Representation is critical in the ARL system since it must be
carefully selected to achieve learning. Invariance in the system must
be introduced a priori into the representation. For example, if a
feature is spurious and contributes no information about the
interaction, it will waste resources. The system will wastefully
attempt to model and predict it inside the
vector. In
addition, the representations must be smooth and must not have
ambiguities. For example, during initial phases of development, the
ARL system employed a different representation of the head and hand
blobs. The head and hands were described by their mean, major axis,
minor axes and rotation (in radians). Unlike the square-root
covariance shape descriptor (our current representation), the rotation
value had some unusual singularities. The 0 and
values are
identical in radian notation. Thus, the system would generate
non-linear steps from 0 to
as the blobs would rotate and these
transitions were difficult to span using the eigenspace temporal
processing techniques. Thus, it is critical to pick a representation
which contains the desired invariants, is well behaved, is smooth and
has no spurious components.