Generated Bumper Learning State Machine
close
This
is a diagram of an actual Learned State Machine from one robot in one
corner learning experiment. (Note: these data are from a different
experiment series than the Distance and Entropy graphs shown before).
At the beginning of the experiment the machine was defined as the two
top states s6 and s7. As time goes on and experience collects the s7
state is split into copies of itself and the copies are 'allowed' to
learn on their own. What they learn is a different set of responses to
sequential bumper inputs. This can be seen in the sub-states
by following the colored paths. Instead of "blindly" using the
Back-up-and-turn-away response, the robot has learned that the sequence
of inputs that indicate a corner should initiate a strategy of
Back-up-and-turn-the-same-direction until it escapes.
Another
measure of structure is how many states a machine has. This diagram is
not entirely indicative because it hasn't been reduced to a minimal
set, but it does show that order and structure have increased.