These results may indicate that more competition is needed. Another solution for

segmentation is to look at the output history of the gate in Figure 4-11: the maximum

value is for the winner, but the waveform gives good segmentation information.










 0 500 1000 1500 2000 2500 3000 3500

Figure 4-11 History of the gate: full line is for expert 1, dotted line is for expert 2

 OLS clustering. Figure 4-12 shows the results of approach II for OLS clustering.

The prediction error is in the exact same range as with k-means clustering, but the

segmentation on that approach is performed far better: all mandatory breaths are detected

well. Actually, this model exhibits a behavior opposite that of the previous model: the

mandatory breaths are well detected, but the detection of return to regime 2 takes longer

(10 to 90 samples more): this is a first hint that the switching mechanism is not

symmetric, an issue we will further develop in chapter 5.

 Also, there is a fundamental difference between those two clustering algorithms:

the k-means algorithm tries to find similarities by assigning a data point to the closest

cluster, whereas the OLS algorithm tries to finds each added center with respect to how

much more than the others it explains, or dissimilar from the others it is.