Yep, thanks Rob... We've been keeping up on this stuff, and one of our team (Moshe) met with Pelikan a few months ago to discuss various directions for future work
What we've done is to apply BOA to the optimization of a special kind of "function trees" that represent computer programs using combinatory logic. So while BOA takes its inspiration from GA's, our use of BOA is more like Koza'z "genetic programming" (though our tree representation is very different from Koza's, using combinators and other functional-programming constructs to allow the creation of compact program trees encapsulating abstractions like loops and recursion). We've also had to extend BOA to deal with combinator trees whose inputs and internal constant-terms are not necessarily discrete variables, but may be floating-point variables or else semantic nodes or links drawn from Novamente's dynamic knowledge base. So we've needed to do some work embedding Novamente nodes and links in an appropriate metric space, so as to be able to do probabilistic instance generation on combinator trees with node/link inputs and/or constant terms. The big extension, which we haven't done yet but will do in late 2004, is to allow BOA to interact with Novamente's long-term memory, so that it uses probabilistic models learned on one problem, to help it figure out how to learn models for another problem. This will entail integration of BOA with our Probabilistic Term Logic framework, which is conceptually straightforward (they're both probabilistically based) but may require a bunch of fiddling... In sum, since we're concerned with integrating BOA into our overall Novamente framework and with using it for pattern-recognition and procedure-learning rather than generic optimization, the ways in which we're improving/extending it are kinda special... - Ben G > -----Original Message----- > From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] > Behalf Of Robert Stewart > Sent: Friday, August 06, 2004 12:08 PM > To: [EMAIL PROTECTED] > Subject: RE: [agi] Experiential interactive learning and Novamente > > > --- Ben Goertzel <[EMAIL PROTECTED]> wrote: > > > * an efficient algorithm for searching this subspace > > (an improvement of > > Pelikan and Goldberg's Bayesian Optimization > > Algorithm, enhanced to make use > > of long-term memory via invocation of probabilistic > > term logic) > > Pelikan et al. have been making significant > enhancements to BOA of their own recently. The > following papers look particularly promising: > > Pelikan, M., Tz-Kai Lin (2004). Parameter-less > hierarchical BOA. Genetic and Evolutionary Computation > Conference 2004 (GECCO-2004), Springer-Verlag, pp. > 24-35. > > Sastry, K., Goldberg, D.E., Pelikan, M. (2004). > Efficiency Enhancement of Probabilistic Model Building > Genetic Algorithms. IlliGAL Report No. 2004020, > Illinois Genetic Algorithms Laboratory, University of > Illinois at Urbana-Champaign, IL. > > http://www.cs.umsl.edu/~pelikan/publications.html > > Best, > > Rob > > > > > __________________________________ > Do you Yahoo!? > Yahoo! Mail - 50x more storage than other providers! > http://promotions.yahoo.com/new_mail > > ------- > To unsubscribe, change your address, or temporarily deactivate > your subscription, > please go to http://v2.listbox.com/member/[EMAIL PROTECTED] > > --- > Incoming mail is certified Virus Free. > Checked by AVG anti-virus system (http://www.grisoft.com). > Version: 6.0.733 / Virus Database: 487 - Release Date: 8/2/2004 > --- Outgoing mail is certified Virus Free. Checked by AVG anti-virus system (http://www.grisoft.com). Version: 6.0.733 / Virus Database: 487 - Release Date: 8/2/2004 ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
