Information Technology Reference
In-Depth Information
The SOM contains all possible states of the robot, distributed across the
map in a topologically ordered fashion and clustered according to similarity of
the states. Initially the SOM is trained on a set of known problematic as well
as stable states. This gives the map an informed starting point, from which
it can evolve and adapt over the lifetime of the robot. A major feature of a
SOM is the clustering effect which means that general robotic states can be
identified in the maps produced when trained in this way. An example of this is
the stable/homeostatic state; this state will be represented within the SOM by
a cluster of similar nodes in which most of the TLR responses are zero. This can
be seen in figure 3 in the top left corner of the map. In contrast the dark region
in the lower right quadrant of the map has clustered all the states in which two
motors are overheating and can be considered to be a stressed state of the robot,
and if the robot remains in this state for long periods then inflammation will
result and spread the activation throughout the map.
The input into the SOM is the TLR vector, which contains all TLR responses.
This vector is presented to the SOM and the algorithm finds the node within
the map which is closest to the input feature vector. In our case this is measured
using the Euclidean distance.
3.4
Neuro-endocrine Control
The system then passes on the responses, which correspond to the winning node
within the SOM, in order to influence the higher level control mechanism's be-
haviour. This response could be achieved in a number of ways, but perhaps a
good candidate would be using a neuro-endocrine control system [7,8] where the
artificial hormone is simply the inflammation level. These neuro-endocrine con-
trollers rely on standard multi-layer perceptron neural networks with the simple
addition of sensitivity to hormone concentrations built into their synapses. Thus
the neural networks in the control system could be selectively (selection being
performed by the SOM) suppressed by the application of the inflammation level
as an artificial hormone at their synapses in the (now standard) neuro-endocrine
way:
nx
u
=
w
i
·
x
i
·
inf
t
(3)
i
=0
where
n
is the number of synapses at the artificial neuron,
w
i
is the weight
associated with the
i
'th synapse,
x
i
is the input to that synapse and
inf
t
is the
inflammation level at the time
t
. This new activation level is then used with the
standard output function:
1
1+
e
−u
o
=
(4)
where
o
is the output from the neuron in question. This provides a simple but
effective way of affecting the higher level control systems of the robot.
