The intelligence of environments 151
realm of established application-oriented technologies (SL and, quite recently and
also very successfully, MMMOG). Some will follow some paradigm of AI (to different degrees, cognitive robotics, social robotics, and AR), whereas others will
not. What emerges, I believe, is an instructive mosaic image of what informational
environments are, and what the interplay between informational environments and
cognitive artefacts is.
Evolutionary and cognitive robotics
The research field of evolutionary robotics addresses the question of how adaptive patterns may be established – or rather, may establish themselves – in the
course of some evolution-like development.
2
In contrast to behaviour-based AI,
with which it shares the notions of decentralised processing and of bottom-up
modelling of basic adaptive behaviours, there is no pre-defined task whose fulfilment is broken down into smaller sub-routines which, in self-organising fashion,
constitute the global behaviour. In evolutionary robotics, the self-organising characteristics are located at an even more basic level than that, namely in random
variation of an artificial “genome” that produces behavioural mechanisms, and
quasi-natural selection of the behavioural patterns, as the corresponding “phenotype”, in the robot’s environment. The robot will be moving within a simplified
artificial environment, and he will be provided with only a few optical sensors that
are coupled with his actuators by means of a rather modest neural network as his
control system. The robot is not equipped with inner models, maps or images of
his environment.
In a classic experiment in this field (Harvey et al. 1994), a population of robots
encounters rectangles and triangles in its environment, and is supposed to evolve
an ability to distinguish between them. The robots are neither provided with the
task of making that distinction, nor are they equipped with sensors that could
distinguish between rectangles and triangles per se. They can only distinguish
between straight vs. oblique orientation of edges. Instead, a fitness function is
defined for them that consists of rewarding the robots for turning away from the
rectangles and moving towards the triangles, and punishing them for moving
towards the rectangles. The algorithms responsible for the robots’ behaviours vary
in accordance with a random function from generation to generation. In repeated
rounds of the experiment, and hence over a number of “generations”, the variant
algorithms are selected on the grounds of how they link the robots’ sensory input
to their behavioural output. Only algorithms for behaviours that, in practice, map
onto the distinction between triangles and rectangles with some degree of reliability will be reproduced in the next generation. If a robot fails at tracking the right
object at one stage of the experiment, the algorithm will not be reproduced in the
next generation, where a different variant, producing a different set of behaviours,
will get its chance.
However, the activation state of the sensors typical for encounters with triangles will also obtain when approaching a rectangle from some specific angles.
Only some variants of behaviour, namely taking a partial turn and thus positioning
realm of established application-oriented technologies (SL and, quite recently and
also very successfully, MMMOG). Some will follow some paradigm of AI (to different degrees, cognitive robotics, social robotics, and AR), whereas others will
not. What emerges, I believe, is an instructive mosaic image of what informational
environments are, and what the interplay between informational environments and
cognitive artefacts is.
Evolutionary and cognitive robotics
The research field of evolutionary robotics addresses the question of how adaptive patterns may be established – or rather, may establish themselves – in the
course of some evolution-like development.
2
In contrast to behaviour-based AI,
with which it shares the notions of decentralised processing and of bottom-up
modelling of basic adaptive behaviours, there is no pre-defined task whose fulfilment is broken down into smaller sub-routines which, in self-organising fashion,
constitute the global behaviour. In evolutionary robotics, the self-organising characteristics are located at an even more basic level than that, namely in random
variation of an artificial “genome” that produces behavioural mechanisms, and
quasi-natural selection of the behavioural patterns, as the corresponding “phenotype”, in the robot’s environment. The robot will be moving within a simplified
artificial environment, and he will be provided with only a few optical sensors that
are coupled with his actuators by means of a rather modest neural network as his
control system. The robot is not equipped with inner models, maps or images of
his environment.
In a classic experiment in this field (Harvey et al. 1994), a population of robots
encounters rectangles and triangles in its environment, and is supposed to evolve
an ability to distinguish between them. The robots are neither provided with the
task of making that distinction, nor are they equipped with sensors that could
distinguish between rectangles and triangles per se. They can only distinguish
between straight vs. oblique orientation of edges. Instead, a fitness function is
defined for them that consists of rewarding the robots for turning away from the
rectangles and moving towards the triangles, and punishing them for moving
towards the rectangles. The algorithms responsible for the robots’ behaviours vary
in accordance with a random function from generation to generation. In repeated
rounds of the experiment, and hence over a number of “generations”, the variant
algorithms are selected on the grounds of how they link the robots’ sensory input
to their behavioural output. Only algorithms for behaviours that, in practice, map
onto the distinction between triangles and rectangles with some degree of reliability will be reproduced in the next generation. If a robot fails at tracking the right
object at one stage of the experiment, the algorithm will not be reproduced in the
next generation, where a different variant, producing a different set of behaviours,
will get its chance.
However, the activation state of the sensors typical for encounters with triangles will also obtain when approaching a rectangle from some specific angles.
Only some variants of behaviour, namely taking a partial turn and thus positioning