>
> Ben, I'm guessing you've thought alot about how to structure the
> reward/goal system of Novamente.  I'd love to hear more about it.
>  It seems that designing a system that forces itself to expand
> its knowledge base is a fairly non-trivial task.  We as entities
> (demonstrated also in rats) have a certain prediliction for
> exploring novel situations, environments, objects, ideas, etc.
> Have you implemented a similar drive for seeking novelty in Novamente?
>
> -Brad

The whole story of Novamente motivations is complicated, and summarizing it
optimally in general nontechnical terms would take some real thought...

Here is a brief attempt though...

Firstly, Novamente will start out with a lot of cognitive processes running.
There are GoalNodes, which contain (rather roughly speaking) predicates that
they "want" to see evaluate to True.  These GoalNodes stimulate the learning
of procedures that, the system estimates, are likely to cause the GoalNodes'
predicates to evaluate to true.  This GoalNode-stimulated procedure learning
process is only one part of the dynamics of the system.

Now, the non-GoalNode processes could still be interpreted as acting in the
service of various "goals."  Even though they are not explicitly implemented
this way, nor viewed by the system this way.  They could be viewed as acting
towards goals such as: seeking to recognize patterns in the world, seeking
novelty, creating a coherent world-model, etc.

When an advanced Novamente rewrites its cognitive processes, it could
rewrite them so that the whole system is explicity goal-oriented, OR so that
none of it is (it could delete GoalNode and create nothing comparable).  It
could create new GoalNodes inconsistent with its previous ones, and not just
through error, but through influence of non-GoalNode-driven processes --
though it's not likely to create new GoalNodes *wildly* inconsistent with
its previous ones.

Specific GoalNodes that we plan to include initially include a "seek reward"
goal like AIXI's (our initial teaching interface, recently coded but not to
be in serious use for a while, has a reward/punishment slider), a goal for
maximizing perceived human happiness, a goal for seeking novelty, and a goal
for learning a lot.  Plenty of more specific ideas for initial GoalNodes
have been tossed around, but we're not experimenting with this stuff yet.
In the Webmind AI system, we had slightly different GoalNodes, and did
experiment with them some.

A deep, important issue is the temporal nature of goals embodied in
Goalnodes.  Does the system seek to make a predicate (embodied in a
GoalNode) true in the short term, or true in the long term, in the medium
term, etc.  GoalNodes may have parameters temporally biasing the time scale
on which they seek to satisfy their internal predicate.  (This is done by
making the predicate depend on the current time...).  Thus there may be many
GoalNodes that are sorta seeking the same thing, but on different (fuzzy)
time scales....

-- Ben


-------
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Reply via email to