THE AUTONOMOUS SYSTEMS LABORATORY

In order to have the plant in a certain state at a defined time, we need other, more
powerful approaches that can anticipate error and prevent it. Due to the inherent
dynamics of the plant, the only possibility of acting to make it reach a final state sf
at tf from an intial state si at ti is to act at ta before tf .

This kind of control is anticipatory in this strict sense of (t_a < tf)⁷. The de-
termination of the action cannot come from the final state (as with classical error
feedback) because of anticipation and we need an estimate of this state Sf at time
ta.

These two alternative approaches were described by Conant Conant (1969) as
error-controlled regulation and cause-controlled regulation. The advange of this second
approach is that in certain conditions, it is often possible for the regulation to be
completely succesful at maintaining the proper outcome. Needless to say is that
due to the non-identity between model and reality, this last one may depart from
what the model says. In these conditions only error-driven control will be able to
eliminate the error. This is the reason why, in real industrial practice, model-based
controllers are implemented as mixed model-driven and error-driven controllers.

The previous analysis take us into the formulation of another principle:

Principle 3: Anticipatory behavior — Except in degenerate cases, maximal
timely performance can only be achieved using predictive models.

These predictive models can be explicit or implicit in the proper machinery of
the action generation mechanism Camacho and Bordons (2007). Obviously the de-
gree to which a particular part of reality can be included in a model will depend
on the possibility of establishing the adequate mappings from/to reality to/from
model and the isomorphims between entailments at the model level and at the
reality level (according to a particular model exploitation policy). The problems as-
socited to inferred model quality have been widely studied in relation with proper-
ties of statistical modelling, where we seek a good model to approximate the effects
or factors supported by the empirical data in the recognition that the model can-
not fully capture reality Burnham and Anderson (2004). This is also the world of
systems identification but in this case, the target model typically belongs to a very
reduced and precise class of models Ljung (1998); Nelles (2000).

4 Integrated Cognitive Control

Reactive and anticipatory control are the core building blocks of complex con-
trollers. Reactive controllers are simpler and more easily tuneable. These are the

⁷This could be seen as acausal because the cause of the action —final cause in aristotelian sense—
is the final state sf , that is a future state.

ASLab.org / Principles for Consciousness / A-2007-011 v 1.0 Final

<<   5   6   7   8   9   10   11   >>

More intriguing information