Figure 4-3 shows results for approach IV, where both experts are adapted with a

learning rate dependent on the output of gate at each time step.

 In both cases, only expert 1 wins and keeps winning, even with approach IV when

the two networks are adapted: this proves further that a linear model is not a good expert

in this case: there is a conflict between the high dimension of the reconstruction space

needed to be able to reproduce the data accurately in order to segment on shape

discrimination, and the necessity for a short input size to achieve sharp detection of the

changes.


Figure 4-3 Results: (a) Flow: desired in blue, predicted in red; (b) MSE criterion, full line
 is for expert 1, dotted line is for expert 2; (c) Winner; (d) Pressure.


1
0

 500 1000 1500 2000 2500 3000

02


0 1
 500 1000 1600 2000 2600 3000
 2-


 1L
 500 1000 1600 2000 2600 3000


50

 500 1000 1600 2000 2600 3000