International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 6. Bandung-lndonesia 1999 
116 
THE COMMUNICATION CONTROL IN MUTUAL CONNECTED NETWORK BY RC-RBFN 
Koji Okuhara*, Haruhiro Fujita** and Toshijiro Tanaka* 
* Hirhoshima Prefectural University 
567 Nanatuka, Shyobara, Hiroshima, Japan 727-0023 
** Shikoku National Agricultural Experiment Station 
Ministry of Agriculture, Forestry and Fisheries 
1-3-1 Senyu, Zentsuji, Kagawa, Japan 765-8508 
e-mail: okuhara@bus.hiroshima-pu.ac.jp 
ABSTRACT 
The mutual connected network can be controlled by the decentralized method. The network, generally speaking, has 
some restriction of the resource. We must control network parameters relating to such resources. In this paper we 
consider that the communication control in mutual connected network by reproductive and competitive radial basis 
function network. 
We first propose a competitive radial basis function network as learning algorithm. To solve the optimal problem of 
network system, we further extend the learning algorithm to reproductive criteria of network parameter. Compared with 
the usual radial basis function network it is shown that the reproductive competitive radial basis function network can 
faster find optimal value of parameter. 
I. INTRODUCTION 
Networks consist of nodes as elements, line connecting 
these nodes and transition. Many kinds of neural networks 
have been proposed by differences in the transition. 
Information processing which deeply concerns 
communications is executed in the nodes and lines. The 
problem of the network is the restriction of the resources 
at there. We propose the neural network algorithms to 
solve such optimization problem. 
Compared with multi-layer neural networks, by the way, a 
radial basis function network (RBFN) has excellent 
properties such as executing the local learning about each 
neuron. Thus, it is applied to the function approximate 
problem, the optimization problem and so on. In the 
initial state, however, the RBFN require the redundant 
neurons in order to approximate an uncertain nonlinear 
function because no one can not know the number of 
required neurons in advance. It is generally known that 
increasing the number of neurons implies the delay of 
learning and the over learning. To avoid this problem 
many methods have been proposed. But on decreasing 
dimensions of the RBFN, the criteria determining how 
many neurons should be left has not been established. We 
focus on this point and try to clarify the criteria. It is 
desirable that the criteria is naturally derived. For this 
purpose, we propose the reproductive and competitive 
radial basis function network (RC-RBFN). 
This paper is organized as follows. The outline of 
CRBFN described in the next section. In section 3 we 
propose the RC-RBFN determining the network 
parameters by learning algorithm. The experimental 
results appear in section4, followed by concluding 
remarks. 
2. THE OUTLINE OF CRBFN 
The RBFN is one of the neural networks approximating 
to a nonlinear function by adding up the radial basis 
function. One can consider the normalized Gaussian 
active function as the radial basis function. The structure 
of the RBFN with M input neurons and 1 output neuron is 
drawn by Figure 1.