Applied Soft Computing 11 (2011) 2326–2333
Contents lists available at ScienceDirect
Applied Soft Computing
journal homepage: www.elsevier.com/locate/asoc
MIMO CMAC neural network classifier for solving classification problems
Jui-Yu Wu
∗
Department of Business Administration, Lunghwa University of Science and Technology, No. 300, Sec. 1, Wanshou Rd., Guishan, Taoyuan County 33306, Taiwan
article info
Article history:
Received 16 January 2010
Accepted 2 August 2010
Available online 12 August 2010
Keywords:
Neural network classifier
Cerebellar model articulation controller
Classification
Machine learning
abstract
Developing an efficient classification method is a challenge task in many research domains, such as neural
network (NN) classifiers, statistical classifiers and machine learning. This study focuses on NN classifiers,
which aredata-driven analytical techniques. This study presents a cerebellar model articulation controller
NN (CMAC NN) classifier, which has the advantages of very fast learning, reasonable generalization ability
and robust noise resistance. To increase the accuracies of training and generalization, the CMAC NN clas-
sifier is designed with multiple-input and multiple-output (MIMO) network topology. The performance
of the proposed MIMO CMAC NN classifier is evaluated using PROBEN1 benchmark datasets (such as for
diabetes, cancer and glass) taken from the UCI Machine Learning Repository. Numerical results indicate
that the proposed CMAC NN classifier is efficient for tested datasets. Moreover, this study compares the
experimental results of the CMAC NN classifier with those in the published literature, indicating that the
CMAC NN classifier is superior to some published classifiers. Therefore, the CMAC NN classifier can be
considered as an analytical tool for solving classification tasks, such as medical decision making.
© 2010 Elsevier B.V. All rights reserved.
1. Introduction
Various problems in science, business, industry and medicine,
including the prediction of bankruptcy, medical diagnosis, hand-
written character recognition and speech recognition, can be
considered as classification problems [26]. In classification prob-
lems, an instance is assigned to a predefined class based on various
features. Many classification methods have been presented. They
can be classified into the following categories: decision tree clas-
sifiers (such as ID3, C4.5, CART), Bayesian classifiers based on
statistics, instance-based learners (including case-based reasoning
and minimum distance classifier), support vector machines, fuzzy
decision tree classifiers and neural network (NN) classifiers [18,19].
In the StatLog project, NN classifiers, statistical classifiers and
machine learning approaches were applied to more than 20 real
datasets. The analytical results of StatLog demonstrated that no
unique classifier was likely to perform best on all datasets [16].
The conclusion of StatLog corresponds to the No Free Lunch
Theorem [25], which states that if algorithm A outperforms algo-
rithm B on average for one class of problems, then it must be
worse than B on average over the remaining problems. There-
fore, this study focuses on NN classifiers, since NNs can learn
highly nonlinear patterns using learning algorithms. Several NN-
based classifiers have been successfully used to many domains. For
instance, Mazurowski et al. [15] developed NN classifiers based
∗
Tel.: +886 2 8209 3211x6509; fax: +886 2 8209 3211x6510.
E-mail address: jywu@mail.lhu.edu.tw.
on the classical back-propagation (BP) learning algorithm and
particle swarm optimization (PSO), and applied them to medical
decision making, Lisboa and Taktak [13] surveyed NN classifiers
used in decision support in the clinical domain, and Misra et al.
[17] designed an improved polynomial NN classifier based on two
learning algorithms (BP algorithm and PSO) for many benchmark
classification problems. Although hybrid systems may improve
the accuracy of classification, they are more complex. Moreover,
some back-propagation network-based classifiers have some lim-
itations, including a slow training time, difficulty of interpretation
and difficulty of implementation in terms of the optimal number
of neurons [18]. Fortunately, the cerebellar model articulation con-
troller (CMAC) NN has the benefits of very fast learning, reasonable
generalization ability and robust noise resistance. Albus [1,2] first
introduced the CMAC NN based on the functions of the human
cerebellum, which is responsible for muscle control and motor
coordination. A cerebellum works as follows. An input signal to the
cerebellum activates numerous mossy fibers, each of which touches
a granule cell. The output of the cerebellum is the sum of the out-
put of the activated granule cells. A CMAC NN performs cerebellum
functions via a series of mappings, and acts as a clever look-up table.
The CMAC NN has been successfully applied in several applications,
such as control [11,12,14,20] and fault detection [7,23]. In classi-
fication tasks, Wen et al. [24] presented a self-organizing CMAC
NN classifier for electrocardiogram classification, and Lin et al. [10]
developed a parametric fuzzy CMAC NN with a hybrid parameter
learning algorithm that consists of a self-clustering genetic algo-
rithm (GA) and a modified GA, for face detection and breast cancer
diagnosis.
1568-4946/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2010.08.013