118 
3. SIMULATION AND CONSIDERATION 
2(a) shows both the teaching signal and the radial basis 
function prepared at the initial state. Figure 2(b) shows 
We show the ability of proposed method. For this the squared error function when the radial basis function 
purpose, the function is generating the teaching signal. It is removed, 
is supposed that one radial basis function is prepared at 
the initial state, the synaptic weight is given by 1. Figure 
Figure 2: The distribution of the teaching signals and the initial radial basis function. 
To paying attention to the efficiency of the reproduction, the 
updates for the synaptic weight and the parameter are omitted 
at Step 1. That is, added radial basis function with the same 
synaptic weight and the parameter is set to the parameter 
when the new stable convergence is detected. This means that 
such model is equivalent to RCRBFN whose parameter can 
be adjusted but the synaptic weight and the parameter can not 
Figure 3 shows that the change of the stable convergence of 
the parameter during learning. The horizontal axis denotes 
the parameter and the vertical axis the free energy, where the 
constant is a coefficient for easily viewing the figure. The 
stable convergence is shown by the black filled circle. F'igure 
4 shows the change of the squared error function during 
learning. 
Figure 3: The stable fixed point of parameter m during learning.