For training, two humans interact with each other in a somewhat natural way while the system accumulates information about the actions and reactions. This is depicted in detail in Figure 8.1.
Here, the operation is straightforward. The two users are interacting
and the learning system is being fed
on one end and
on the other. Then, once many pairs of data are accumulated,
the system uses CEM to optimize a conditioned Gaussian mixture model
which represents
.
Once again, we note the role
of the integration symbol which indicates the pre-processing of the
past time-series via an attentional window over the past T samples
of measurements. This window can be represented compactly in an
eigenspace with
.
Typically the two users interact for a few minutes indicating to the
system some specific patterns of behaviour via a distribution of
pairs. Once the CEM converges to a maximum
conditional likelihood solution that approximates this distribution,
the interaction of the two users has been learned and can be used to
generate predictions.