Re: What did that thread indicate?

HMS Beagle <bgates@xxxxxxxxxxxxx> wrote:

>If you scroll up on your thread list, you will see the thread I
>started called "Ghost of GOFAI".
>
>The purpose of that thread was to confirm something I had been
>suspicious of since I came here.
>
>I have concluded that the behaviorists in this channel (spearheaded by
>Sizemore and Welch) have a theory of the brain from the early 20th
>century. And that they and their behaviorist cohorts defend its
>tenets to the death.
>
>But although they have a theory they defend so adamantly, when it
>comes down to the point in which they have to demonstrate SPECIFIC,
>PARTICULAR BEHAVIORAL POWERS this framework gives to a robot, they
>all fall conspicuously silent.
>
>The only response I got was a little peep from Welch, saying it
>"eliminates the programmer". My challenge is basically this, if
>you say that you can get a robot to do some behavior B using neural
>networks, I say I can produce an algorithm that does behavior B as
>well, using good old-fashioned programming. If you think you have
>an exemplar B that you think fails my claim, your time to post it was
>in the FORMER thread, not this one. Do not respond this post if you
>have an exemplar behavior B. Instead... post it where it belongs, in
>my initial challenge post, under the heading "Ghost of GOFAI".

I'm not sure if this is the correct way to address the problem. The
power of an intelligent system doesn't lie in its ability to
generate behaviour that you cannot program directly. Its power lies
in the ability to generate secondary behaviours that you don't have
to put in.

Let me illustrate with an example of an NLP system. You can directly
program the system to answer a particular question. You can also
program it to answer all the possible variants of the same question,
but you don't want to do that: You expect an intelligent system to
do that by itself.

The ability to generate secondary behaviours is where GOFAI fall
flat on its face. You cannot program directly all the behaviour that
an intelligent machine will ever need. You have to rely on the
machine's ability to generate behaviour.

A reinforcement-learning machine can, at least in principle,
generate secondary behaviour by combining primary behaviours,
testing that in the environment and learning from the feedback. The
challenge here is, of course, to devise suitable strategies for
generating secondary behaviour. Curt's ideas about this are so
silly that they don't deserve further consideration.

Antonio Esteves

--
Corby - A new approach to Artificial Intelligence
http://futalgo.planetaclix.pt/corby/index.htm

.

Quantcast