Re: Innoivationm and the Curse of Knowledge, etc

casey <jgkjcasey@xxxxxxxxxxxx> wrote:

The problem with building a robot that can
follow you instead of a simple light source
is recognizing you as a stimulus. A simple
photocell will respond to a variable light
source but will not respond to something as
complex as you.

Instead you need an array of photocells the
patterns of which have to be processed to
produce an invariant output called "Wolf".
The interesting question is how a system
might learn to recognize "Wolf" as a stimulus.

You might not need the array at all.

Even in a system with only two photocells, there's a very rich source of
information hidden in how the light levels change over time that a good
temporal pattern recognizing machine should be able to take advantage of.

For example, if you have two people with flashlights, the light levels will
change based on how you move, and point the flashlight, and on the type of
flashlight - or how it might be focused. A flashlight with a narrow and
bright beam will create very different temporal light level changes as you
wave it at the robot from a flashlight with a very wide but dimmer beam.

One person might wave the light in a way to beckon the robot to follow,
where as another person might point it at the ground to try and indicate
where the robot should go. Even though the robot only has two light
sensors and won't be able to "understand" what is happening at the level a
human with normal complex eyes can understand, these sorts of subtle
difference in behavior will create unique "fingerprints" (so to say) of
temporal light level changes which will most likely make it possible for a
robot with only two light level sensors to learn to follow one person, and
not the other (with enough training).

Of course, if you give it a million sensors instead of two, and give it an
eye to focus images on those sensors, then it will have a lot more data to
work with and far more unique patterns to pick up on.

But if you do the processing right, you should be able to do a lot more
with only two sensors than any of our current robots are doing - and it's
all a problem of extracting temporal patterns from multiple sensors and
corrolating those patterns to behaviors and rewards.

I don't see the complexity in connecting two
stimuli, I see the complexity in getting the
two stimuli to connect.

Making them connect is trivial so I don't get what point you are making
here. I don't see that as a problem at all. Connect two sensory inputs
with _any_ function (sum, product, AND, OR, NOR, whatever), and produce an
output of the function, and use that output to control behavior, and you
have _connected_ the two sensory signals. What's so hard about that?
Making them connect is trivial.

Making it find the right connection function is where all the interesting
problems show up, not in making them connect. Making them connect is
trivial.

It's also trivial to understand how to make them find the right connection
function for small problems as well. If you have two single bit binary
input sensors then you only have 4 different states the inputs can be in
and 2^4 different truth tables (aka functions) possible if you use only
spatial functions instead of temporal functions. The search space of all
possible functions to try is only 16. For an RL problem where you are
trying to find the function which maximizes reward, the approach is kinda
obvious. You collect data about how much reward each function has produced
in time and bias the selection of behavior to the functions which have
produced the most reward. End of problem.

None of that is complex and that is _exactly_ how sensory data is combined.

It only gets interesting (or hard) when the search space of all functions
possible becomes very large. When there are billions and billions of
possible functions, how to you track what works and what doesn't work? How
do you know what has been tried and what hasn't? When you include temporal
pattern functions instead of just spatial patterns, how large of a temporal
pattern do you include? Minutes, hours, years? The father back in time
you go, the larger the search space becomes. How does the search system
trade off depth of temporal pattern vs width of spatial pattern as it
attempts to search the space of all possible functions to find the ones
which produce the most reward?

How you connect two stimuli is trivial. How you find the _right_
connection function, is where all the fun starts.

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

Relevant Pages

  • Reply to Wolf
    ... pulses in response to the current energy inputs and recent ... happens to them as they travel from the sensors to the ... different keys close contacts in different patterns, ... primitive behaviors that could be chained and combined" (how ...
    (comp.ai.philosophy)
  • Re: Counting patters
    ... second sheet ... single row and delete data from three cells as you described. ... know how many patterns of this type are there. ... there are 32 sensors that reads a part and if the part ...
    (microsoft.public.excel.misc)
  • Re: Counting patters
    ... second sheet ... single row and delete data from three cells as you described. ... know how many patterns of this type are there. ... there are 32 sensors that reads a part and if the part ...
    (microsoft.public.excel.misc)
  • Re: Counting patters
    ... second sheet ... single row and delete data from three cells as you described. ... know how many patterns of this type are there. ... there are 32 sensors that reads a part and if the part ...
    (microsoft.public.excel.misc)
Loading