completely unsupervised, we are tackling a particular problem with its own specificities,

and they should be included in the design of the system.

 One particularity is our goal to achieve online segmentation, which means only one

epoch through the data and training at the same time as segmentation. Choosing the MSE

estimate helps preventing spurious switching during segmentation by adding memory to

the performance criterion, its memory depth has to be set for each expert. Since we have

no reason for a priori distinguishing one expert or from another in our case, and

considering the expert's return map have close trajectories as we show in the next

chapter, Ye, is chosen to be the same for each expert, and since the real data is non-

stationary, it will not be annealed. Therefore, after testing several values on the test set

defined in the next chapter, yjs =0.01 is chosen to define a memory depth of one hundred

sample points, which seems logical because it corresponds roughly the length of a breath

cycle. But such a long memory depth also accounts for the "error buildup" issue evoked

in approach III. Also, the more memory about the experts' performance is available, the

more information the gate has about the experts' performance, but the less accurate the

segmentation becomes, because the relative importance of a change point in the criterion

decreases, so the switch might come a few samples late.

 Lastly, concerning approach IV, the same competition parameter is set for all

models, for a really soft competition at M=2. This allows the non-winning expert to still

adapt a little to the data without loosing its specialization (the learning rates are chosen so

that the experts do not adapt too much).