On Sun, Oct 8, 2017 at 10:54 AM, Nil Geisweiller <ngeis...@googlemail.com> wrote:
> On 10/07/2017 04:39 PM, Linas Vepstas wrote> So here's a completely > different but related idea: First, use a crisp > >> reasoner to deduce what happens whenever strength>0.9999. Next, do it >> again, but now for strength>0.8. (but still using the crisp reasoner: just >> take strength>0.8 to mean "true"). This should have a "broader" set of >> consequences. Do it again for strength>0.6 - this causes even more >> possibilities to be explored. >> >> It seems like these three cases can be treated as "lower bounds" of what >> we might expect PLN to find. That these could be used to guide/limit what >> PLN explores. >> >> Alternately, if this was fast enough, you could do this 100 times for 100 >> different truth cutoffs, and build up a distributional TV... >> > > That's an interesting idea. You could > > 1. Sample the probability of each atom in the KB (axioms) according to > their TV > 2. Sample, according to this probabilities, whether the axiom is true or > false > 3. Run crisp-PLN over the dicretized theory, save the output > 4. repeat 2. N times to obtain a probability of the output > 5. repeat 1. M times to obtain a second order probability to regenerate > the output TV > > I suppose this type of crisp-PLN monte-carlos simulation should converge > to PLN. The advantage could be real though, assuming PLN complexity grows > with exp(alpha*L), and the complexity of crisp-PLN grows with exp(beta*L), > with beta < alpha, L being the length of the proof, we'd reach a point > where where M*N*exp(beta*L) < exp(alpha*L). > Ah, well, in this rephrasing, you've converted it into a probabilisitic-programming problem. I recall both you and I were at AGI 2015 and there were several papers on this, and clearly work stretching back a decade or more ... but that work, those papers were always focused on programming languages, and not on logic. I want to smack my forehead and say "but of course!" and wonder/marvel how it is that this hasn't been done before (maybe it has been, and we don't know?) Ben was mumbling something to me about adding probabilistic programming to opencog, but I did not understand what he was trying to achieve. This, by contrast, seems to be a well-defined, well-contained problem, which could give some decent results. It also has the benefit of allowing you to start with low values of M,N to get a rough estimate, and refine over time. So, the question is: what's the base tech? Starting with SAT solvers seems like too low a level. I like answer-set programming (ASP) because it explicitly deals with first-order logic and therefore is a natural fit for PLN. (and of course, the ASP solvers are now blazingly fast). A third possibility would be a theorem prover, like Coq or whatever, but these might be a poor fit for PLN. I dunno. > > Certainly an idea to keep in mind. More, my knee-jerk reastion is to say "its an idea we should prusue". To return to the original thread: in pursuing this idea, how much of it should be developed as in "independent module", and how much should be integrated with opencog? Certainly, it would be dumb to re-invent atoms yet again, but the idea of a stand-alone module seems to ask for that. --linas > > > Nil > > > >> I find this idea exciting! It seems plausible, doable ... >> >> --linas >> >> >> >> Nil >> >> -- *"The problem is not that artificial intelligence will get too smart and take over the world," computer scientist Pedro Domingos writes, "the problem is that it's too stupid and already has." * -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to opencog+unsubscr...@googlegroups.com. To post to this group, send email to opencog@googlegroups.com. Visit this group at https://groups.google.com/group/opencog. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAHrUA35D6dfy0YOXckfAxQ8oNf-sg3jWvvCxNDZMaurBxx1qyw%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
