Re: Temporal Learning

On 7 Oct 2005 15:38:47 -0700, humiguel@xxxxxxx wrote:

>Traveler <traveler@xxxxxxxxxx> wrote:
>
>>On 4 Oct 2005 10:35:30 -0700, "feedbackdroids"
>><feedbackdroids@xxxxxxxxx> wrote:
>>
>>>I think what might be a reasonable approach regards adding something
>>>like your net to a creature such as Aibo, would be to follow the same
>>>course that evolution/genetics already followed.
>>
>>To tell you the truth, I have found evolution to be rather useless in
>>my quest to understand intelligence, especially human intelligence.
>>It's like understanding Windows XP. Do I need to know that XP evolved
>>from Windows 1.0? I don't think so.
>
>What Dan suggests makes a lot of sense. Why waste time trying to
>solve a problem that is already solved by other means?

We don't know what the solution is, IMO. Otherwise, we would have
general AI by now.

>Would the AI
>solution to the low-level stuff add something to the existing one? I
>don't think so, on the contrary: Using AI algorithms to solve what
>is essentially a control-engineering problem would almost certainly
>require more computing power and that could be a serious hindrance,
>especially in mobile applications where space, weight and power are
>at a premium.

Computing power has nothing to do with solving the AI problem, IMO. If
we knew the solution, we would accomplish amazing things even with one
or two million neurons.

>Another concern is that even if you manage to solve the low-level
>stuff, not a small task in itself, there is no guarantee whatsoever
>that what you get will be usable at higher levels.

This is precisely my point. Innate behavior is made of neurons and
their connections. Learned behavior is exactly the same thing, once it
is learned. Innate behavior in animals is there for survival purposes:
procreation, eating, chasing prey, avoiding predators, etc... With
reinforcement learning, we can make robots that learn
self-preservation and do what we want them to do. A RL mechanism is
not a behavior, it is a mechanism of learning and adaptation, i.e.,
for creating behaviors. We know it works from observing ourselves and
other animals.

> If that is the
>case, you would have spent a lot of effort for nothing.

Well, we have. GOFAI was and is an awful waste of time and money.

> On the
>contrary, if you start at a higher level, even if the solution isn't
>usable at lower levels, at least you got a useful system. If you
>really got the theoretical part right, it should be no more difficult
>to work at the higher level.

This does not make sense to me. If we figure out how to make a machine
learn and adapt, why start at the high level? We would have found the
solution to the problem.

>[snip]
>
>>I see your point but I don't think it's needed. We don't need
>>pre-existing behaviors, IMO. We need principles of learning,
>>conditioning, goal-seeking, concept formation, etc... In fact, the
>>highest intelligence on the planet happens to be the one with the
>>least number of genetically programmed behaviors: humans.
>
>I have some difficulty in accepting this view. Most of the
>high-level behaviours exhibited by humans, the really interesting
>stuff, namely language, are acquired from the environment though
>imitation of our peers.

They are acquired through learning. IMO, over 90% of what a human
adult knows is learned after birth. The part that is genetically
programmed can also be learned by our robots. Again, let me emphasize
that a learning mechanism is not a behavior. It is the engine that
builds behaviors. That's what people like Curt and I are after.

> This is a trait common to other animals (it
>is clearly visible especially in young animals) and as I understand
>it that is an innate behaviour. Do you have an alternative strategy
>to collect behaviour from the environment?

No. Why should I? The holy grail of AI is the creation a mechanism of
general learning and adaptation through conditioning. This is should
be our only focus, IMO. Everything else is a waste of time.

Louis Savain

Why Software Is Bad and What We Can Do to Fix It:
http://www.rebelscience.org/Cosas/Reliability.htm
.

Relevant Pages

  • Re: Is there a purpose to our existense
    ... >> A goat who accidentally eats a thistle and gets a sore mouth will learn ... There are many ways in which a goat, or most other animals, ... system containing only a few thousand neurons, is capable of learning. ... Learning, including learning from one's own mistakes, is a commonplace ...
    (talk.origins)
  • Re: Darwin, Evolution, the Animal Kingdom, and Man
    ... >> offspring who can imitate what adults do have an advantage. ... >> You haven't watched young animals much, ... I like this as a way of understanding some learning. ...
    (sci.cognitive)
  • Re: OT -- Dog lovers in the house?
    ... I know vet schools take ... in very sick animals much like any university hospital for interns to ... I imagine learning how to give a physical examination requires a very ... pressing against the stomach to save my life. ...
    (rec.pets.cats.anecdotes)
  • Re: Biggest Mistakes Learning Guitar, Episode 1
    ... glaring mistakes. ... MISTAKE in learning you've ever encountered, ... Anyone who's read a 'good' tab of Voodoo ... I found (imo) that the key to starting off is learning the CHORDS (and/or ...
    (alt.guitar.beginner)
  • Re: recognition of individuals theory
    ... humans and other mammals begin with a VC that preferentially ... learning (IMO, that's still a debatable point), but it's not learning. ... The post to which I responded had nothing to do with robotic vehicles. ...
    (sci.cognitive)
Loading