Re: Goal of AI: Perfect or Bounded Rationality

"emotioncube" <rlong19@xxxxxxxxx> wrote:
Curt,

Would you subscribe to the proposition that there exists:

1. A reinforcement learning algorithm
2. A set of conditional reinforcements (possibly non-stationary)
3. A sequence of sensory input

such that a machine which learns according to the above would exhibit
behavior
indistiguishable from human behavior?

Well, yes and no.

Or are there behaviors that humans exhibit
which will cannot be accounted for by reinforcement learning alone?

Humans will kick their leg if you tap on it. They will bleed if you cut
them. They show unique facial expressions if you scare them. They are
likely to make noise if you break their fingers. They will grow if you
feed them. They might turn red if you expose them to too much UV light.
All these are behaviors of a human that has nothing to do with the
reinforcement learning I'm talking about.

However, all those and a million other things are innate behaviors produced
by our body. They are not the high level behaviors we create under the
control of reinforcement learning.

I believe that the right reinforcement learning algorithm, combined with
the the right set of reinforcers, placed in the correct environment (i.e. a
typical human environment), with the right set of sensors and effectors,
will produce all the interesting behaviors we tend to talk about as
intelligent human behavior. I believe it will create a C-3P0 like machine.
It will still be easy to test the machine to see if it's human (stick it
with a knife and see if it bleeds), and even a turning test will be hard
because you can ask it endless questions about human bodies and what it's
like to grow up as a human that it will have a hard time answering (unless
it has already spent 20 years learning to fool humans at that game). But
all the high level stuff we do (reasoning, planing, dreaming about the
future) as well as all the typical emotions we have will be present in the
machine if you get the reinforcement algorithm and the reinforcers correct.
(and the sensors and effectors to roughly match what humans have).

I think creating something with the general intelligent powers of a human
will be a lot easier than shaping it with similar enough learning skills
and similar enough reinforcers to allow it to act really human. Just like
different humans all tend to develop different personalities because of
some complex combination of nurture and nature, our first intelligent
machines will have very odd personalities because of they will have both a
very different set of reinforcers, different learning strengths, and a very
non-human experience (people won't treat it like they treat humans). All
that will add up to creating a very non-human personality even if it's got
roughly human level learning skills.

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

Relevant Pages