152 Environments of intelligence
the sensors in a different angle towards the object, will allow the robots to attain
the disjunct activation states for their sensors that are correlated with the difference between triangles and rectangles. The variant algorithms resulting in this
behaviour will be positively selected. In a number of runs (or “generations”) of
this setting, the robot’s control system will have learned to practically distinguish
triangles from rectangles with a high degree of reliability, thus establishing an
adaptive fit between his behavioural mechanisms and the features of his environment that serve as his fitness function. A very basic functional analogy to the
development of adaptive behaviours in natural organisms is thus established.
Still, despite the capability of practically distinguishing between triangles and
rectangles, even the most successful robots will not acquire even the most basic
of concepts of triangles or rectangles. Instead, the robots’ behaviour is directly
guided by the natural information pertaining to the shapes of triangles and rectangles, to the effect that the difference in shape between them constitutes the only
relevant variable in their informational environment. The invariant tracked by the
robot is the angular orientation of edges that covaries with the presence of triangles and rectangles, for which there exists one ambiguous state that has to be, and
can be, disambiguated by moving in relation to those edges, so that transformations of the robot’s relative position match transformations in the detected angular
orientation of the edges.
If there are affordances, in Gibsonian terms, that would be provided by that one
trackable invariant, these could be circumscribed as a drive-to-ability or favourable proximity to triangles and avoidance of rectangles respectively, despite the
absence of concepts of triangles or rectangles that would be available to the robot.
To the extent that seeking the proximity of triangles and seeking distance from
rectangles is the fitness function for the robots, activities of information uptake
that match this function will be activities of furnishing the respective affordances
on the most basic level.
The robot’s informational environment is composed of the one relevant variable that he is capable of tracking, namely the difference between triangles and
rectangles. Given the experimental setting described earlier, he is bound to the
domain in which the difference between angular orientation of edges, given a certain relative position of his sensors towards them, reliably and regularly covaries
with the difference between triangles and rectangles. That difference being the
one and only ecologically and selectively relevant condition to him, his ecological
and selective environment is composed of that spatio-temporal region in which
the correlation between angular orientation and the objects in question holds,
which happens to coincide with the room and the time in which the experiment
is conducted – unless tampering with the fitness function becomes part of the
experiment.
However, the experiments in evolutionary robotics described here are not supposed to provide models of ecological perception in the first place, but to provide,
in elementary embodied fashion, a model of processes of biological evolution,
with the elements of variation, selection and adaptation. These processes are that
model’s primary target system. Notably, neither the specific variations produced
the sensors in a different angle towards the object, will allow the robots to attain
the disjunct activation states for their sensors that are correlated with the difference between triangles and rectangles. The variant algorithms resulting in this
behaviour will be positively selected. In a number of runs (or “generations”) of
this setting, the robot’s control system will have learned to practically distinguish
triangles from rectangles with a high degree of reliability, thus establishing an
adaptive fit between his behavioural mechanisms and the features of his environment that serve as his fitness function. A very basic functional analogy to the
development of adaptive behaviours in natural organisms is thus established.
Still, despite the capability of practically distinguishing between triangles and
rectangles, even the most successful robots will not acquire even the most basic
of concepts of triangles or rectangles. Instead, the robots’ behaviour is directly
guided by the natural information pertaining to the shapes of triangles and rectangles, to the effect that the difference in shape between them constitutes the only
relevant variable in their informational environment. The invariant tracked by the
robot is the angular orientation of edges that covaries with the presence of triangles and rectangles, for which there exists one ambiguous state that has to be, and
can be, disambiguated by moving in relation to those edges, so that transformations of the robot’s relative position match transformations in the detected angular
orientation of the edges.
If there are affordances, in Gibsonian terms, that would be provided by that one
trackable invariant, these could be circumscribed as a drive-to-ability or favourable proximity to triangles and avoidance of rectangles respectively, despite the
absence of concepts of triangles or rectangles that would be available to the robot.
To the extent that seeking the proximity of triangles and seeking distance from
rectangles is the fitness function for the robots, activities of information uptake
that match this function will be activities of furnishing the respective affordances
on the most basic level.
The robot’s informational environment is composed of the one relevant variable that he is capable of tracking, namely the difference between triangles and
rectangles. Given the experimental setting described earlier, he is bound to the
domain in which the difference between angular orientation of edges, given a certain relative position of his sensors towards them, reliably and regularly covaries
with the difference between triangles and rectangles. That difference being the
one and only ecologically and selectively relevant condition to him, his ecological
and selective environment is composed of that spatio-temporal region in which
the correlation between angular orientation and the objects in question holds,
which happens to coincide with the room and the time in which the experiment
is conducted – unless tampering with the fitness function becomes part of the
experiment.
However, the experiments in evolutionary robotics described here are not supposed to provide models of ecological perception in the first place, but to provide,
in elementary embodied fashion, a model of processes of biological evolution,
with the elements of variation, selection and adaptation. These processes are that
model’s primary target system. Notably, neither the specific variations produced