Applied Soft Computing 35 (2015) 52–65
Contents lists available at ScienceDirect
Applied Soft Computing
j ourna l ho me page: www.elsevier.com/locate /asoc
A novel hybrid learning algorithm for full Bayesian approach of
artificial neural networks
Ozan Kocada˘ glı
*
Department of Statistics, Mimar Sinan Fine Arts University, Silahsör Cad., No: 71, Bomonti/Sisli, Istanbul, Turkey
a r t i c l e i n f o
Article history:
Received 16 November 2013
Received in revised form 24 April 2015
Accepted 2 June 2015
Available online 25 June 2015
Keywords:
Bayesian neural networks
Bayesian learning
Hierarchical Bayesian models
Genetic algorithms
Markov chain Monte Carlo
Hybrid Monte Carlo
a b s t r a c t
The Bayesian neural networks are useful tools to estimate the functional structure in the nonlinear sys-
tems. However, they suffer from some complicated problems such as controlling the model complexity,
the training time, the efficient parameter estimation, the random walk, and the stuck in the local optima
in the high-dimensional parameter cases. In this paper, to alleviate these mentioned problems, a novel
hybrid Bayesian learning procedure is proposed. This approach is based on the full Bayesian learning,
and integrates Markov chain Monte Carlo procedures with genetic algorithms and the fuzzy member-
ship functions. In the application sections, to examine the performance of proposed approach, nonlinear
time series and regression analysis are handled separately, and it is compared with the traditional training
techniques in terms of their estimation and prediction abilities.
© 2015 Elsevier B.V. All rights reserved.
1. Introduction
Within Bayesian perspective, it is assumed that uncertainty of
any quantity of interest can be expressed and measured by prob-
abilistic distributions. This framework provides a natural way to
estimate the functional structure in the nonlinear systems. For this
reason, the Bayesian learning is mostly treated in the regression, the
time series, the classification and the density estimation applica-
tions of the artificial neural networks (ANNs). Bayesian learning in
the ANNs are typically based on Gaussian approximation, ensem-
ble learning and Markov chain Monte Carlo (MCMC) simulations
known as full Bayesian approach. For ANNs, Gaussian approxima-
tion was introduced by Buntine and Weigend [1] and MacKay [2]
in which one well-established procedure for approximating the
integrals over parameter space, known as Laplace’s method. This
approach is to model the posterior distribution by a Gaussian dis-
tribution, centered locally at a mode of posterior distribution of
parameters [2]. The ensemble learning was introduced by Hinton
and Camp [3] in which the approximating distribution is fitted
globally by minimizing a Kullback–Leibler divergence rather than
locally. In the context of Full Bayes approach, Neal [4] introduced
advanced Bayesian simulation methods in which MCMC simula-
tions are used to generate samples from the posterior distribution.
*
Tel.: +90 212 246 0011x5511; fax: +90 2122611121.
E-mail address: ozankocadagli@gmail.com
However, MCMC techniques can be computationally expensive,
and also suffer from assessing the convergence. For this reason, Neal
[4] integrated Bayesian learning with Hybrid Monte Carlo (HMC)
method introduced by Duane et al. [5] to overcome the mentioned
shortcomings. Afterwards, the Bayesian applications to ANNs was
reviewed thoroughly in [6–8].
In the literature, there are the remarkable studies that deal
with the specific problems related to ANNs from Bayesian per-
spective. For instances, Insua and Müller [9], Marrs [10], Holmes
and Mallick [11] worked on the issue of selecting the number
of hidden neurons with the growing and the pruning algorithms
for the dimensionality problem in the ANNs. In these studies,
they applied the reversible jump MCMC algorithm introduced by
Green [12], Richardson and Green [13]. Freitas [14] incorporated
the particle filters and the sequential Monte Carlo (MC) methods
in the BNNs. Liang and Wong [15] proposed to the evolutionary
MC algorithm which samples the parameters in the ANNs from
the Boltzmann distribution using the mutation, the crossover and
the exchange operations defined in the Genetic Algorithms (GAs).
Chua and Goh [16] used the Gaussian approach for the multivariate
modeling in the ANNs. They used the stochastic gradient descent
algorithm integrated with the evolutionary operators to produce
the parameters in the ANNs. Liang [17] and Xie et al. [18] used
the truncated Poison priors to determine the neuron numbers in
the hidden layers, and estimated the parameters in the ANNs via
the evolutionary MC algorithms proposed by Liang and Wong [15].
Lampinen and Vehtari [19], Vanhatalo and Vehtari [20] improved
http://dx.doi.org/10.1016/j.asoc.2015.06.003
1568-4946/© 2015 Elsevier B.V. All rights reserved.