Page 32 - Dynamic Vision for Perception and Control of Motion

16 1 Introduction

do not have just one single task to perform. Depending on their circumstances,
quite different tasks may predominate.
Systems will be labeled intelligent if they are able to:
x recognize situations readily that require certain behavioral capabilities and
x trigger this behavior early and correctly, so that the overall effort to deal with
the situation is lower than for direct reaction to some combination of values
measured but occurring later (tactical – strategic differentiation).
This “insight” into processes in the real world is indicative of an internal temporal
model for this process in the interpretation system. It is interesting to note that the
word “intelligent” is derived from the Latin stem “inter-legere”: To read in be-
tween the lines. This means to understand what is not explicitly written down but
what can be inferred from the text, given sufficient background knowledge and the
capability of associative thinking. Therefore, intelligence understood in this sense
requires background knowledge about the processes to be perceived and the capa-
bility to recognize similar or slightly different situations in order to be able to ex-
tend the knowledge base for correct use.
Since the same intelligent system will have to deal with many different situa-
tions, those individuals will be superior which can extract information from actual
experience not just for the case at hand but also for proper use in other situations.
This type of “knowledge transfer” is characteristic of truly intelligent systems.
From this point of view, intelligence is not the capability of handling some abstract
symbols in isolation but to have symbolic representations available that allow ade-
quate or favorable decisions for action in different situations which have to be rec-
ognized early and reliably.
These actions may be feedback control laws with very fast implementations
gearing control output directly to measured quantities (reflex-like behavior), or
stereotypical feed-forward control time histories invoked after some event, known
to achieve the result desired (rule-based instantiation). To deal robustly with per-
turbations common in the real world, expectations of state variable time histories
corresponding to some feed-forward control output may be determined. Differ-
ences between expected and observed states are used in a superimposed feedback
loop to modify the total control output so that the expected states are achieved at
least approximately despite unpredictable disturbances.
Monitoring these control components and the resulting state variable time histo-
ries, the triggering “knowledge-level” does have all the information available for
checking the internal models on which it based its predictions and its decisions. In
a distributed processing system, this knowledge level need not be involved in any
of the fast control implementation and state estimation loops. If there are system-
atic prediction errors, these may be used to modify the models. Therefore, predic-
tion error minimization may be used not just for state estimation according to some
model but also for adapting the model itself, thereby learning to better understand
behavioral characteristics of a body or the perturbation environment in the actual
situation. Both of these may be used in the future to advantage. The knowledge
thus stored is condensed information about the (material) world including the body
of the vehicle carrying the sensors and data processing equipment (its “own” body,