predictor to spike (assuming the predictor did not model the noise) and might create
segmentation problems.
3.4 Data Models
Dynamic modeling implies a two-step process: transforming the observed data into
a path in the state space (i.e., determine an embedding), and then from this path build the
predictive model [22]. Now that the embedding is chosen, we consider several design
options for the experts: the traditional linear approach and two nonlinear ones using
global dynamic models that are universal mappers, one with global basis (MLP) the other
with local basis (RBFN).
3.4.1 Degrees of Freedom
Whereas all structural parameters are now set for the linear model, there is still
another parameter to consider for the nonlinear models. As explained in the next sections,
both models have a hidden layer of processing elements (PE), and the number of PEs
determines the number of degrees of freedom of the experts. Minimizing the network size
is a important issue in signal modeling, because a network of smaller size has less risks of
learning the noise in the data (and thus generalizes better), and also because a smaller size
means fewer weights and a smaller computational load. The optimal network size can be
determined by using "network growing or pruning" methods, but it is also dependent on
the embedding dimension.
A second data set recorded from a patient under assisted ventilation is presented in
Figure 3-7. This data set shares the waveform shapes and other particularities with the set
used for the simulations, so the expert's structure should be similar for both sets, and it is
interesting to consider designing some of the expert's structure from this second set,