On 02/06/06, Richard Loosemore <[EMAIL PROTECTED]> wrote:
Will,
Comments taken, but the direction of my critique may have gotten lost in
the details:
Suppose I proposed a solution to the problem of unifying quantum
mechanics and gravity, and suppose I came out with a solution that said
that the unified theory involved (a) a specific interface to quantum
theory, which I spell out in great detail, and (b) ditto for an
interface with geometrodynamics, and (c) a linkage component, to be
specified.
I have argued with behaviourists that we are not at a point to explain
intelligence and that the operant and classical conditioning do not
fully explain intelligence either. So your analogies with GUT do not
apply to me so much as most behaviourists.
Physicists would laugh at this. What linkage component?! they would
say. And what makes you *believe* that once you sorted out the linkage
component, the two interfaces you just specified would play any role
whatsoever in that linkage component? They would point out that my
"linkage component" was the meat of the theory, and yet I had referred
to in such a way that it seemed as though it was just an extra, to be
sorted out later.
I agree the stuff to be sorted out later needs is essential and
important and distinctly non-trivial. I never claimed my architecture
was a complete design for an AGI. Rather it is a path of research,
that if you read it should strike you as different from other RL
approaches.
This is exactly what happened to Behaviorism, and the idea of
Reinforcement Learning. The one difference was that they did not
explicitly specify an equivalent of my (c) item above: it was for the
cognitive psychologists to come along later and point out that
Reinforcement Learning implicitly assumed that something in the brain
would do the job of deciding when to give rewards, and the job of
deciding what the patterns actually were .... and that that something
was the part doing all the real work. In the case of all the
experiments in the behaviorist literature, the experimenter substituted
for those components, making them less than obvious.
Exactly the same critique bears on anyone who suggests that
Reinforcement Learning could be the basis for an AGI.
It depends what you mean by basis. To me saying an agi is based on an
architecture that has a slight resemblance to a RL algorithm is about
the same as saying it is based on a turing complete machine. I.e. next
to nothing about the actual behaviour/programming of the system.
That is because each general intelligence needs different initial
programming dependent upon how it is embodied. One that gets its
energy from food will need to know how to acquire food so it can
continue to function while learning about the environment. Similarly a
electronic robot will need to know how and when to come back to a base
station. Then you get into all sorts of things like whether to imprint
on someone, how to recognise them and what sort of social learning is
important.
This is all important but at a higher level to the reinforcement I am
interested in and environment/body specific. A new robot on mars will
have to have vastly different expectations of the world around it
compared to a human baby on earth. So any theory of AGI cannot make
explicit any of these important things that the system does, else it
loses generality and only applies to baby-like intelligences or
mars/robot-like.
I do not believe
there is still any reply to that critique.
I think about the reinforcement architecture I am interested as a
keystone, that is useless on its own but vital to any particular AGI
effort. So I degrade the importance of RL, compared to other RL
people, and see it as a tiny step towards creating an AGI.
Whether you accept that as a reply I don't know.
Will
-------
To unsubscribe, change your address, or temporarily deactivate your subscription,
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
