Re: Goal of AI: Perfect or Bounded Rationality

jalegris@xxxxxxxxxxxx wrote:
Curt Welch wrote:

All this type of behavior is easily explained by understanding that
humans are simply reinforcement learning machines.

We know how to build simple reinforcement learning machines - we just
don't yet know how to build ones with the full learning skills of
humans.

How do you reconcile such a broad-stroke characterization of behaviour
with the evidence that humans exhibit classical conditioning,
habituation and sensitization as well?

All those can be seen as implementation details of a reinforcement learning
machine.

The problem here is that you can use many different languages to describe a
complex object. There's no end to the number of ways people have created
to describe learning for example. (just check out the hundreds of human
learning theories). The question however is which language does the most
complete job of describing everything? I believe the language of
reinforcement learning is the most complete (and therefore the most
fundamental) of all the options.

You can for example easily explain classical conditioning in terms of the
operation of a reinforcement learning machine, but you can't explain
reinforcement learning in terms of classical conditioning.

And what about still
undiscovered aspects of behaviour?

We can only deal with those when they are discovered. I would expect any
new aspect of behavior to be explainable in terms of reinforcement. I
would expect that anyone who thinks they have created yet another theory of
learning to have just once again, reinvented the wheel by giving old ideas
new names.

But, unless it's something radically new, I'm inclined to believe that I
would just classify it as an implementation detail of how the human
reinforcement learning machine works.

You must keep in mind that "reinforcement learning" is more of a problem
description than an implementation. So to say that humans are
reinforcement learning machines is only a small step towards explaining
what type machines we are. It's like calling a car a self powered
locomotion machine. It hardly narrows down the scope of the problem at
all.

But, with the car, it makes it clear that you need an internal power
source, and a way to use that power, to create motion. If you haven't
addressed that, then you are a long way from building a car. If someone
for example was saying the correct way to build a car was to build a box
with seats inside, along with a steering wheel, gas pedal, and brake pedal,
I would ask them, "but where is the power source"? And if they didn't have
an answer to that, all there talk about a gas peddle would be a waste of
time.

Likewise, when building an intelligent machine, when people talk about all
sorts of minor stuff like memory and reasoning, and rational vs irrational
behavior, but never once acknowledge the fact that they must first be
building a reinforcement learning machine, I feel they are just as lost as
the person who talked about building a box with seats in it to create a car
who has no understanding that it must first have an internal power source
and drive train to convert the power to motion. All the other stuff is
added on top of that.

Where do you get your facts?

Who ever said they were facts? :) It's just my opinion.

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

Relevant Pages

  • Re: Goal of AI: Perfect or Bounded Rationality
    ... humans are simply reinforcement learning machines. ... We know how to build simple reinforcement learning machines - we ... future use of behaviors that our rewarded. ...
    (comp.ai.philosophy)
  • Re: Goal of AI: Perfect or Bounded Rationality
    ... humans are simply reinforcement learning machines. ... We know how to build simple reinforcement learning machines - we ... future use of behaviors that our rewarded. ...
    (comp.ai.philosophy)
  • Re: Existential risks
    ... Reinforcement learning machines have no desire to survive. ... They only have the innate desire to produce behaviors which ... Why on earth would I ever suggest the RLM didn't have the power to ...
    (comp.ai.philosophy)
  • Re: Goal of AI: Perfect or Bounded Rationality
    ... humans are simply reinforcement learning machines. ... We know how to build simple reinforcement learning machines - we just ... The question however is which language does the most ...
    (comp.ai.philosophy)
  • Re: Curts AI epiphany
    ... All your mentalese maps straight to reinforcement learning algorithms ... They would then try to build hardware which was based on the ... Instead, look at the behavior of humans, and forget about what might be ... reward signal is generated by hardware separate from the learning hardware. ...
    (comp.ai.philosophy)