Trade-Offs in Sensoric Information Acquisition and Processing

			    Daniel Polani
		   Adaptive Systems Research Group
		    Department of Computer Science
		     University of Hertfordshire

		  Preferred presentation mode: oral

			       Abstract


From the large amount of information in an agent's environment, only a
limited part is relevant to the agent's purposes. In settings where a
concept of utility can be defined, it is possible to give the notion
of 'relevant information' a precise meaning in the spirit of of
Shannon. The amount of this information can be extracted, and for
given systems it is also (at least in principle; present methods are
restricted to systems of limited size) possible to identify the
environmental structures where this information comes from.

In addition, it turns out that one can trade off utility for
'relevant' information to be processed. In certain circumstances this
means that by forgoing a relatively small amount of utility, one may
experience significant savings in relevant information. This means
that there is a potential for the sensoric apparatus to concentrate on
certain types of information and for the information processing
channels to remain small. Given that for real-world agents suboptimal
solutions are often sufficient, a judicious reduction of the
achievable utility can minimize the effort of information
processing. Coupled with a suitable morphological or sensorical
equipment, it offers a principled path towards the development of
minimal agents.

Under this view, an agent with several different goals and different
incompatible utility functions may distribute its available 'relevant'
information processing resources to balance the utilities belonging to
the different tasks. This provides a formal description of how a
limited amount of processing power can be used to solve different
problems. In addition, this may act as a formalism which allows to
tackle the intriguing question is how existing sensoric interfaces
which have evolved under certain conditions and requirements can be
reused in a new function for alternative tasks. 

In fact, in certain scenarios, the formalism itself predicts that even
minimal sensoric interfaces actually have to capture more information
from the environment than is actually relevant to solve a given
problem. We believe that in these cases this extra information then
offers an entry point for selection pressure towards the development
of new information processing capabilities that take into accout this
unexploited information; this, in turn could provide insight about the
forces that drive evolution to create increasingly complex
perception/action systems.