|dc.description.abstract||Based on the view of an agent as an information processing system, and the
premise that for such a system it is evolutionary advantageous to be parsimonious
with respect to informational burden, an information-theoretical
framework is set up to study behaviour under information minimisation pressures.
This framework is based on the existing method of relevant information,
which is adopted and adapted to the study of a range of cognitive aspects.
Firstly, the model of a simple reactive actor is extended to include layered decision
making and a minimal memory, in which it is shown that these aspects
can decrease some form of bandwidth requirements in an agent, but at the
cost of an increase at a different stage or moment in time, or for the system as
a whole. However, when combined, they do make it possible to operate with
smaller bandwidths at each part of the cognitive system, without increasing
the bandwidth of the whole or lowering performance.
These results motivate the development of the concept of look-ahead information,
which extends the relevant information method to include time, and
future informational effects of immediate actions in a more principled way. It
is shown that this concept can give rise to intrinsic drives to avoid uncertainty,
simplify the environment, and develop a predictive memory.
Next, the framework is extended to incorporate a set of goals, rather than deal
with just a single task. This introduces the task description as a new source
of relevant information, and with that the concept of relevant goal information.
Studying this quantity results in several observations: minimising goal
information bandwidth results in ritualised behaviour; relevant goal and state
information may to some point be exchanged for one another without affecting
the agent’s performance; the dynamics of goal information give rise to a
natural notion of sub-goals; bottlenecks on goal memory, and a measure of
efficiency on the use of these bottlenecks, provide natural abstractions of the
environment, and a global reference frame that supersedes local features of
Finally, it is shown how an agent or species could actually arrive at having
a large repertoire of goals and accompanying optimal sensors and behaviour,
while under a strong information-minimisation pressure. This is done by introducing
an informational model of sensory evolution, which indicates that
a fundamental information-theoretical law may underpin an important evolutionary
catalyst; namely, even a fully minimal sensor can carry additional
information, dubbed here concomitant information, that is required to unlock
the actual relevant information, which enables a minimal agent to still explore,
enter and acquire different niches, accelerating a possible evolution to
higher acuity and behavioural abilities.||en_US