should be a winner does not compensate fast enough, the other expert stays a winner and

adapts to the new regime, so that there the system barely or never detects a switch (Figure

4-7). Therefore the determination of the memory depth is a compromise between

unwarranted switches and detection delays or misses.

 The gate. The gate is the part of the model that decides the assignment of a specific

regime to a particular expert, by ensuring that most successful experts receive the most

update. The gate can be input-based (i.e., an adaptable function of the input, that learns to

forecast which expert will perform best), or output-based (i.e., a calculated function of

the output of the experts, or of their performance). All gates studied here are output

based.

 In this framework, the competition mechanism and the gate can be coupled: the

history of the output of the gate does provide segmentation information, but since we are

attempting online segmentation concurrent with the training of the experts, the gate's

output is not necessarily binary, so for the sake of clarity the competition mechanism for

the segmentation output is defined separately.

 The winner is chosen with respect to its performance as shown in (2.9), p (n)

being the value of the performance criterion for the pth expert at sample n.

 winner(n) = argmin [p (n)] (2.9)

 In Figure 2-1 is depicted the common structure used along this thesis, with several

types of gates and competition, and several types of experts such as linear filters,

multilayer perception, or radial basis function network (another example of interesting

experts is PCA networks [14], where then the competition is based on how well the

experts can compress and decompress the data rather than predict it).