Stephen Reed wrote:
> > What kind of approach does JavaBayes use for handling very
> large Bayesian
> > nets? Gibbs sampling and the other standard MCMC methods seem
> not to scale
> > very well.
>
> I was not aware until now of the scaling issue you describe, but do
> envision that the Bayes nets used for causal deliberation will be modest
> in size and JavaBayes works very fast with the sample problems provided.
The "modest in size" criterion applies to many cognitive causal deliberation
problems, but when you get into the perceptual domain (the lower levels of
Albus's hierarchy, basically), you're going to have very large Bayes net
models, I believe. There is a lot of perceptual data out there. Or do you
plan to have low-level and mid-level perception handled by methods other
than Cyc-based inference?
I guess you're a long way from thinking detail about doing perception from
real, information-rich sensors?
> > I guess that's an OK behavioral framework... the big problem
> though is how
> > you're going to *learn* perceptual and behavioral schemata
> ("subprograms")
> > within Cyc.
>
> This is obvious to me but I am *so* naive in this area - Cyc will learn by
> asking for help when it reaches an impasse, and Cyc will learn by
> experience when making choices about goals to pursue and actions to
> accomplish those goals.
Learning by experience is definitely the right philosophy, but the history
of reinforcement learning in AI is very weak.
Flexible incorporation of diverse background knowledge (which Cyc should
excel at) is part of the answer, but I feel it's only *part* ....
This is the stickiest issue we've come up against in Novamente, and we have
some deep, creative ideas here that are not yet really tested...
> > I strongly suspect that Cyc's collection of logical inference
> methods are
> > all badly inadequate for this purpose.... Do you disagree?
>
> Yes, I agree that passive behavior, and an expressive deductive inference
> engine are inadequate, but I hope that Cyc's inference engine will provide
> an excellent foundation upon which to construct the Cognitive Cyc
> application - which brings in deliberative goal seeking, fuzzy logic,
> Bayesian modeling, multi-resolutional hierarchical control structure,
> and reflective (self-aware) behavior. I expect that a very large set of
> probabilistic relevance meta-assertions will be learned by Cyc
> in order to provide the focus behavior called for by the NIST/Albus
> Reference Model Architecture. These meta assertions would provide focus
> during deductive inference by filtering the permitted set of ground facts
> and backchaining rules according to the current level of resolution (e.g.
> planet, continent, region, city, or neighborhood) and current situation.
I am not impressed with the performance of the logic-based planning engines
that I've read about or played with. Given complex planning problems based
on large volumes of highly uncertain data, they all basically fail.
However, as I said above, I think that integration of broad background
knowledge into the process is one important ingredient needed for success.
And your approach with "probabilistic relevant meta-assertions" should allow
this nicely.
I have to say, I was very unimpressed with Cyc until I started hearing about
Cognitive Cyc from you. I'm still not in love with the Cyc approach, but
the Cognitive Cyc project does seem like a sensible attempt to build towards
AGI on the foundation of the Cyc knowledge base & inference engine. I look
forward to playing with & conversing with the system one day,
Ben G
-------
To unsubscribe, change your address, or temporarily deactivate your subscription,
please go to http://v2.listbox.com/member/
