Re: Cool visual illusion

"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:

> It is in the implementing that you realize
> just how vague these descriptions are. That
> is why I suspect Curt can't translate his
> RL into a working net.

Well, it's already translated into a working net. So that's not the issue.
The issue is does it work anything like humans. :)

> Another reason is I
> think he is plain wrong to think all learning
> is RL.

My network includes non-RL learning. So where is it that I have ever said
that "all learning is RL"?

What I stress about RL is that it's the key type of learning that creates
AI. Without it, you just don't have what I believe people are trying to
talk about when they talk about intelligence.

> I suspect we can learn just by being
> exposed to an input with constraint as an
> automatic result of the way we are wired.

Which is exactly why my network includes the non RL learning system for
adjusting the sorting gap. It's there to adjust the behavior of the
network to the characteristics of the input data. It does it in order to
maximize the amount of information extracted from the signal to base it's
RL behavior on.

The point is that you can't explain intelligent behavior in terms of
adapting to the constraints of the input data alone. It just gets you
nowhere. If you have two inputs, and the "constrants" of the data make
input 1 twice as active as input 2, then all the system can "learn" is that
input 1 is twice as active as input 2. Nothing in that knowledge tells the
machine how to behave in response to the data. No matter how much
information is extracted about the constants of the input data, or the
constrains about how the outputs effect the inputs, the machine still has
no clue about what to do with with that knowledge.

That's where RL comes in. It gives the machine a purpose. Without a
purpose the machine has no reason to pefer lifting an arm, chasing food, or
sitting on a couch watching TV. It has no way to make a choice on it's
own.

The most general way to explain all systems that have the power to change
their behavior in response to the environment is RL.

So, any type of system you make which "learns" the constraints of the
system, but also include something that tells it what to do with that
knowledge. And that system, will always be, a RL system.

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

Relevant Pages

  • Re: Cool visual illusion
    ... > My network includes non-RL learning. ... > to the characteristics of the input data. ... > behavior in terms of adapting to the constraints ...
    (comp.ai.philosophy)
  • Re: What did that thread indicate?
    ... > problems at hand when there exist preprocessor modules that can ... the dimensional space is enormous. ... Change the input data, and they quickly crash. ... learning in GOF-connectionist nets - esp when dealing with temporal ...
    (comp.ai.philosophy)
  • JMLR: Bayesian Network Learning with Parameter Constraints
    ... The Journal of Machine Learning Research is pleased to ... Bayesian Network Learning with Parameter Constraints ... incorporating domain knowledge into learning algorithms, ...
    (comp.ai)
  • Re: Common Lisp from a Unix perspective - barriers to using CL
    ... Best then to choose a learning environment with a built ... in lisp-aware editor and not waste time with configuration issues. ... with constraints that can't be cleanly factored out of the learning ...
    (comp.lang.lisp)
  • Restricting to real solutions in a maximization problem
    ... I'm just learning to use matlab and I'm trying to maximize a function ... subject to some constraints using 'fminunc'. ...
    (comp.soft-sys.matlab)