Measuring Adaptation
Distance Traveled |
Controller Entropy |
|
|
direct measure of behavior
|
measure of structure
|
from unpublished
work with Dr. J. P. Crutchfield at the Santa Fe Institute
|
circle experiment
video
corner experiment
video
next
The
graphs show the combined results of the three experiment series in two
different ways.
Distance Traveled
On the left is the total distance traveled over time for each series.
This is a pretty direct measurement of the the instincts given to the robots as
evaluation criteria. As can be seen, the Designed machine travels the
furthest from the start and the Random machine does the worst. The
slope of the lines is the average speed of the robot, so we also see
that the Designed machine goes fastest and the Random one slowest.
Looking at the Learned machine's results one can see that it starts out
going about the same speed as the Random one but then inflects to the
speed of the Designed machine. At the inflection point is where the
learned evaluations take over from the random explorations. After that
point the robot's speed is nearly the same as the Designed machine so
we can infer that the learning algorithm is working fairly well. In the
individual experiments the inflection point occurs at various times,
the data displayed is an average over all the learning experiment runs.
Controller Entropy
On the right is a plot of the Entropy
of the controlling state machine over
time for each experiment series. In this context Entropy is a big word
for how
random something is, where lower values are less random. One can see
that after some time, the randomness of each controller parallels the
measurements made using the Total Distance. Where the Random machine
was slow, it's entropy is high and it is disordered; and, where the
Designed machine was fast, its entropy is low and it is very ordered.
The Learned machine shows a transition from high to low entropy, from
disorder to order. Interestingly, this measure has nothing to do
with the learning evaluation criteria and is thus more general as it
indicates the amount structure in the controller itself.
(Way Short) Discourse on Information Theory
Information is an
unfortunate name for this because it gets confused with Meaning. In
this context Information can be thought of as Surprise,
i.e., how Surprised are you to get a new bit of information. If you had
no way to know what the new event would be you are very surprised, and
vice versa. In this sense then, random sequences are more surprising
and thus contain more Information than predictable sequences. The
measure of Information is Entropy, so a high Entropy means a lot of
surprising Information.
The word Entropy is also a
confusing term because it is used in the context of thermodynamics to
mean energy loss. However the mathematics for calculating each is
identical in form, and Johnny von Neumann (again...) suggested that
Claude Shannon use entropy to describe his Information Theory. It turns
out that one can be stated in terms of the other in many cases, so it
wasn't such a bad idea.