Re: Goal of AI: Perfect or Bounded Rationality

"adityar7" <adityar7@xxxxxxxxx> wrote:
[quote]Thus I do not believe morality is an irrational behavior
when we are unable to make a more reasoned decision.

Because of our limitations and the complexity of the
world we have a set of emotions that guide us most
of the time toward optimal behavior both against the
elements (such as fear of heights) and toward our
fellow human being (caring for each other under
certain circumstances). [/quote]

But a situation where one can steal from one's neighbor is not a case
of our limitations or the complexity of the world. One can easily steal
and gain something, but chooses not to do so. That seems to be
irrational behavior.

You can say that one does not steal due to fear of punishment, but how
does that apply to AI? How do you punish machines? In fact, a machine
may not even understand pain despite being intelligent.

So how to explain irrational behavior such as the above example?

With reinforcement learning machines.

They produce behavior based on a long history of past experience. You
punish a reinforcement learning machine by sending it a punishment signal
which will reduce the probability of it producing similar behaviors in the
future for similar situations.

We don't steal because most of us learned that taking things that don't
belong to us leads to less rewards in the long run. We don't reach over
and grab food from our friend because we learned long ago that bad things
happen to us when we do things like that. If nothing else, they stop being
our friend and they stop doing all the nice things they have done in the
past for us (like sharing their food with us when we don't have any).

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.

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