On 1/20/07, Benjamin Goertzel <[EMAIL PROTECTED]> wrote:
A) This is just not true, many commonsense inferences require significantly more than 5 applications of rules
OK, I concur. Long inference chains are built upon short inference steps. We need a mechanism to recognize the "interestingness" of sentences. So we only keep the interesting/relevant ones and build more deductions upon them. It's not easy, OK. The bottomline is that the knowledge acquisition project is *separable* from specific inference methods.
B) Even if there are only 5 applications of rules, the combinatorial explosion still exists. If there are 10 rules and 1 billlion knowledge items, then there may be up to 10 billion possibilities to consider in each inference step. So there are (10 billion)^5 possible 5-step inference trajectories, in this scenario ;-) Of course, some fairly basic pruning mechanisms can prune it down a lot, but, one is still left with a combinatorial explosion that needs to be dealt with via subtle means...
This is really solved. Use a simple hashing of predicates, which is part of the rete algorithm. For example, you want to deduce whether "dead birds can fly". There may be 5000 rules/facts about birds, and 10000 rules/facts about dead things. (Reasonable?) So only these rules/facts would be checked, the other parts of the KB (about flowers, cats, etc) are completely untouched (unsearched).
Please bear in mind that we actually have a functional uncertain logical reasoning engine within the Novamente system, and have experimented with feeding in knowledge from files and doing inference on them. (Though this has been mainly for system testing, as our primary focus is on doing inference based on knowledge gained via embodied experience in the AGISim world.)
Do you think the problems you encounter in Novamente are really due to combinatorial explosion or rather the *lack* of the right rules/facts?
The truth is that, if you have a lot of knowledge in your system's memory, you need a pretty sophisticated, context-savvy inference control mechanism to do commonsense inference.
I'm thinking about simple questions like "can dead birds fly?" etc. It shouldn't involve more than 5 steps. What you're talking about seems to be chaining many such steps to solve a detective story, that kind of thing. And yes, for that you need sophisticated inference mechanisms.
Also, temporal inference can be quite tricky, and introduces numerous options for combinatorial explosion that you may not be thinking about when looking at atemporal examples of commonsense inference. Various conclusions may hold over various time scales; various pieces of knowledge may become obsolete at various rates, etc.
Think 4D. Time is just another dimension. If you can do spatial reasoning you can do temporal reasoning. It *has* got to be the same, thanks to Einstein. If your approach uses special tricks to deal with temporal, then at least it is not an elegant solution. I'm not arrogant, and I admit I have not fully solved this 4D problem. It is kind of tricky, but I'm optimistic about it.
I imagine you will have a better sense of these issues once you have actually built an uncertain reasoning engine, fed knowledge into it, and tried to make it do interesting things.... I certainly think this may be a valuable exercise for you to do. However, until you have done it, I think it's kind of silly for you to be speaking so confidently about how you are so confident you can solve all the problems found by others in doing this kind of work!! I ask again, do you have some theoretical innovation that seems probably to allow you circumvent all these very familiar problems??
I've given brief answers to your earlier questions, hope they're convincing. The point is that I think a simple inference engine combined with a good, *densely* populated knowledgebase can accomplish a lot. Let me stress again that this project per se is only a collection of facts/rules. Other, more intelligent people may come up with a better AGI to use this database. As for myself, I do use some innovative ideas (eg use of uncertain logic (not my invention though)) in my AGI that makes it different from GOFAI. My knowledge representation is not just a bunch of logic formulae. The logic formulae can reference each other so they form an intricate network similar to your graphical representation. I'd love to talk more about these things. But I think it's better to actually start a project and do some damned programming. So far I still believe the project is worth doing; and if my inference engine sucks, the database could still be of use to others... YKY ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
