atal
ion
ain
int
ect
In
| is
| dS
ion
CO-
are
the
ent
of
ter
me
ich
| is
ur-
is
ata
cal
he
re-
his
ic-
ng
ats
he
zw
tal
je-
jis
on
ht
vil
T3
d.
n-
ls
nd
nt
lal
nd
18
C-
ft
short — term future
P
——— ——— histor axpec-
y Ai
details of internal
representation
1
arse
ongterm
expectations
This type of data compression into valid
models is symbolized in figure 7 by the
formation of a reduced tail on the past time
axis (left). Quasi-static knowledge resulting
from this is used lateron for triggering
proper control activities depending on the
situation encountered. Standard perturba-
history of
tie objects
12
>
abstracted experience v =
————— treet) >, TO 7
e modal elements for the in- S, s
terpretation of sensor data
e goal functions
e control mode actuation
e instantiation control extended
long term memory presence
' here and now’
Fig.7: Representation density over time
Once this is represented, predictions of the state evolu-
tion over time may be obtained at relatively low cost. Since
usually neither the control nor the perturbation inputs of
the future are known, prediction usually stops at one cycle
(for the normal prediction-error-feedback state estimation
process) or after only a few cycles in order not to incur too
much uncertainty. For well known feedforward control
time history inputs in order to achieve some maneuver
element (for example lane change in road vehicle guidance
with a sine-like steering angle input over time using proper
parameters for period T and amplitude A) reliable predic-
tions over longer temporal ranges (seconds) are possible.
Taking standard perturbation statistics into account, even
longer ranges over entire maneuver sequences may be
meaningful (like prediction of the time needed to go from
point A to B). In the average, however, the number of
predicted events will vanish on the future time scale to the
right.
If good internal models are available for generating rich
actual internal representations from the actual data
measured, it will be impossible to store all these data as a
'personal history of adventures'; it is not necessary,
though. Since the time histories of the state variables may
be regenerated from stored initial conditions and control
as well as perturbation time history inputs once a proper
model for the dynamic behavior is available, only the latter
ones need be stored. For these again, instead of pointwise
storing each individual time history, parameterized
generic models would allow very efficient storage since a
dense data input vector may be replaced by a few parame-
ters needed to feed the proper function call. This shows
that proper temporal models may be very efficient in
reducing memory requirements if things are properly or-
ganized. Past process state time histories and events may
then be reconstructed actively from combining only a few
stored historical data with stored model knowledge. This
principle is the basic advantage of the 4D approach com-
bining space and time in an integrated manner.
ee
— tions are counter-acted by feedback control
laws which are implemented by a direct
loop from the sensory data to the corre-
sponding actuators (see center of fig.7) via
internal state variables of recognized ob-
jects; this allows stable behavior under per-
turbed conditions without the explicit
knowledge levels having to interact withthe
high frequency data stream. Only unforseen
situations and unpredicted new features dis-
covered lead to an activation of the more
knowledge based hypothesis generation
partcontrolling the active set of internal dynamical models
(lower left in fig.7).
Seen from this point of view, the entire mental’ internal
world of representations has as its purpose to provide the
system with capabilities of data interpretation well suited
for control outputs which enable the system to achieve its
goals; previous experience may be exploited for this pur-
pose contributing to the rating of a system as being intel-
ligent or not.
The 4D-approach
For areliable description of mechanical processes in our
everyday environment science has found the framework
of threedimensional space and time to be well suited.
Objects are defined in this environment as units having
special properties or functions. For simplicity, we confine
ourselves at present to rigid objects which may be moved
as units having constant shape over time (e.g. vehicles,
obstacles) or which are static parts of the environment
(roads,buildings,installations etc.). Each object has a spa-
tial shape, a position and an angular orientation in a
framework relative to the observer, all in 3-D. Objects are
classified according to their mobility: 1. Environmental
objects are fixed to an environment and determine its
visual appearance, like roads, road shoulders, trees and
buildings, walls; 2. static objects are presently at rest,
however, they may be moved or may even belong to the
last class; 3. objects able of autonomous locomotion. The
vehicle itself is an object of class 3, for which a model of
its locomotion capabilities and of some basic geometrical
properties are known. This includes the cause-and-effect
relationships with respectto activating the controls and the
state transition over time. In addition, the position and
orientation of the vision sensor relative to those parts of
the body interacting with the environment, i.e. the wheel
base, are assumed to be known.