of S maximizes smoothness and <math> \lambda </math> is known as a [[regularization (machine learning)|regularization]] parameter.
A third optional [[backpropagation]] step can be performed to fine-tune all of the RBF net's parameters.<ref name="schwenker">{{cite journal
|last1 = Schwenker
|first1 = Friedhelm
|last2 = Kestler
|first2 = Hans A.
|last3 = Palm
|first3 = Günther
|title = Three learning phases for radial-basis-function networks
|journal = Neural Networks
|volume = 14
|issue = 4–5
|pages = 439–458
|year = 2001
|citeseerx = 10.1.1.109.312
|doi=10.1016/s0893-6080(01)00027-2
|pmid = 11411631
}}</ref>
===Interpolation===
RBF networks can be used to interpolate a function <math>y: \mathbb{R}^n \to \mathbb{R}</math> when the values of that function are known on finite number of points: <math>y(\mathbf x_i) = b_i, i=1, \ldots, N</math>. Taking the known points <math>\mathbf x_i</math> to be the centers of the radial basis functions and evaluating the values of the basis functions at the same points <math>g_{ij} = \rho(|| \mathbf x_j - \mathbf x_i ||)</math> the weights can be solved from the equation
:<math>\left[ \begin{matrix}
g_{11} & g_{12} & \cdots & g_{1N} \\
Line 155 ⟶ 228:
\end{matrix} \right]</math>
It can be shown that the interpolation matrix in the above equation is non-singular, if the points
:<math>\mathbf{w} = \mathbf{G}^{-1} \mathbf{b}</math>
where <math>G = (g_{ij})</math>.
===Function approximation===
Line 170 ⟶ 244:
====Pseudoinverse solution for the linear weights====
After the centers <math>c_i</math> have been fixed, the weights that minimize the error at the output
:<math>\mathbf{w} = \mathbf{G}^+ \mathbf{b}</math>,
where the entries of ''G'' are the values of the radial basis functions evaluated at the points <math>x_i</math>: <math>g_{ji} = \rho(||x_j-c_i||)</math>.
The existence of this linear solution means that unlike
====Gradient descent training of the linear weights====
Another possible training algorithm is [[gradient descent]]. In gradient descent training, the weights are adjusted at each time step by moving them in a direction opposite from the gradient of the objective function (thus allowing the minimum of the objective function to be found),
:<math> \mathbf{w}(t+1) = \mathbf{w}(t) - \nu \frac {d} {d\mathbf{w}} H_t(\mathbf{w}) </math>
Line 219 ⟶ 293:
==Examples==
===Logistic map===
The basic properties of radial basis functions can be illustrated with a simple mathematical map, the [[logistic map]], which maps the [[unit interval]] onto itself. It can be used to generate a convenient prototype data stream. The logistic map can be used to explore [[function approximation]], [[time series prediction]], and [[control theory]]. The map originated from the field of [[population dynamics]] and became the prototype for [[chaos theory|chaotic]] time series. The map, in the fully chaotic regime, is given by
:<math> x(t+1)\ \stackrel{\mathrm{def}}{=}\ f\left [ x(t)\right ] = 4 x(t) \left [ 1-x(t) \right ] </math>
Line 227 ⟶ 302:
where t is a time index. The value of x at time t+1 is a parabolic function of x at time t. This equation represents the underlying geometry of the chaotic time series generated by the logistic map.
Generation of the time series from this equation is the [[forward problem]]. The examples here illustrate the [[inverse problem]]; identification of the underlying dynamics, or fundamental equation, of the logistic map from exemplars of the time series. The goal is to find an estimate
:<math> x(t+1) = f \left [ x(t) \right ] \approx \varphi(t) = \varphi \left [ x(t)\right ] </math>
Line 234 ⟶ 309:
===Function approximation===
====Unnormalized radial basis functions====
Line 241 ⟶ 317:
:<math> \varphi ( \mathbf{x} ) \ \stackrel{\mathrm{def}}{=}\ \sum_{i=1}^N a_i \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) </math>
where
:<math> \rho \big ( \left \Vert \mathbf{x} - \mathbf{c}_i \right \Vert \big ) = \exp \left[ -\
Since the input is a [[Scalar (mathematics)|scalar]] rather than a [[Vector (
:<math> a_i (t+1) = a_i(t) + \nu \big [ x(t+1) - \varphi \big ( \mathbf{x}(t), \mathbf{w} \big ) \big ] \frac {\rho \big ( \left \Vert \mathbf{x}(t) - \mathbf{c}_i \right \Vert \big )} {\sum_{i=1}^N \rho^2 \big ( \left \Vert \mathbf{x}(t) - \mathbf{c}_i \right \Vert \big )} </math>
where the [[learning rate]] <math> \nu </math> is taken to be 0.3. The training is performed with one pass through the 100 training points. The [[Mean squared error|rms error]] is 0.15.
[[Image:Normalized basis functions.png|thumb|350px|right|Figure 8: Normalized basis functions. The Logistic map (blue) and the approximation to the logistic map (red) after one pass through the training set. Note the improvement over the unnormalized case.]]
Line 270 ⟶ 346:
:<math> a_i (t+1) = a_i(t) + \nu \big [ x(t+1) - \varphi \big ( \mathbf{x}(t), \mathbf{w} \big ) \big ] \frac {u \big ( \left \Vert \mathbf{x}(t) - \mathbf{c}_i \right \Vert \big )} {\sum_{i=1}^N u^2 \big ( \left \Vert \mathbf{x}(t) - \mathbf{c}_i \right \Vert \big )} </math>
where the [[learning rate]] <math> \nu </math> is again taken to be 0.3. The training is performed with one pass through the 100 training points. The [[Mean squared error|rms error]] on a test set of 100 exemplars is 0.084, smaller than the unnormalized error. Normalization yields accuracy improvement. Typically accuracy with normalized basis functions increases even more over unnormalized functions as input dimensionality increases.
[[
===Time series prediction===
Line 284 ⟶ 360:
:<math> {x}(t+1) \approx \varphi(t)=\varphi [\varphi(t-1)]</math>.
A comparison of the actual and estimated time series is displayed in the figure. The estimated times series starts out at time zero with an exact knowledge of x(0). It then uses the estimate of the dynamics to update the time series estimate for several time steps.
Note that the estimate is accurate for only a few time steps. This is a general characteristic of chaotic time series. This is a property of the sensitive dependence on initial conditions common to chaotic time series. A small initial error is amplified with time. A measure of the divergence of time series with nearly identical initial conditions is known as the [[Lyapunov exponent]].
Line 290 ⟶ 366:
===Control of a chaotic time series===
[[
We assume the output of the logistic map can be manipulated through a control parameter <math> c[ x(t),t] </math> such that
:<math> {x}^{ }_{ }(t+1) = 4 x(t) [1-x(t)] +c[x(t),t] </math>.
The goal is to choose the control parameter in such a way as to drive the time series to a desired output <math> d(t) </math>. This can be done if we choose the control
:<math> c^{ }_{ }[x(t),t] \ \stackrel{\mathrm{def}}{=}\ -\varphi [x(t)] + d(t+1) </math>
where
:<math> y[x(t)] \approx f[x(t)] = x(t+1)- c[x(t),t] </math>
Line 314 ⟶ 390:
==See also==
* [[Radial basis function kernel]]
* [[instance-based learning]]
* [[In Situ Adaptive Tabulation]]
* [[Predictive analytics]]
* [[Chaos theory]]
* [[Hierarchical RBF]]
* [[Cerebellar model articulation controller]]
* [[Instantaneously trained neural networks]]
* [[Support vector machine]]
==References==
{{reflist}}
==Further reading==
* J. Moody and C. J. Darken, "Fast learning in networks of locally tuned processing units," Neural Computation, 1, 281-294 (1989). Also see [https://web.archive.org/web/20070302175857/http://www.ki.inf.tu-dresden.de/~fritzke/FuzzyPaper/node5.html Radial basis function networks according to Moody and Darken]
* T. Poggio and F. Girosi, "[http://courses.cs.tamu.edu/rgutier/cpsc636_s10/poggio1990rbf2.pdf Networks for approximation and learning]," Proc. IEEE 78(9), 1484-1487 (1990).
* Roger D. Jones, Y. C. Lee, C. W. Barnes, G. W. Flake, K. Lee, P. S. Lewis, and S. Qian, [https://ieeexplore.ieee.org/Xplore/home.jsp;jsessionid=1BAA8854614AFC21D2C29CDB4FC7DBEB Function approximation and time series prediction with neural networks], Proceedings of the International Joint Conference on Neural Networks, June 17–21, p. I-649 (1990).
* {{cite book | author=
*
* {{cite book|first1=John R.|last1=Davies|first2=Stephen V.|last2=Coggeshall |author3-link=Roger Jones (physicist and entrepreneur)|first3=Roger D.|last3=Jones|first4= Daniel|last4=Schutzer|contribution=Intelligent Security Systems|editor1-last=Freedman|editor1-first=Roy S.|editor2-last= Flein|editor2-first= Robert A.|editor3-last= Lederman|editor3-first= Jess | title=Artificial Intelligence in the Capital Markets | ___location= Chicago | publisher=Irwin| year=1995 | isbn=1-55738-811-3}}
* {{cite book | author=Simon Haykin | title=Neural Networks: A Comprehensive Foundation | edition=2nd | ___location=Upper Saddle River, NJ | publisher=Prentice Hall| year=1999 | isbn=0-13-908385-5}}
* S. Chen, C. F. N. Cowan, and P. M. Grant, "[https://eprints.soton.ac.uk/251135/1/00080341.pdf Orthogonal Least Squares Learning Algorithm for Radial Basis Function Networks]", IEEE Transactions on Neural Networks, Vol 2, No 2 (Mar) 1991.
[[Category:Neural
[[Category:Computational statistics]]
[[Category:Classification algorithms]]
[[Category:
[[Category:Regression analysis]]
[[Category:1988 in artificial intelligence]]
| |||