Information Technology Reference
In-Depth Information
Fig. 1.
(a-h) Dynamic of the training stage presenting the growing and the re-labeling
process. (i) Performance of the different networks.
2.4
Convergence Criterion
The learning procedure involves a constructive network, and it is a challenging
task to automatically decide when the network should stop growing. To perform
this task, an approach was developed based on two concepts. The first one is the
reference network
, that represents the network with the best performance so far.
The other concept is called
convergence window
and is related to the number
of networks that will have its performances compared to the
reference network
.
These concepts are illustrated in Fig. 2 and in Fig. 3. The size of the convergence
window defines the number of networks to be compared to the
reference network
,
and it is the unique user-defined parameter of the algorithm. The performance
evaluation, at this point, is achieved by using only the training dataset.
