Figure 4-9 shows the results for approach III, and the effect of the spike detection

of the winner is important: all mandatory breaths are now detected accurately, but there

are still two unwarranted switches, due to the fact that expert 2 seems to attain better

prediction of data (the amplitude of the criterion for expert 2 decreases slightly).

 Figure 4-10 presents approach IV, which does not work well for that model: the

two experts are adapted proportionally to the gate output and the winner gets most of the

update, but expert 1 is still updated for several consecutive spontaneous breaths, which

could explain why the criterion does not decrease as fast as expected when a switch

occurs (the first one as previously explained is detected only because it is at the very

beginning of the series). Also, expert 2 is updated during that first mandatory breath, so it

starts to explain regime 1 (we can notice that the spikes in criterion for expert 2 are not as

high for a switch to regime 1 as for the other approaches).


Figure 4-10 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 is for.


 500 1000 1500 2000 2500 3000

 0 02

 500 1000 1500 2000 2500 3000

41.5 -
 1 7 11