In my opinion, Chapter 5 in the most important chapter in this dissertation
to show what had the author done? 

That ends decomposition tree building phase. Then the models 1, 2, 3 and
4 are created and trained using the learning sub-databases that have arrived to their
location in the tree. That ends T-DTS learning phase.

=> how the 4 models are created?

5.1.2 Model identification - drilling rubber problem


This section uses data from real industry relating the proccess of drilling rubber.
Patterns to be input here have M-ARMAX shape. 
DU's in this example are constructed by Kohone SOM with 4x3 grid which
leads to 12 feature sub-spaces. Hence 12 PU's based on NN are created and trained. 

The results shown are 
(1) 4 out of 12 sub-databases are shown as examples


Kohonen's SOM is popular enough in our comunity but in mostcases
the reference is added

(Multi inputs ARMAX model). Figure 5.6
Learned process output identification (right)

The models then were built for each
sub-database resulting in good estimation properties of resulting system.

5.2.1
The data is composed
of two classes, where decision boundary is a spiral