Re: Curtnetrons Don't Do Parity

Michael Olea <oleaj@xxxxxxxxxxxxx> wrote:
Curt Welch wrote:

And I really don't grasp the connections between thermodynamic entropy
and information entropy. But I'll figure that out as some point
here... :)

Think about the game of clue. At the start of the game the possibilities,
who, where, and with what, are least constrained. When you ask a question
and get an answer you gain information to the extent that it narrows the
possibilities. You have complete information, all the information there
is to be had, when the possibilities have been narrowed to one: Colonel
Mustard, in the library, with the lead pipe. If you ask a question and
get an answer that does not narrow the possibilities then it was useless
- you wasted your turn.

Ok, well, that's a problem of transferring information. And I did see some
comments about thermodynamic entropy being looked at as transferring
information about internal structure. So I kinda get how there could be a
link there.

And in relation to reinforcement learning, I could see how the problem
might be looked at in terms of the learning system trying to get
information about rewards from the environment. And how some ability to
quantify the transfer of information about rewards into the system would be
a way to quantify it's learning ability. But I'm having a harder time
seeing how it would be useful to look at the information flow from sensory
inputs to behavior output in terms of entropy. This is because the goal of
a reinforcement learning machine is not to perform some type of data
compression or optimization on the sensory information. Instead, it goal
is to calculate the value of the data so that it can throw out what is not
important and make use of what is important in the generation of behavior.
I see as a programmable information filter. The entropy of the input can
be measured, but that's just a measure of the bandwidth of the sensory
data. It seems to just be a way to quantify the rate of data flowing into
the filter. You can likewise quantify the flow of the data out of the
filter, but you can't determine anything about how close to optimal the
system is performing by only looking at those numbers. This is because
"good" is not defined by information flow. It's defined by the rewards
generated by the current configuration of the "filter". And the filter
configuration which is mostly heavily rewarded might have a very low
entropy, or it might have a very high entropy, or something in between. So
you can't tell anything about how "well" the filter is functioning by only
looking at its entropy output. But there does seem to be value in using
these concepts to understand what the system is doing. So I've got lots to
work on. :)

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Relevant Pages

  • Re: Energy
    ... I thought about using entropy as a "digital world" analog ... the physical world reduces energy or power. ... Filter this sequence, and entropy reduces. ...
    (comp.dsp)
  • Re: Energy
    ... that is not possible with analog filters. ... the physical world reduces energy or power. ... This has maximum entropy. ... Filter this sequence, and entropy reduces. ...
    (comp.dsp)
  • Re: Where does information come from?
    ... In general, then, we compute the information inherent in any given arrangement of matter from the number of choices we must make to arrive at that particular arrangement among all equally possible ones. ... "Information thus becomes a concept equivalent to entropy, and any system can be described in terms of one or the other. ... that information in biology doesn't have an objective definition ... The problem is not that you can't objectively quantify order - it's that you can quantify *any amount* of order you like depending on the description. ...
    (talk.origins)
  • Re: Lizard engines and rat engines
    ... I need to do a lot more reading and thinking about entropy. ... >> about entropy whereby something has to be put into order by an intelligent ... > While there is some pretty fuzzy thinking about entropy in this group, ... The fourth one is a bit fuzzier - there is no way to quantify ...
    (sci.bio.evolution)