Re: [agi] would anyone want to use a commonsense KB?
On 3/5/08, david cash [EMAIL PROTECTED] wrote: In my opinion, instead of having to cherry-pick desirable and undesirable traits in an unconscious AGI entity, that we, of course, wish to have consciousness and cognitive abilites like reasoning, deductive and inductive logic comprehension skills, emotional traits, compassion, ethics, street smarts, and the like, requires the following protocols: We model AGI after the most successful living (and non-living) humans. That way, we will be able to pretty much predict how they will react in various situations; for example, we would not want to make a George W. Bush AGI. We would, however, make the Bill Clinton, David m. Cash, Angelina Jolie, and so on. Why reinvent the wheel when we have perfect examples of successful intelligence in front of us everyday and in our history books. There will have to be some ground rules though: no major world religious figures, no rapists or child molestors, etc. I will begin work on compiling a list of likely candidates for the new project, which is already underway in Finland with several of my colleagues. By the way, my name is David M. Cash, otherwise known on the Internet as DC or DMC. On a final note, I feel that each AGI should have an independent personality, not some slavelike borgish slave mind that is easily manipulated or controlled. Furthermore, in creation of this industry standard framework for the AGI mind, I elect we only use Commercial Open Source Software or regular Open Source Software. This project is too important to be left in the hands of just one or a few corporations (i.e. Microsoft, Cisco, Sun) who would inevitably taint the project with corporate greed and shareholder interests, which would hamper and ultimately destroy the proper development of the new Atlantis Project. So let's call it Project Atlantis from now on if that is ok with everyone... Calm down -- no one is a rapist here (AFAIK). I just used that as an example. Secondly, a commercial AGI scenario may not be that bad, it may even be more stable than a scenario where AGI is free for all. The latter is not very likely since the AGI needs to run on hardware, and hardware is not free. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 3/4/08, Mark Waser [EMAIL PROTECTED] wrote: But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. An excellent point. But what if the representation is natural language with pointers to the specific intended meaning of any words that are possibly ambiguous? That would seem to be the best of both worlds. Yes, that's the very same strategy I can think of. Not sure if there're better ways. The problem here is that the decompression algorithm seems to be very complex. The algorithm to compute the combination of two concepts A and B goes like this: 1. generate random sentences containing A and B 2. test these sentences to see if they make sense (to make sense means to be supported by, or be consistent with, other facts/rules). Such an algorithm may be very time-consuming -- this may explain why *reading* NL is a slow task for humans -- we need to find abductive explanations for the texts. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 3/4/08, Ben Goertzel [EMAIL PROTECTED] wrote: Rather, I think the right goal is to create an AGI that, in each context, can be as ambiguous as it wants/needs to be in its representation of a given piece of information. Ambiguity allows compactness, and can be very valuable in this regard. What exactly does ambiguity mean? If it means that sentences in the KR have multiple meanings, then all KRs are necessarily ambiguous. This is because information content of the world information content of the representation. The question is whether we should use a NL-like KR. If we use NL as the KR, every thought would involve abductive interpretation, which is very time-consuming. An alternative is to decompress NL texts and then compress the information again using *another* format. This new format may be easier to decompress later. (NL is difficult to decompress). Maybe the human brain's KR is also different from NL... YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. An excellent point. But what if the representation is natural language with pointers to the specific intended meaning of any words that are possibly ambiguous? That would seem to be the best of both worlds. - Original Message - From: YKY (Yan King Yin) To: agi@v2.listbox.com Sent: Monday, March 03, 2008 5:03 PM Subject: Re: [agi] would anyone want to use a commonsense KB? On 3/4/08, Mike Tintner [EMAIL PROTECTED] wrote: Good example, but how about: language is open-ended, period and capable of infinite rather than myriad interpretations - and that open-endedness is the whole point of it?. Simple example much like yours : handle. You can attach words for objects ad infinitum to form different sentences - handle an egg/ spear/ pen/ snake, stream of water etc. - the hand shape referred to will keep changing - basically because your hand is capable of an infinity of shapes and ways of handling an infinity of different objects. . And the next sentence after that first one, may require that the reader know exactly which shape the hand took. But if you avoid natural language, and its open-endedness then you are surely avoiding AGI. It's that capacity for open-ended concepts that is central to a true AGI (like a human or animal). It enables us to keep coming up with new ways to deal with new kinds of problems and situations - new ways to handle any problem. (And it also enables us to keep recognizing new kinds of objects that might classify as a knife - as well as new ways of handling them - which could be useful, for example, when in danger). Sure, AGI needs to handle NL in an open-ended way. But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. Otherwise, the knowledge stored in episodic memory would be open to interpretations and may need to errors in recall, and similar problems. YKY -- agi | Archives | Modify Your Subscription --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 04/03/2008, Mark Waser [EMAIL PROTECTED] wrote: But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. An excellent point. But what if the representation is natural language with pointers to the specific intended meaning of any words that are possibly ambiguous? That would seem to be the best of both worlds. This is fine provided that the AGI lives inside a chess-like ambiguity free world, which could be a simulation or maybe some abstract data mining environment. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Dead right (in an ambiguous way :) ) Basically an AGI without open-ended concepts will never live in the real world. I should add that I don't believe early, true AGI's *will* be anywhere near capable of natural language. All they will need is one or more systems of open-ended concepts. Emotions are one such system. A drive/urge of hunger for food is an open-ended concept that allows an animal to seek/eat any of a whole range of foods, (unless you're pregnant and it's 1 a.m. and only one particular form of chocolate will do - but even then it could be any of many brands). A body can be regarded as itself a system of open-ended concepts - for effecting concepts of how to seek goals. Hands and other limbs and indeed a torso offer a potentially infinite range of ways to effect commands to handle objects. They offer a roboticist many degrees of freedom - true mobility. (Do you think about the body, natural or robotic, from this POV, Bob?) Bob M: Mark Waser [EMAIL PROTECTED] wrote: But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. An excellent point. But what if the representation is natural language with pointers to the specific intended meaning of any words that are possibly ambiguous? That would seem to be the best of both worlds. This is fine provided that the AGI lives inside a chess-like ambiguity free world, which could be a simulation or maybe some abstract data mining environment. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
In my opinion, instead of having to cherry-pick desirable and undesirable traits in an unconscious AGI entity, that we, of course, wish to have consciousness and cognitive abilites like reasoning, deductive and inductive logic comprehension skills, emotional traits, compassion, ethics, street smarts, and the like, requires the following protocols: We model AGI after the most successful living (and non-living) humans. That way, we will be able to pretty much predict how they will react in various situations; for example, we would not want to make a George W. Bush AGI. We would, however, make the Bill Clinton, David m. Cash, Angelina Jolie, and so on. Why reinvent the wheel when we have perfect examples of successful intelligence in front of us everyday and in our history books. There will have to be some ground rules though: no major world religious figures, no rapists or child molestors, etc. I will begin work on compiling a list of likely candidates for the new project, which is already underway in Finland with several of my colleagues. By the way, my name is David M. Cash, otherwise known on the Internet as DC or DMC. On a final note, I feel that each AGI should have an independent personality, not some slavelike borgish slave mind that is easily manipulated or controlled. Furthermore, in creation of this industry standard framework for the AGI mind, I elect we only use Commercial Open Source Software or regular Open Source Software. This project is too important to be left in the hands of just one or a few corporations (i.e. Microsoft, Cisco, Sun) who would inevitably taint the project with corporate greed and shareholder interests, which would hamper and ultimately destroy the proper development of the new Atlantis Project. So let's call it Project Atlantis from now on if that is ok with everyone... On Tue, Mar 4, 2008 at 7:58 AM, Bob Mottram [EMAIL PROTECTED] wrote: On 04/03/2008, Mark Waser [EMAIL PROTECTED] wrote: But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. An excellent point. But what if the representation is natural language with pointers to the specific intended meaning of any words that are possibly ambiguous? That would seem to be the best of both worlds. This is fine provided that the AGI lives inside a chess-like ambiguity free world, which could be a simulation or maybe some abstract data mining environment. -- *agi* | Archives http://www.listbox.com/member/archive/303/=now http://www.listbox.com/member/archive/rss/303/ | Modifyhttp://www.listbox.com/member/?;Your Subscription http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
RE: [agi] would anyone want to use a commonsense KB?
I have been an admirer of your robotic works for many years but I just couldn't help ask you a question. I have created a few robots using sonar and other sensors and I totally agree that the real world is very messy. However, most of my thinking is at a much higher level than what my senses provide. The messy world is turned into my symbolic world through a complex maze of systems but in the end, at least my conscious brain, uses the symbols created, not messy reality. The idea that the whole AGI puzzle can be solved by exclusively using predicate logic, I agree doesn't fit with what we both know about sensor data in the real world. On the other hand, I don't see why a huge chunk (and I would argue the more difficult part) of the intelligence part of an AGI can't be done using words, numbers and pictures (idealized pictures I believe are more valuable than real ones) in a vast series of models. I can't see why AGI researchers would have to take an either/or approach to constructing their AGI when dealing with the real world AND intelligently correlating the symbols created. Having both seems like an obviously better approach. Neither approach by itself seems even remotely plausible to me. If both areas can't be had in a single researcher or group then I see no reason why each group can't do what they do best and then develop an interface so that the final AGI has the benefit of both. I am not proposing that all AGI designs are equal or useful and I am not proposing an AGI should be created using an amalgamation of all types of AGI. I am specifically referring to only symbolic intelligence (in some form) and the systems that would turn messy real world data into symbols that could be used by the symbolic system. I have great respect for people that are working on turning pictures and vision into useful symbolic objects. I believe this and speech recognition etc are greatly needed and would be very helpful to an AGI but I don't believe these problems need to be solved as a first step to AGI or that they are entirely necessary to at least get close to human level intelligence. Is a blind man who is also a paraplegic necessarily considered less intelligent than an able bodied person? David Clark From: Bob Mottram [mailto:[EMAIL PROTECTED] Sent: March-04-08 8:58 AM To: agi@v2.listbox.com Subject: Re: [agi] would anyone want to use a commonsense KB? On 04/03/2008, Mark Waser [EMAIL PROTECTED] wrote: But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. An excellent point. But what if the representation is natural language with pointers to the specific intended meaning of any words that are possibly ambiguous? That would seem to be the best of both worlds. This is fine provided that the AGI lives inside a chess-like ambiguity free world, which could be a simulation or maybe some abstract data mining environment. _ agi | http://www.listbox.com/member/archive/303/=now Archives http://www.listbox.com/member/archive/rss/303/ | http://www.listbox.com/member/?; Modify Your Subscription http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/28/08, Mark Waser [EMAIL PROTECTED] wrote: I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts There is a substantial amount of literature that claims that *humans* can't generate new ideas de novo either -- and that they can only build up new ideas from existing pieces. That's fine, but the way our language builds up new ideas seems to be very complex, and it makes natural language a bad knowledge representation for AGI. For example: An apple pie is a pie with apple fillings. A door knob is a knob attached to a door. A street prostitute is prostitute working in the streets. So the meaning of AB depends on the *interactions* of A and B, and it violates the principle of compositionality -- where the meaning of AB would be somehow combined from A and B in a *fixed* way. An even more complex example: spread the jam with a knife draw a circle with a knife cut the cake with a knife rape the girl with a knife stop the train with a knife (with unclear meaning) So the simple concept do X with a knife can be interpreted in myriad ways -- it generates new ideas in complex ways. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
YKY: the way our language builds up new ideas seems to be very complex, and it makes natural language a bad knowledge representation for AGI. An even more complex example: spread the jam with a knife draw a circle with a knife cut the cake with a knife rape the girl with a knife stop the train with a knife (with unclear meaning) So the simple concept do X with a knife can be interpreted in myriad ways -- it generates new ideas in complex ways. YKY, Good example, but how about: language is open-ended, period and capable of infinite rather than myriad interpretations - and that open-endedness is the whole point of it?. Simple example much like yours : handle. You can attach words for objects ad infinitum to form different sentences - handle an egg/ spear/ pen/ snake, stream of water etc. - the hand shape referred to will keep changing - basically because your hand is capable of an infinity of shapes and ways of handling an infinity of different objects. . And the next sentence after that first one, may require that the reader know exactly which shape the hand took. But if you avoid natural language, and its open-endedness then you are surely avoiding AGI. It's that capacity for open-ended concepts that is central to a true AGI (like a human or animal). It enables us to keep coming up with new ways to deal with new kinds of problems and situations - new ways to handle any problem. (And it also enables us to keep recognizing new kinds of objects that might classify as a knife - as well as new ways of handling them - which could be useful, for example, when in danger). - Original Message - From: YKY (Yan King Yin) To: agi@v2.listbox.com Sent: Monday, March 03, 2008 7:14 PM Subject: Re: [agi] would anyone want to use a commonsense KB? On 2/28/08, Mark Waser [EMAIL PROTECTED] wrote: I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts There is a substantial amount of literature that claims that *humans* can't generate new ideas de novo either -- and that they can only build up new ideas from existing pieces. That's fine, but the way our language builds up new ideas seems to be very complex, and it makes natural language a bad knowledge representation for AGI. For example: An apple pie is a pie with apple fillings. A door knob is a knob attached to a door. A street prostitute is prostitute working in the streets. So the meaning of AB depends on the *interactions* of A and B, and it violates the principle of compositionality -- where the meaning of AB would be somehow combined from A and B in a *fixed* way. An even more complex example: spread the jam with a knife draw a circle with a knife cut the cake with a knife rape the girl with a knife stop the train with a knife (with unclear meaning) So the simple concept do X with a knife can be interpreted in myriad ways -- it generates new ideas in complex ways. YKY -- agi | Archives | Modify Your Subscription -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.516 / Virus Database: 269.21.3/1308 - Release Date: 3/3/2008 10:01 AM --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 3/4/08, Mike Tintner [EMAIL PROTECTED] wrote: Good example, but how about: language is open-ended, period and capable of infinite rather than myriad interpretations - and that open-endedness is the whole point of it?. Simple example much like yours : handle. You can attach words for objects ad infinitum to form different sentences - handle an egg/ spear/ pen/ snake, stream of water etc. - the hand shape referred to will keep changing - basically because your hand is capable of an infinity of shapes and ways of handling an infinity of different objects. . And the next sentence after that first one, may require that the reader know exactly which shape the hand took. But if you avoid natural language, and its open-endedness then you are surely avoiding AGI. It's that capacity for open-ended concepts that is central to a true AGI (like a human or animal). It enables us to keep coming up with new ways to deal with new kinds of problems and situations - new ways to handle any problem. (And it also enables us to keep recognizing new kinds of objects that might classify as a knife - as well as new ways of handling them - which could be useful, for example, when in danger). Sure, AGI needs to handle NL in an open-ended way. But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. Otherwise, the knowledge stored in episodic memory would be open to interpretations and may need to errors in recall, and similar problems. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Sure, AGI needs to handle NL in an open-ended way. But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. Otherwise, the knowledge stored in episodic memory would be open to interpretations and may need to errors in recall, and similar problems. Rather, I think the right goal is to create an AGI that, in each context, can be as ambiguous as it wants/needs to be in its representation of a given piece of information. Ambiguity allows compactness, and can be very valuable in this regard. Guidance on this issue is provided by the Lojban language. Lojban allows extremely precise expression, but also allows ambiguity as desired. What one finds when speaking Lojban is that sometimes one chooses ambiguity because it lets one make ones utterances shorter. I think the same thing holds in terms of an AGI's memory. An AGI with finite memory resources must sometimes choose to represent relatively unimportant information ambiguously rather than precisely so as to conserve memory. For instance, storing the information A is associated with B is highly ambiguous, but takes little memory. Storing logical information regarding the precise relationship between A and B may take one or more orders of magnitude more information. -- Ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Ben Goertzel wrote: yet I still feel you dismiss the text-mining approach too glibly... No, but text mining requires a language model that learns while mining. You can't mine the text first. Agreed ... and this gets into subtle points. Which aspects of the language model need to be adapted while mining, and which can remain fixed? Answering this question the right way may make all the difference in terms of the viability of the approach... ben Given the history of evolution of language... ALL aspects of the language model need to be adaptive, but some need to be more easily adapted than others. E.g., adding words needs to be something that's easy to do. Combining words and eliding pieces more difficult (but that's how languages transition from forms without verb endings to forms with verb endings). E.g., the -ed past tense suffix of verbs is derived from the word did (as in derive did instead of derived in the previous sentence). If you go looking you find transitions where the order of subject, verb and object flip, and many other permutations. If you don't find a permutation, this doesn't mean it never happened and will never happen, but rather that most of the evidence is missing, so many rare events aren't recorded. There probably actually *are* some transitions that have zero probability, but we don't know what they are. So just make some transitions extremely improbable. (Who would have predicted l33tspeak ahead of time?) --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
My latest thinking tends to agree with Matt that language and common sense are best learnt together. (Learning langauge before common sense is impossible / senseless). I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts. I want to stress that AGI needs to be able to think at the WORD/CONCEPT level. In order to do this, we need some rules that *rewrite* sentences made up of words, such that the AGI can reason from one sentence to another. Such rewrite rules are very numerous and can be very complex -- for example rules for auxillary words and prepositions, etc. I'm not even sure that such rules can be expressed in FOL easily -- let alone learn them! The embodiment approach provides an environment for learning qualitative physics, but it's still different from the common sense domain where knowledge is often verbally expressed. In fact, it's not the environment that matters, it's the knowledge representation (whether it's expressive enough) and the learning algorithm (how sophisticated it is). YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Hi, I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts. Text mining is not an AGI approach, it's merely a possible way of getting knowledge into an AGI. Whether the AGI can generate new ideas is independent of whether it gets knowledge via text mining or via some other means... I want to stress that AGI needs to be able to think at the WORD/CONCEPT level. In order to do this, we need some rules that *rewrite* sentences made up of words, such that the AGI can reason from one sentence to another. Such rewrite rules are very numerous and can be very complex -- for example rules for auxillary words and prepositions, etc. I'm not even sure that such rules can be expressed in FOL easily -- let alone learn them! This seems off somehow -- I don't think reasoning should be implemented on the level of linguistic surface forms. The embodiment approach provides an environment for learning qualitative physics, but it's still different from the common sense domain where knowledge is often verbally expressed. I don't get your point... Most of common sense is about the world in which we live, as embodied social organisms... Embodiment buys you a lot more than qualitative physics. It buys you richly shared social experience, among other things. In fact, it's not the environment that matters, it's the knowledge representation (whether it's expressive enough) and the learning algorithm (how sophisticated it is). I think that all three of these things matter a lot, along with the overall cognitive architecture. -- Ben G --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts There is a substantial amount of literature that claims that *humans* can't generate new ideas de novo either -- and that they can only build up new ideas from existing pieces. Such rewrite rules are very numerous and can be very complex -- for example rules for auxiliary words and prepositions, etc The epicycles that the sun performs as it moves around the Earth are also very numerous and complex -- until you decide that maybe you should view it as the Earth moving around the sun instead. Read some Pinker -- the rules of language tell us *a lot* about the tough-to-discern foundations of human cognition. - Original Message - From: YKY (Yan King Yin) To: agi@v2.listbox.com Sent: Thursday, February 28, 2008 4:37 AM Subject: Re: [agi] would anyone want to use a commonsense KB? My latest thinking tends to agree with Matt that language and common sense are best learnt together. (Learning langauge before common sense is impossible / senseless). I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts. I want to stress that AGI needs to be able to think at the WORD/CONCEPT level. In order to do this, we need some rules that *rewrite* sentences made up of words, such that the AGI can reason from one sentence to another. Such rewrite rules are very numerous and can be very complex -- for example rules for auxillary words and prepositions, etc. I'm not even sure that such rules can be expressed in FOL easily -- let alone learn them! The embodiment approach provides an environment for learning qualitative physics, but it's still different from the common sense domain where knowledge is often verbally expressed. In fact, it's not the environment that matters, it's the knowledge representation (whether it's expressive enough) and the learning algorithm (how sophisticated it is). YKY -- agi | Archives | Modify Your Subscription --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/27/08, Ben Goertzel [EMAIL PROTECTED] wrote: YKY I thought you were talking about the extraction of information that is explicitly stated in online text. Of course, inference is a separate process (though it may also play a role in direct information extraction). I don't think the rules of inference per se need to be learned. In our book on PLN we outline a complete set of probabilistic logic inference rules, for example. What needs to be learned via experience is how to appropriately bias inference control -- how to sensibly prune the inference tree. So, one needs an inference engine that can adaptively learn better and better inference control as it carries out inferences. We designed and partially implemented this feature in the NCE but never completed the work due to other priorities ... but I hope this can get done in NM or OpenCog sometime in late 2008.. I'm not talking about inference control here -- I assume that inference control is done in a proper way, and there will still be a problem. You seem to assume that all knowledge = what is explicitly stated in online texts. So you deny that there is a large body of implicit knowledge other than inference control rules (which are few in comparison). I think that if your AGI doesn't have the implicit knowledge, it'd only be able to perform simple inferences about statistical events -- for example, calculating the probability of (lung cancer | smoking). The kind of reasoning I'm interested in is more sophisticated. For example, I may ask the AGI to open a file and print the 100th line (in Java or C++, say). The AGI should be able to use a loop to read and discard the first 99 lines. We need a step like: read 99 lines - use a loop but such a step must be based on even simpler *concepts* of repetition and using loops. What I'm saying is that your AGI does NOT have such rules and would be incapable of thinking about such things. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
I'm not talking about inference control here -- I assume that inference control is done in a proper way, and there will still be a problem. You seem to assume that all knowledge = what is explicitly stated in online texts. So you deny that there is a large body of implicit knowledge other than inference control rules (which are few in comparison). I think that if your AGI doesn't have the implicit knowledge, it'd only be able to perform simple inferences about statistical events -- for example, calculating the probability of (lung cancer | smoking). For instance, suppose you ask an AI if chocolate makes a person more alert. It might read one article saying that coffee makes people more alert, and another article saying that chocolate contains theobromine, and another article saying that theobromine is related to caffeine, and another article saying that coffee contains caffeine ... and then put the pieces together to answer YES This kind of reasoning may sound simple but getting it to work systematically on the large scale based on text mining has not been done... And it does seem w/in the grasp of current tech without any breakthroughs... The kind of reasoning I'm interested in is more sophisticated. For example, I may ask the AGI to open a file and print the 100th line (in Java or C++, say). The AGI should be able to use a loop to read and discard the first 99 lines. We need a step like: read 99 lines - use a loop but such a step must be based on even simpler *concepts* of repetition and using loops. What I'm saying is that your AGI does NOT have such rules and would be incapable of thinking about such things. Being incapable of thinking about such things is way too strong a statement -- that has to do with the AI's learning/reasoning algorithms rather than about the knowledge it has. I think there would be a viable path to AGI via 1) Filling a KB up w/ commensense knowledge via text mining and simple inference, as I described above 2) Building an NL conversation system utilizing the KB created in 1 3) Teaching the AGI the implicit knowledge you suggest via conversing with it As noted I prefer to introduce embodiment into the mix, though, for a variety of reasons... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
--- Ben Goertzel [EMAIL PROTECTED] wrote: For instance, suppose you ask an AI if chocolate makes a person more alert. It might read one article saying that coffee makes people more alert, and another article saying that chocolate contains theobromine, and another article saying that theobromine is related to caffeine, and another article saying that coffee contains caffeine ... and then put the pieces together to answer YES This kind of reasoning may sound simple but getting it to work systematically on the large scale based on text mining has not been done... And it does seem w/in the grasp of current tech without any breakthroughs... It could be done with a simple chain of word associations mined from a text corpus: alert - coffee - caffeine - theobromine - chocolate. But that is not the problem. The problem is that the reasoning would be faulty, even with a more sophisticated analysis. By a similar analysis you could reason: - coffee makes you alert. - coffee contains water. - water (H20) is related to hydrogen sulfide (H2S). - rotten eggs produce hydrogen sulfide. - therefore rotten eggs make you alert. Long chains of logical reasoning are not very useful outside of mathematics. I think there would be a viable path to AGI via 1) Filling a KB up w/ commensense knowledge via text mining and simple inference, as I described above 2) Building an NL conversation system utilizing the KB created in 1 3) Teaching the AGI the implicit knowledge you suggest via conversing with it I think adding common sense knowledge before language is the wrong approach. It didn't work for Cyc. Natural language evolves to the easiest form for humans to learn, because if a language feature is hard to learn, people will stop using it because they aren't understood. We would be wise to study language learning in humans and model the process. The fact is that children learn language in spite of a lack of common sense. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
It could be done with a simple chain of word associations mined from a text corpus: alert - coffee - caffeine - theobromine - chocolate. That approach yields way, way, way too much noise. Try it. But that is not the problem. The problem is that the reasoning would be faulty, even with a more sophisticated analysis. By a similar analysis you could reason: - coffee makes you alert. - coffee contains water. - water (H20) is related to hydrogen sulfide (H2S). - rotten eggs produce hydrogen sulfide. - therefore rotten eggs make you alert. There is a produce predicate in here which throws off the chain of reasoning wildly. And, nearly every food contains water, so the application of Bayes rule within this inference chain of yours will yield a conclusion with essentially zero confidence. Since fewer foods contain caffeine or theobromine, the inference trail I suggested will not have this problem. In short, I claim your similar analysis is only similar at a very crude level of analysis, and is not similar when you look at the actual probabilistic inference steps involved. Long chains of logical reasoning are not very useful outside of mathematics. But the inference chain I gave as an example is NOT very long. The problem is actually that outside of math, chains of inference (long or short) require contextualization... I think there would be a viable path to AGI via 1) Filling a KB up w/ commensense knowledge via text mining and simple inference, as I described above 2) Building an NL conversation system utilizing the KB created in 1 3) Teaching the AGI the implicit knowledge you suggest via conversing with it I think adding common sense knowledge before language is the wrong approach. It didn't work for Cyc. I agree it's not the best approach. I also think, though, that one unsuccessful attempt should not be taken to damn the whole approach. The failure of explicit knowledge encoding by humans, does not straightforwardly imply the failure of knowledge extraction via text mining (as approaches to AGI) Natural language evolves to the easiest form for humans to learn, because if a language feature is hard to learn, people will stop using it because they aren't understood. We would be wise to study language learning in humans and model the process. The fact is that children learn language in spite of a lack of common sense. Actually, they seem to acquire language and common sense together. But, wild children and apes learn common sense, but never learn language beyond the proto-language level. But I agree, study of human dev psych is one thing that has inclined me toward the embodied approach ... yet I still feel you dismiss the text-mining approach too glibly... -- Ben G --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
--- Ben Goertzel [EMAIL PROTECTED] wrote: It could be done with a simple chain of word associations mined from a text corpus: alert - coffee - caffeine - theobromine - chocolate. That approach yields way, way, way too much noise. Try it. I agree that it does to the point of uselessness. But that is not the problem. The problem is that the reasoning would be faulty, even with a more sophisticated analysis. By a similar analysis you could reason: - coffee makes you alert. - coffee contains water. - water (H20) is related to hydrogen sulfide (H2S). - rotten eggs produce hydrogen sulfide. - therefore rotten eggs make you alert. There is a produce predicate in here which throws off the chain of reasoning wildly. And, nearly every food contains water, so the application of Bayes rule within this inference chain of yours will yield a conclusion with essentially zero confidence. Since fewer foods contain caffeine or theobromine, the inference trail I suggested will not have this problem. But you cannot conclude from - coffee contains caffeine - coffee makes you alert that caffeine makes you alert either. I think adding common sense knowledge before language is the wrong approach. It didn't work for Cyc. I agree it's not the best approach. I also think, though, that one unsuccessful attempt should not be taken to damn the whole approach. Well, there are many examples of AI failures where we add knowledge first, then language (BASEBALL, Eliza, SHRDLU, hundreds of expert systems, etc). We do this because abstract knowledge is easy to represent efficiently on a computer. But the brain doesn't work this way. A language model has to work like the brain, because language is adapted to it. We don't do it that way because it requires such a huge amount of computation. yet I still feel you dismiss the text-mining approach too glibly... No, but text mining requires a language model that learns while mining. You can't mine the text first. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
yet I still feel you dismiss the text-mining approach too glibly... No, but text mining requires a language model that learns while mining. You can't mine the text first. Agreed ... and this gets into subtle points. Which aspects of the language model need to be adapted while mining, and which can remain fixed? Answering this question the right way may make all the difference in terms of the viability of the approach... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/25/08, Ben Goertzel [EMAIL PROTECTED] wrote: Hi, There is no good overview of SMT so far as I know, just some technical papers... but SAT solvers are not that deep and are well reviewed in this book... http://www.sls-book.net/ But that's *propositional* satisfiability, the results may not extend to first-order SAT -- I've no idea. Secondly, the learning of an entire KB from text corpus is much, much harder than SAT. Even the learning of a single hypothesis from examples with background knowledge (ie the problem of inductive logic programming) is harder than SAT. Now you're talking about inducing the entire KB, and possibly involving theory revision -- this is VERY impractical. I guess I'd focus on learning simple rules, one at a time, from NL instructions. IMO this is one of the most feasible ways of acquiring the AGI KB. But it also involves the AGI itself in the acquisition process, not just a passive collection of facts like MindPixel... YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Obviously, extracting knowledge from the Web using a simplistic SAT approach is infeasible However, I don't think it follows from this that extracting rich knowledge from the Web is infeasible It would require a complex system involving at least 1) An NLP engine that maps each sentence into a menu of probabilistically weighted logical interpretations of the sentence (including links into other sentences built using anaphor resolution heuristics). This involves a dozen conceptually distinct components and is not at all trivial to design, build or tune. 2) Use of probabilistic inference rules to create implication links between the different interpretations of the different sentences 3) Use of an optimization algorithm (which could be a clever use of SAT or SMT, or something else) to utilize the links formed in step 2, to select the right interpretation(s) for each sentence The job of the optimization algorithm is hard but not THAT hard because the choice of the interpretation of one sentence is only tightly linked to the choice of interpretation of a relatively small set of other sentences (ones that are closely related syntactically, semantically, or in terms of proximity in the same document, etc.). I don't know any way to tell how well this would work, except to try. My own approach, cast in these terms, would be to -- use virtual-world grounding to help with the probabilistic weighting in step 1 and the link building in step 2 -- use other heuristics besides SAT/SMT in step 3 ... but, using these techniques within NM/OpenCog is also a possibility down the road, I've been studying the possibility... -- Ben On Tue, Feb 26, 2008 at 6:56 AM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 2/25/08, Ben Goertzel [EMAIL PROTECTED] wrote: Hi, There is no good overview of SMT so far as I know, just some technical papers... but SAT solvers are not that deep and are well reviewed in this book... http://www.sls-book.net/ But that's *propositional* satisfiability, the results may not extend to first-order SAT -- I've no idea. Secondly, the learning of an entire KB from text corpus is much, much harder than SAT. Even the learning of a single hypothesis from examples with background knowledge (ie the problem of inductive logic programming) is harder than SAT. Now you're talking about inducing the entire KB, and possibly involving theory revision -- this is VERY impractical. I guess I'd focus on learning simple rules, one at a time, from NL instructions. IMO this is one of the most feasible ways of acquiring the AGI KB. But it also involves the AGI itself in the acquisition process, not just a passive collection of facts like MindPixel... YKY agi | Archives | Modify Your Subscription -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/26/08, Ben Goertzel [EMAIL PROTECTED] wrote: Obviously, extracting knowledge from the Web using a simplistic SAT approach is infeasible However, I don't think it follows from this that extracting rich knowledge from the Web is infeasible It would require a complex system involving at least 1) An NLP engine that maps each sentence into a menu of probabilistically weighted logical interpretations of the sentence (including links into other sentences built using anaphor resolution heuristics). This involves a dozen conceptually distinct components and is not at all trivial to design, build or tune. 2) Use of probabilistic inference rules to create implication links between the different interpretations of the different sentences 3) Use of an optimization algorithm (which could be a clever use of SAT or SMT, or something else) to utilize the links formed in step 2, to select the right interpretation(s) for each sentence Gosh, I think you've missed something of critical importance... The problem you stated above is about choosing the correct interpretation of a bunch of sentences. The problem we should tackle instead, is learning the rules that make up the KB. To see the difference, let's consider this example: Suppose I solve a problem (eg a programming exercise), and to illustrate my train of thoughts I clearly write down all the steps. So I have, in English, a bunch of sentences A,B,C,...,Z where Z is the final conclusion sentence. Now the AGI can translate sentences A-Z into logical form. You claim that this problem is hard because of multiple interpretations. But I think that's relatively unimportant compared to the real problem we face. So let's assume that we successfully -- correctly -- translate the NL sentences into logic. Now let's imagine that the AGI is doing the exercise, not me. Then it should have a train of inference that goes from A to B to C ... and so on... to Z. But, the AGI would NOT be able to make such a train of thoughts. All it has is just a bunch of *static* sentences from A-Z. What is missing? What would allow the AGI to actually conduct the inference from A-Z? The missing ingredient is a bunch of rules. These are the invisible glue that links the thoughts between the lines. This is the knowledge that I think should be learned, and would be very difficult to learn. You know what I'm talking about?? YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
YKY I thought you were talking about the extraction of information that is explicitly stated in online text. Of course, inference is a separate process (though it may also play a role in direct information extraction). I don't think the rules of inference per se need to be learned. In our book on PLN we outline a complete set of probabilistic logic inference rules, for example. What needs to be learned via experience is how to appropriately bias inference control -- how to sensibly prune the inference tree. So, one needs an inference engine that can adaptively learn better and better inference control as it carries out inferences. We designed and partially implemented this feature in the NCE but never completed the work due to other priorities ... but I hope this can get done in NM or OpenCog sometime in late 2008.. -- Ben On Tue, Feb 26, 2008 at 3:02 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 2/26/08, Ben Goertzel [EMAIL PROTECTED] wrote: Obviously, extracting knowledge from the Web using a simplistic SAT approach is infeasible However, I don't think it follows from this that extracting rich knowledge from the Web is infeasible It would require a complex system involving at least 1) An NLP engine that maps each sentence into a menu of probabilistically weighted logical interpretations of the sentence (including links into other sentences built using anaphor resolution heuristics). This involves a dozen conceptually distinct components and is not at all trivial to design, build or tune. 2) Use of probabilistic inference rules to create implication links between the different interpretations of the different sentences 3) Use of an optimization algorithm (which could be a clever use of SAT or SMT, or something else) to utilize the links formed in step 2, to select the right interpretation(s) for each sentence Gosh, I think you've missed something of critical importance... The problem you stated above is about choosing the correct interpretation of a bunch of sentences. The problem we should tackle instead, is learning the rules that make up the KB. To see the difference, let's consider this example: Suppose I solve a problem (eg a programming exercise), and to illustrate my train of thoughts I clearly write down all the steps. So I have, in English, a bunch of sentences A,B,C,...,Z where Z is the final conclusion sentence. Now the AGI can translate sentences A-Z into logical form. You claim that this problem is hard because of multiple interpretations. But I think that's relatively unimportant compared to the real problem we face. So let's assume that we successfully -- correctly -- translate the NL sentences into logic. Now let's imagine that the AGI is doing the exercise, not me. Then it should have a train of inference that goes from A to B to C ... and so on... to Z. But, the AGI would NOT be able to make such a train of thoughts. All it has is just a bunch of *static* sentences from A-Z. What is missing? What would allow the AGI to actually conduct the inference from A-Z? The missing ingredient is a bunch of rules. These are the invisible glue that links the thoughts between the lines. This is the knowledge that I think should be learned, and would be very difficult to learn. You know what I'm talking about?? YKY agi | Archives | Modify Your Subscription -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
RE: [agi] would anyone want to use a commonsense KB?
Ben, Thanks for the info. If you knew of anything relatively simple that was on-line that would be preferred. But, if not, I guess I could try to Google for something myself. (Of if I wait long enough there probably will be a simple Wikipedia explanation.) Ed Porter -Original Message- From: Ben Goertzel [mailto:[EMAIL PROTECTED] Sent: Sunday, February 24, 2008 9:13 PM To: agi@v2.listbox.com Subject: Re: [agi] would anyone want to use a commonsense KB? Hi, There is no good overview of SMT so far as I know, just some technical papers... but SAT solvers are not that deep and are well reviewed in this book... http://www.sls-book.net/ -- Ben On Sun, Feb 24, 2008 at 4:38 PM, Ed Porter [EMAIL PROTECTED] wrote: Ben or anyone, Do you know of an explanation or reference that is a for Dummies explanation of how SAT (or SMT) handles computations in spaces with and 100,000 variables and/or 10^300 states in practically computable time. I assume it is by focusing only on that part of the space through which relevant and/or relatively short inferences paths pass, or something like that. Ed Porter -Original Message- From: Ben Goertzel [mailto:[EMAIL PROTECTED] Sent: Wednesday, February 20, 2008 5:54 PM To: agi@v2.listbox.com Subject: Re: [agi] would anyone want to use a commonsense KB? And I seriously doubt that a general SMT solver + prob. theory is going to beat a custom probabilistic logic solver. My feeling is that an SMT solver plus appropriate subsets of prob theory can be a very powerful component of a general probabilistic inference framework... I can back this up with some details but that would get too thorny for this list... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
YKY (Yan King Yin) [EMAIL PROTECTED] wrote: Comparing the problem at hand with SAT may not be very accurate. First, we need to formulate the problem more clearly -- what exactly are we trying to do. Then we can estimate whether it's feasible with available computing power. Also, modern complexity theory has moved beyond P and NP -- to things that I'm still struggling to learn. For example, there are complexity classes beyond NP, in the polynomial hierarchy. Also there is the analytical hierarchy which is different from the polynomial one. Very often I see logical problems being described in the analytical hierarchy. For example, abduction and induction are both higher in the analytical hierarchy than SAT. I guess our problem would involve abductive and inductive learning, so it would be strictly harder than SAT. No doubt that we'd employ heuristics, so the worst-case complexity is not a show-stopper. But still, there is the possibility that the problem would be too hard using realistic computing power. YKY AGI might even turn out to be impossible, but I feel that there is a greater chance that eventually a program that is built to handle more complexity, in just the right way, will succeed. I had pretty much rejected the possibility that more logical or rational systems would result in a significant leap in the field until I became interested in the feasibility of a general polytime SAT solver. Now I see it as a logical next step in the trial and error method of research. In other words, something has been missing in the field logical-rational methods. The lack of a more efficient general solver actually represents an unbalance in the use of logical methods in computing. Most of us did not think of in just this way, but if a reasonable polytime general solver is feasible then it means that that we can significantly boost computing power through software. Even if this doesn't produce a significant leap in AI it might produce the overdue next step. Jim Bromer - Looking for last minute shopping deals? Find them fast with Yahoo! Search. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
RE: [agi] would anyone want to use a commonsense KB?
Ben or anyone, Do you know of an explanation or reference that is a for Dummies explanation of how SAT (or SMT) handles computations in spaces with and 100,000 variables and/or 10^300 states in practically computable time. I assume it is by focusing only on that part of the space through which relevant and/or relatively short inferences paths pass, or something like that. Ed Porter -Original Message- From: Ben Goertzel [mailto:[EMAIL PROTECTED] Sent: Wednesday, February 20, 2008 5:54 PM To: agi@v2.listbox.com Subject: Re: [agi] would anyone want to use a commonsense KB? And I seriously doubt that a general SMT solver + prob. theory is going to beat a custom probabilistic logic solver. My feeling is that an SMT solver plus appropriate subsets of prob theory can be a very powerful component of a general probabilistic inference framework... I can back this up with some details but that would get too thorny for this list... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Pp Sent from my phone. On Feb 24, 2008, at 1:38 PM, Ed Porter [EMAIL PROTECTED] wrote: Ben or anyone, Do you know of an explanation or reference that is a for Dummies explanation of how SAT (or SMT) handles computations in spaces with and 100,000 variables and/or 10^300 states in practically computable time. I assume it is by focusing only on that part of the space through which relevant and/or relatively short inferences paths pass, or something like that. Ed Porter -Original Message- From: Ben Goertzel [mailto:[EMAIL PROTECTED] Sent: Wednesday, February 20, 2008 5:54 PM To: agi@v2.listbox.com Subject: Re: [agi] would anyone want to use a commonsense KB? And I seriously doubt that a general SMT solver + prob. theory is going to beat a custom probabilistic logic solver. My feeling is that an SMT solver plus appropriate subsets of prob theory can be a very powerful component of a general probabilistic inference framework... I can back this up with some details but that would get too thorny for this list... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Hi, There is no good overview of SMT so far as I know, just some technical papers... but SAT solvers are not that deep and are well reviewed in this book... http://www.sls-book.net/ -- Ben On Sun, Feb 24, 2008 at 4:38 PM, Ed Porter [EMAIL PROTECTED] wrote: Ben or anyone, Do you know of an explanation or reference that is a for Dummies explanation of how SAT (or SMT) handles computations in spaces with and 100,000 variables and/or 10^300 states in practically computable time. I assume it is by focusing only on that part of the space through which relevant and/or relatively short inferences paths pass, or something like that. Ed Porter -Original Message- From: Ben Goertzel [mailto:[EMAIL PROTECTED] Sent: Wednesday, February 20, 2008 5:54 PM To: agi@v2.listbox.com Subject: Re: [agi] would anyone want to use a commonsense KB? And I seriously doubt that a general SMT solver + prob. theory is going to beat a custom probabilistic logic solver. My feeling is that an SMT solver plus appropriate subsets of prob theory can be a very powerful component of a general probabilistic inference framework... I can back this up with some details but that would get too thorny for this list... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 20/02/2008, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: So, looking at the moon, what color would you say it was? As Edwin Land showed colour perception does not just depend upon the wavelength of light, but is a subjective property actively constructed by the brain. http://en.wikipedia.org/wiki/Color_constancy http://youtube.com/watch?v=ZiTg4kRt13w --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/19/08, Ben Goertzel [EMAIL PROTECTED] wrote: If we need a KB orders of magnitude larger to make that approach work, doesn't that mean we should use another approach? But do you agree that a KB orders of magnitude larger is required for all AGI, regardless of *how* the knowledge is acquired? The debate is on how to acquire the knowledge. Like, er, embodied learning or NL information extraction / conversation ... which have the potential to allow rules to be learned implicitly from experience rather than explicitly via human hard-coding... Let me list all the ways of AGI knowledge acquisition: A) manual encoding in logical form B) manual teaching in NL and pictures C) learning in virtual reality (eg Second Life) D) embodied learning (eg computer vision) E) inductive learning / extraction from existing texts I'm originally proposing A, but I'm also considering B, as Bob suggests. C is not very viable as of now. The physics in Second Life is simply not *rich* enough. SL is mainly a space for humans to socialize, so the physics will not get much richer in the near future -- is anyone interested in emulating cigarette smoke in SL? D is hard. I think I know how to do it, but it'd require $$$. E is also hard, but you seem to be *unaware* of its difficulty. In fact, the problem with E is the same as that with AIXI -- the thoery is elegant, but the actual learning would take forever. Can you explain, in broad terms, how the AGI is to know that water runs downhill instead of up, and that the moon is not blue, but a greyish color? It just doesn't seem a pragmatically feasible approach, setting aside all my doubts about the AI viability of it (i.e., I'm not so sure that even if you spent a billion dollars on hand-coding of rules, this would be all that helpful for AGI, in the absence of a learning engine radically different in nature from typical logical reasoning engines...) The KB can be used in conjunction with any learning algorithm you have. In fact, all A-E can be used together to build the KB. You seem to think that method E is so superior that you don't need other sources of knowledge, but the problem is the efficiency of the learning algorithm. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 20/02/2008, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: E is also hard, but you seem to be *unaware* of its difficulty. In fact, the problem with E is the same as that with AIXI -- the thoery is elegant, but the actual learning would take forever. Can you explain, in broad terms, how the AGI is to know that water runs downhill instead of up, and that the moon is not blue, but a greyish color? You might be able to extract the knowledge that the moon is grey from text analysis, but ultimately this knowledge comes from D. Also, there is nothing intrinsically grey about the moon. It's only grey in the context of a certain kind of physical system observing it. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
C is not very viable as of now. The physics in Second Life is simply not *rich* enough. SL is mainly a space for humans to socialize, so the physics will not get much richer in the near future -- is anyone interested in emulating cigarette smoke in SL? Second Life will soon be integrating the Havok 4 physics engine. I agree that game-world physics is not yet very realistic, but it's improving fast, due to strong economics in the MMOG industry. E is also hard, but you seem to be *unaware* of its difficulty. In fact, the problem with E is the same as that with AIXI -- the thoery is elegant, but the actual learning would take forever. Can you explain, in broad terms, how the AGI is to know that water runs downhill instead of up, and that the moon is not blue, but a greyish color? Water does not always run downhill, sometimes it runs uphill. To learn commonsense information from text requires parsing the text and mapping the parse-trees into semantic relationships, which are then reasoned on by a logical reasoning engine. There is nothing easy about this, and there is a hard problem of semantic disambiguation of relationships. Whether the disambiguation problem can be solved via statistical/inferential integration of masses of extracted relationships, remains to be seen. Virtual embodiment coupled with NL conversation is the approach I currently favor, but I think that large-scale NL information extraction can also play an important helper role. And I think that as robotics tech develops, it can play a big role too. I think we can take all approaches at once within an integrative framework like Novamente or OpenCog, but if I have to pick a single focus it will be virtual embodiment, with the other aspects as helpers... -- Ben G --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Water does not always run downhill, sometimes it runs uphill. But never without a reason. - Original Message - From: Ben Goertzel [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, February 20, 2008 9:47 AM Subject: Re: [agi] would anyone want to use a commonsense KB? C is not very viable as of now. The physics in Second Life is simply not *rich* enough. SL is mainly a space for humans to socialize, so the physics will not get much richer in the near future -- is anyone interested in emulating cigarette smoke in SL? Second Life will soon be integrating the Havok 4 physics engine. I agree that game-world physics is not yet very realistic, but it's improving fast, due to strong economics in the MMOG industry. E is also hard, but you seem to be *unaware* of its difficulty. In fact, the problem with E is the same as that with AIXI -- the thoery is elegant, but the actual learning would take forever. Can you explain, in broad terms, how the AGI is to know that water runs downhill instead of up, and that the moon is not blue, but a greyish color? Water does not always run downhill, sometimes it runs uphill. To learn commonsense information from text requires parsing the text and mapping the parse-trees into semantic relationships, which are then reasoned on by a logical reasoning engine. There is nothing easy about this, and there is a hard problem of semantic disambiguation of relationships. Whether the disambiguation problem can be solved via statistical/inferential integration of masses of extracted relationships, remains to be seen. Virtual embodiment coupled with NL conversation is the approach I currently favor, but I think that large-scale NL information extraction can also play an important helper role. And I think that as robotics tech develops, it can play a big role too. I think we can take all approaches at once within an integrative framework like Novamente or OpenCog, but if I have to pick a single focus it will be virtual embodiment, with the other aspects as helpers... -- Ben G --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
--- YKY (Yan King Yin) [EMAIL PROTECTED] wrote: Let me list all the ways of AGI knowledge acquisition: A) manual encoding in logical form B) manual teaching in NL and pictures C) learning in virtual reality (eg Second Life) D) embodied learning (eg computer vision) E) inductive learning / extraction from existing texts In the distributed query/message posting service I have proposed in http://www.mattmahoney.net/agi.html all of these methods could be used. The system creates a marketplace that rewards intelligence and cooperation with computing resources -- storage and bandwidth. Personally, I believe the system will favor extracting information from existing data on the internet and from interacting with humans in natural language, plus some manual encoding of knowledge. The requirement that peers communicate in natural language is not onerous. It is not necessary for each peer to fully solve the problem. Your calculator uses natural language in that it uses and understands symbols like 3 and +. Likewise, experts in 1959 baseball statistics (the latest year available at the time) have been written to understand queries like how many games did the Yankees win in July? [1]. Nor is it hard for each peer to route messages to the right experts. The system would work (albeit very inefficiently) if peers simply broadcast messages to every other peer that it knows about. A better, but still simple strategy would be to match terms to previously stored messages and forward to the peers that sent them. Peers will reward peers by prioritizing the messages of those that broadcast more selectively, thus increasing their own reputations. There will be an economic pressure to increase intelligence with better language models, i.e. a thesaurus or parsing to improve message understanding. Although I described the protocol using text, it could be extended to speech and images. For example, a peer would be rewarded if it could match a picture of a baseball player to an expert on baseball. I guess the question YKY would have is, how can I make money from this? Well, nobody would have control over it, but it does present opportunities for profit in a market that doesn't yet exist. I think it is worthwhile to investigate it and at least have a head start. References 1. Bert F. Green Jr., Alice K. Wolf, Carol Chomsky, and Kenneth Laughery, Baseball: An Automatic Question Answerer, Proceedings of the Western Joint Computer Conference, 19:219-224, 1961. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 20, 2008 1:34 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: Looking at the moon won't help -- of course it helps, it tells you that something odd is with the expression, as opposed to say yellow sun ... it might be the case that it described a particular appearance that only had a slight resemblance to other blue things (as in red hair), for example. There are some rare conditions (high stratospheric dust) which can make the moon look actually blue. In fact blue moon is generally taken to mean, metaphorically, something very rare (or even impossible) or the second full moon in a given month (which happens about every two-and-a-half years on the average). ask someone is of course what human kids do a lot of. An AI could do this, or look it up in Wikipedia, or the like. All of which are heuristics to reduce the ambiguity/generality in the information stream. The question is do enough heuristics make an autogenous AI or is there something more fundamental to its structure? On Wednesday 20 February 2008 12:27:59 pm, Ben Goertzel wrote: The trick to understanding once in a blue moon is to either -- look at the moon or -- ask someone --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Looking at the moon won't help -- it might be the case that it described a particular appearance that only had a slight resemblance to other blue things (as in red hair), for example. There are some rare conditions (high stratospheric dust) which can make the moon look actually blue. In fact blue moon is generally taken to mean, metaphorically, something very rare (or even impossible) or the second full moon in a given month (which happens about every two-and-a-half years on the average). ask someone is of course what human kids do a lot of. An AI could do this, or look it up in Wikipedia, or the like. All of which are heuristics to reduce the ambiguity/generality in the information stream. The question is do enough heuristics make an autogenous AI or is there something more fundamental to its structure? On Wednesday 20 February 2008 12:27:59 pm, Ben Goertzel wrote: The trick to understanding once in a blue moon is to either -- look at the moon or -- ask someone --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
So, looking at the moon, what color would you say it was? Here's what text mining might give you (Google hits): blue moon 11,500,000 red moon 1,670,000 silver moon 1,320,000 yellow moon 712,000 white moon 254,000 golden moon 163,000 orange moon 122,000 green moon 105,000 gray moon 9,460 To me, the moon varies from a deep orange to brilliant white depending on atmospheric conditions and time of night... none of which would help me understand the text references. On Wednesday 20 February 2008 02:02:52 pm, Ben Goertzel wrote: On Feb 20, 2008 1:34 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: Looking at the moon won't help -- of course it helps, it tells you that something odd is with the expression, as opposed to say yellow sun ... --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
There seems to be an assumption in this thread that NLP analysis of text is restricted to simple statistical extraction of word-sequences... This is not the case... If there were to be a hope for AGI based on text analysis, it would have to be based on systems that parse linguistic expressions into logical relationships, and combine these logical relationships via reasoning. Assessing metaphoric versus literal mentions would be part of that reasoning. Critiquing NLP-based AGI based on Google is a lot like critiquing robotics- based AGI based on the Roomba. Google is a good product implemented very scalably, but in its linguistic sophistication, it is nowhere near the best research systems out there. Let alone what would be possible with further research. I stress that this is not my favored approach to AGI, but I think these discussions based on Google are unfairly dismissing NLP-based AGI by using Google as a straw man. I note also that a web-surfing AGI could resolve the color of the moon quite easily by analyzing online pictures -- though this isn't pure text mining, it's in the same spirit... ben On Feb 20, 2008 2:30 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: So, looking at the moon, what color would you say it was? Here's what text mining might give you (Google hits): blue moon 11,500,000 red moon 1,670,000 silver moon 1,320,000 yellow moon 712,000 white moon 254,000 golden moon 163,000 orange moon 122,000 green moon 105,000 gray moon 9,460 To me, the moon varies from a deep orange to brilliant white depending on atmospheric conditions and time of night... none of which would help me understand the text references. On Wednesday 20 February 2008 02:02:52 pm, Ben Goertzel wrote: On Feb 20, 2008 1:34 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: Looking at the moon won't help -- of course it helps, it tells you that something odd is with the expression, as opposed to say yellow sun ... --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Wednesday 20 February 2008 02:58:54 pm, Ben Goertzel wrote: I note also that a web-surfing AGI could resolve the color of the moon quite easily by analyzing online pictures -- though this isn't pure text mining, it's in the same spirit... U -- I just typed moon into google and at the top of the page it gives three pictures. Two are thin sliver crescents. The third, of a full moon, is distinctly blue. There seems to be an assumption in this thread that NLP analysis of text is restricted to simple statistical extraction of word-sequences... I certainly make no such assumption. I offered the stats to point out the kind of traps that lie in wait for the hapless text-miner. As I am sure you are fully aware, you can't parse English without a knowledge of the meanings involved. (The council opposed the demonstrators because they (feared/advocated) violence.) So how are you going to learn meanings before you can parse, or how are you going to parse before you learn meanings? They have to be interleaved in a non-trivial way. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
As I am sure you are fully aware, you can't parse English without a knowledge of the meanings involved. (The council opposed the demonstrators because they (feared/advocated) violence.) So how are you going to learn meanings before you can parse, or how are you going to parse before you learn meanings? They have to be interleaved in a non-trivial way. True indeed! Feeding all the ambiguous interpretations of a load of sentences into a probabilistic logic network, and letting them get resolved by reference to each other, is a sort of search for the most likely solution of a huge system of simultaneous equations ... i.e. one needs to let each, of a huge set of ambiguities, be resolved by the other ones... This is not an easy problem, but it's not on the face of it unsolvable... But I think the solution will be easier with info from direct experience to nudge the process in the right direction... Ben -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
--- Ben Goertzel [EMAIL PROTECTED] wrote: As I am sure you are fully aware, you can't parse English without a knowledge of the meanings involved. (The council opposed the demonstrators because they (feared/advocated) violence.) So how are you going to learn meanings before you can parse, or how are you going to parse before you learn meanings? They have to be interleaved in a non-trivial way. True indeed! Feeding all the ambiguous interpretations of a load of sentences into a probabilistic logic network, and letting them get resolved by reference to each other, is a sort of search for the most likely solution of a huge system of simultaneous equations ... i.e. one needs to let each, of a huge set of ambiguities, be resolved by the other ones... This is not an easy problem, but it's not on the face of it unsolvable... But I think the solution will be easier with info from direct experience to nudge the process in the right direction... Children solve the problem by learning semantics before grammar. Statistical language models do the same thing. Models like LSA and vector spaces (used for search) do not depend on word order. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Yes, of course, but no human except an expert in lunar astronomy would have a definitive answer to the question either The issue at hand is really how a text-analysis based AGI would distinguish literal from metaphoric text, and how it would understand the context in which a statement is implicitly intended by the speaker/writinger. These are hard problems, which are being worked on by many individuals in the computational linguistics community. I tend to think that introducing (real or virtual) embodiment will make the solution of these problems easier... -- Ben On Wed, Feb 20, 2008 at 3:45 PM, Matt Mahoney [EMAIL PROTECTED] wrote: --- Ben Goertzel [EMAIL PROTECTED] wrote: I note also that a web-surfing AGI could resolve the color of the moon quite easily by analyzing online pictures -- though this isn't pure text mining, it's in the same spirit... Not really. You can get a better answer to what color is the moon? if you google what color is the moon?. Better, but not definitive. Even the photos are not in agreement. Some photos show a mix of orange and blue. If you stood on the moon, it would look black next to your feet, but white in contrast to the even darker sky. During tonight's eclipse, it should look reddish brown. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
--- Ben Goertzel [EMAIL PROTECTED] wrote: I note also that a web-surfing AGI could resolve the color of the moon quite easily by analyzing online pictures -- though this isn't pure text mining, it's in the same spirit... Not really. You can get a better answer to what color is the moon? if you google what color is the moon?. Better, but not definitive. Even the photos are not in agreement. Some photos show a mix of orange and blue. If you stood on the moon, it would look black next to your feet, but white in contrast to the even darker sky. During tonight's eclipse, it should look reddish brown. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
OK, imagine a lifetime's experience is a billion symbol-occurences. Imagine you have a heuristic that takes the problem down from NP-complete (which it almost certainly is) to a linear system, so there is an N^3 algorithm for solving it. We're talking order 1e27 ops. Now using HEPP = 1e16 x 30 years = 1e9 secs, you get a total crunch for the human of 1e25 ops. That's close enough to call even, I think. Learning order is easily worth a couple orders of magnitude in problem complexity. Let's build a big cluster... On Wednesday 20 February 2008 03:51:28 pm, Ben Goertzel wrote: Feeding all the ambiguous interpretations of a load of sentences into a probabilistic logic network, and letting them get resolved by reference to each other, is a sort of search for the most likely solution of a huge system of simultaneous equations ... i.e. one needs to let each, of a huge set of ambiguities, be resolved by the other ones... This is not an easy problem, but it's not on the face of it unsolvable... But I think the solution will be easier with info from direct experience to nudge the process in the right direction... Ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Wed, Feb 20, 2008 at 4:27 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: OK, imagine a lifetime's experience is a billion symbol-occurences. Imagine you have a heuristic that takes the problem down from NP-complete (which it almost certainly is) to a linear system, so there is an N^3 algorithm for solving it. We're talking order 1e27 ops. That's kind of specious, since modern SAT and SMT solvers can solve many realistic instances of NP-complete problems for large n, surprisingly quickly... and without linearizing anything... Worst-case complexity doesn't mean much... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/21/08, Ben Goertzel [EMAIL PROTECTED] wrote: Feeding all the ambiguous interpretations of a load of sentences into a probabilistic logic network, and letting them get resolved by reference to each other, is a sort of search for the most likely solution of a huge system of simultaneous equations ... i.e. one needs to let each, of a huge set of ambiguities, be resolved by the other ones... This is not an easy problem, but it's not on the face of it unsolvable... But I think the solution will be easier with info from direct experience to nudge the process in the right direction... Excellent analogy, I'd like to work on such a system =) Note that it's not just ambiguities that need to be resolved; some facts need to be explained away, for example, water may uphill because of some abnormal conditions. The system needs to be able to find explanations for things... YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
A PROBABILISTIC logic network is a lot more like a numerical problem than a SAT problem. On Wednesday 20 February 2008 04:41:51 pm, Ben Goertzel wrote: On Wed, Feb 20, 2008 at 4:27 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: OK, imagine a lifetime's experience is a billion symbol-occurences. Imagine you have a heuristic that takes the problem down from NP-complete (which it almost certainly is) to a linear system, so there is an N^3 algorithm for solving it. We're talking order 1e27 ops. That's kind of specious, since modern SAT and SMT solvers can solve many realistic instances of NP-complete problems for large n, surprisingly quickly... and without linearizing anything... Worst-case complexity doesn't mean much... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Not necessarily, because --- one can encode a subset of the rules of probability as a theory in SMT, and use an SMT solver -- one can use probabilities to guide the search within an SAT or SMT solver... ben On Wed, Feb 20, 2008 at 5:00 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: A PROBABILISTIC logic network is a lot more like a numerical problem than a SAT problem. On Wednesday 20 February 2008 04:41:51 pm, Ben Goertzel wrote: On Wed, Feb 20, 2008 at 4:27 PM, J Storrs Hall, PhD [EMAIL PROTECTED] wrote: OK, imagine a lifetime's experience is a billion symbol-occurences. Imagine you have a heuristic that takes the problem down from NP-complete (which it almost certainly is) to a linear system, so there is an N^3 algorithm for solving it. We're talking order 1e27 ops. That's kind of specious, since modern SAT and SMT solvers can solve many realistic instances of NP-complete problems for large n, surprisingly quickly... and without linearizing anything... Worst-case complexity doesn't mean much... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Ben Goertzel [EMAIL PROTECTED] wrote: On Wed, Feb 20, 2008 at 4:27 PM, J Storrs Hall, PhD wrote: OK, imagine a lifetime's experience is a billion symbol-occurences. Imagine you have a heuristic that takes the problem down from NP-complete (which it almost certainly is) to a linear system, so there is an N^3 algorithm for solving it. We're talking order 1e27 ops. That's kind of specious, since modern SAT and SMT solvers can solve many realistic instances of NP-complete problems for large n, surprisingly quickly... and without linearizing anything... Worst-case complexity doesn't mean much... ben I think you may both be off the track here. Although many difficult problems for large n can be solved with contemporary solvers (and approximations to those that cannot be can be used in their stead) many of the kinds of problems that need to be solved cannot be approximated. I don't think that J Storrs Hall example of a lifetime's experience (a billion symbols) is truly relevant. No one actually believes that the human mind is capable of analyzing reality into a monotonic (pure hierarchal) logic and there is not much reason to believe that an effective AGI would have to be. But there is a lot of reason to believe that increasing the capacity of logic solvers dramatically would be useful. To get back to Ben's statement: Is the computer chip industry happy with contemporary SAT solvers or would a general solver that is capable of beating n^4 time be of some use to them? If it would be useful, then there is a reason to believe that it might be useful to AGI. Jim Bromer - Never miss a thing. Make Yahoo your homepage. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
To get back to Ben's statement: Is the computer chip industry happy with contemporary SAT solvers Well they are using them, but of course there is loads of room for improvement!! or would a general solver that is capable of beating n^4 time be of some use to them? If it would be useful, then there is a reason to believe that it might be useful to AGI. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
And I seriously doubt that a general SMT solver + prob. theory is going to beat a custom probabilistic logic solver. My feeling is that an SMT solver plus appropriate subsets of prob theory can be a very powerful component of a general probabilistic inference framework... I can back this up with some details but that would get too thorny for this list... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
It's probably not worth too much taking this a lot further, since we're talking in analogies and metaphors. However, it's my intuition that the connectivity in a probabilistic formulation is going to produce a much denser graph (less sparse matrix) than what you find in the SAT problems that the solvers do so well on. And I seriously doubt that a general SMT solver + prob. theory is going to beat a custom probabilistic logic solver. On Wednesday 20 February 2008 05:31:59 pm, Ben Goertzel wrote: Not necessarily, because --- one can encode a subset of the rules of probability as a theory in SMT, and use an SMT solver -- one can use probabilities to guide the search within an SAT or SMT solver... ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/19/08, Stephen Reed [EMAIL PROTECTED] wrote: Pei: Resolution-based FOL on a huge KB is intractable. Agreed. However Cycorp spend a great deal of programming effort (i.e. many man-years) finding deep inference paths for common queries. The strategies were: prune the rule set according to the context substitute procedural code for modus ponens in common query paths (e.g. isa-links inferred via graph traversal) structure the inference engine as a nested set of iterators so that easy answers are returned immediately, and harder-to-find answers trickle out later. establish a battery of inference engine controls (e.g. time bounds, speed vs. completeness - whether to employ expensive inference strategies for greater coverage of answers) and have the inference engine automatically apply the optimal control configuration for queries determine rule utility via machine learning and apply prioritized inference modules within the given time constraints My last in-house talk at Cycorp, in the summer of 2006, described a notion of mine that Cyc's deductive inference engine behaves as an interpreter, and that for a certain set of queries, a dramatic speed improvement (e.g. four orders of magnitude) could be achieved by compiling the query, and possibly preprocessing incoming facts to suit expected queries. The queries that interested me were those embedded in an intelligent application, and which could be viewed as a query template with parameters. The compilation process I described would explore the parameter space with programmer-chosen query examples. Then the resulting proof trees would be compiled into executable code - avoiding entirely the time consuming candidate rule search and their application when the query executes. My notion for Cyc's deductive inference engine optimization is analogous to SQL query optimization technology. I expect to use this technique in the Texai project at the point when I need a deductive inference engine. Thanks a lot for the info. These are very important speed-up strategies. I have not yet studied this aspect in detail. Can you explain what you mean by deep inference? I think resolution theorem proving provides a way to answer yes/no queries in a KB. I take it as a starting point, and try to think of ways to speed it up and to expand its abilities (answering what/where/when/who/how queries). YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/19/08, Matt Mahoney [EMAIL PROTECTED] wrote: Why would this approach succeed where Cyc failed? Cyc paid people to build the knowledge base. Then when they couldn't sell it, the tried giving it away. Still, nobody used it. For an AGI to be useful, people have to be able to communicate with it in natural language. It is easy to manipulate formulas like if P then Q. It is much harder to explain how this knowledge is represented and learned in a language model. Cyc did not solve this problem, and we see the result. I think Cyc failed mainly because their KB is not large enough to make useful inferences. We need a huge KB indeed. Automatic rule generalization can make the KB smaller. Translating NL into the KB form is one way to collect facts/rules easier. But I still think the knowledge representation should be logic, instead of natural language. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/19/08, Pei Wang [EMAIL PROTECTED] wrote: A purely resolution-based inference engine is mathematically elegant, but completely impractical, because after all the knowledge are transformed into the clause form required by resolution, most of the semantic information in the knowledge structure is gone, and the result is equivalent to the original knowledge in truth-value only. It is hard to control the direction of the inference without semantic information. I wonder how you can preserve structural information in NARS? If I say eating sweets will cause cavities it will be translated in clause form as ~ eat_sweets V cavities so the direction of causation is lost. If this directional info is needed, we attach additional information to the clause. Truth maintenance systems do something like that. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 19, 2008 8:49 AM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 2/19/08, Pei Wang [EMAIL PROTECTED] wrote: A purely resolution-based inference engine is mathematically elegant, but completely impractical, because after all the knowledge are transformed into the clause form required by resolution, most of the semantic information in the knowledge structure is gone, and the result is equivalent to the original knowledge in truth-value only. It is hard to control the direction of the inference without semantic information. I wonder how you can preserve structural information in NARS? By supporting various compound terms/statements. If I say eating sweets will cause cavities it will be translated in clause form as ~ eat_sweets V cavities so the direction of causation is lost. If this directional info is needed, we attach additional information to the clause. How? Truth maintenance systems do something like that. No. That is for something else --- update management. BTW, classical resolution can be used to answer wh questions using unification. Pei YKY agi | Archives | Modify Your Subscription --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 19, 2008 2:41 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: I think resolution theorem proving provides a way to answer yes/no queries in a KB. I take it as a starting point, and try to think of ways to speed it up and to expand its abilities (answering what/where/when/who/how queries). Oh my, resolution answers wh-questions as well as decision questions in FOL. You just record the answer substitution. (BTW, Prolog is a positive resolution.) We need to be more technical here. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
According to the in-house Cycorp jargon, deep inference begins at approximately four backchain steps in a deductive inference. As most here know, there is an exponential fanout in the number of separate inference paths with each backchain step, given a large candidate rule set and a large set of facts. One or two backchain steps can usually be accomplished in seconds by Cyc. Deep inference, let's say six backchain steps, may require hours if it completes at all. Cyc is designed to make the best use of the time allocated for answering a query, and the iterative deepening stategy was proposed as a query configuration alternative. It is possible that Cycorp uses it now - I no longer have access to its source code. Although I agree that a resolution based refutation proof may be more efficient for answering a yes/no query, I think, as did Cycorp, that effort is better spent on a single inference engine that operates in a constructive manner, like an ordinary SQL query engine, but with heuristic rule application included. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: YKY (Yan King Yin) [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Tuesday, February 19, 2008 7:41:06 AM Subject: Re: [agi] would anyone want to use a commonsense KB? [snip]Thanks a lot for the info. These are very important speed-up strategies. I have not yet studied this aspect in detail. Can you explain what you mean by deep inference? I think resolution theorem proving provides a way to answer yes/no queries in a KB. I take it as a starting point, and try to think of ways to speed it up and to expand its abilities (answering what/where/when/who/how queries). YKY agi | Archives | Modify Your Subscription Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try it now. http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 19/02/2008, Ben Goertzel [EMAIL PROTECTED] wrote: If we need a KB orders of magnitude larger to make that approach work, doesn't that mean we should use another approach? Yes. Like, er, embodied learning or NL information extraction / conversation ... which have the potential to allow rules to be learned implicitly from experience rather than explicitly via human hard-coding... A compromise scenario might be to go back to the very early days of AI and have the goal of an AGI learning to read books intended for young children containing pictures and relatively simple sentences. This would involve a combination of both automatic knowledge extraction and explicit teaching by a supervisor. The key to success would be establishing some system which allowed teacher and learner to interact in a manner which facilitates efficient and flexible knowledge transfer. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 9:42 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: So far I've been using resolution-based FOL, so there's only 1 inference rule and this is not a big issue. If you're using nonstandard inference rules, perhaps even approximate ones, I can see that this distinction is important. Resolution-based FOL on a huge KB is intractable. Pei --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
All of these rules have exception or implicit condition. If you treat them as default rules, you run into multiple extension problem, which has no domain-independent solution in binary logic --- read http://www.cogsci.indiana.edu/pub/wang.reference_classes.ps for details. Pei On Feb 17, 2008 10:04 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: Yesterday I didn't give a clear explanation of what I mean by rules, so here is a better try: 1. If I see a turkey inside the microwave, I immediately draw the conclusion that it's NOT empty. 2. However, if I see some katchup on the inside walls of the microwave, I'd say it's dirty but it's empty. 3. If I see the rotating plate inside the microwave, I'd still say it's empty 'cause the plate is part of the microwave. etc etc So the AGI may have a rule that sounds like: if X is an object inside the microwave, and X satisfies some criteria, then the microwave is NOT empty. But it would be a very dumb AGI if it has this rule specifically for microwave ovens, and then some other rules for washing machines, bottles, book shelves, and other containers. It would be necessary for the AGI to have a general rule for emptiness for all containers. So I'd say a washing machine with a sock inside is not empty, but if it's just some lint then it's empty. Such a general rule for emptiness is certainly not available on the net, at least not explicitly expressed. One solution is to manually encode them (perhaps with some machine assistance), which is the approach of Cyc. Another solution is to induce them from existing texts on the web -- Ben's suggestion. If given a large enough corpus and a long enough learning period, Ben's solution may work. The key issue is how to speed up the inductive learning of rules. YKY agi | Archives | Modify Your Subscription --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 18/02/2008, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: Well, the idea is to ask lots of people to contribute to the KB, and pay them with virtual credits. (I expect such people to have a little knowledge in logic or Prolog, so they can enter complex rules. Also, they can be assisted by inductive learning algorithms.) The income of the KB will be given back to them. I'll take a bit of administrative fees =) In principle this sounds ok, but this is almost exactly the same as the Mindpixel business model. Once an element of payment is involved (usually with some kind of shares in future profits) participants tend to expect that they're going to be able to realise that value within a relatively short time, like a few years. Inevitably when expectations aren't met things get sticky. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
All of these rules have exception or implicit condition. If you treat them as default rules, you run into multiple extension problem, which has no domain-independent solution in binary logic --- read http://www.cogsci.indiana.edu/pub/wang.reference_classes.ps for details. Pei, Do you have a PDF version? Thanks! --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Just put one at http://nars.wang.googlepages.com/wang.reference_classes.pdf On Feb 18, 2008 9:01 AM, Mark Waser [EMAIL PROTECTED] wrote: All of these rules have exception or implicit condition. If you treat them as default rules, you run into multiple extension problem, which has no domain-independent solution in binary logic --- read http://www.cogsci.indiana.edu/pub/wang.reference_classes.ps for details. Pei, Do you have a PDF version? Thanks! --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
I believe I offered the beginning of a v. useful way to conceive of this whole area in an earlier post. The key concept is inventory of the world. First of all, what is actually being talked about here is only a VERBAL/SYMBOLIC KB. One of the grand illusions of a literature culture is that words/symbols refer to everything. The reality is that we have a v. limited verbal inventory of the world. Words do not describe most parts of your body, for example, only certain key divisions. Check over your hand for a start and see how many bits you can name - minute bit by bit. When it comes to the movements of objects, our vocabulary is breathtakingly limited. In fact, our verbal/symbolic inventory of the world (as provided for by our existing cultural vocabulary - for all its millions of words) is, I suggest, only a tiny fraction of our COMMON SENSE KB/ inventory of the world - i.e. that knowledge we hold purely in sensory image form - and indeed in common-sense form (since as Tye points out, we never actually experience/operate one sense in isolation - even though we have the intellectual illusion that we do). When we learn to respect the extent of our true common sense knowledge of the world as distinct from our formal, verbal knowledge of the world, we will realise another major reason why CYC like projects are doomed. They have nothing to do with common sense. Of course they will never be able to work out, pace Minsky, whether you can whistle and eat at the same time, or whether you can push or pull an object with a string. This is true common sense knowledge. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?.. p.s.
I should add to the idea of our common sense knowledge inventory of the world - because my talk of objects and movements may make it all sound v. physical and external. That common sense inventory also includes a vast amount of non-verbal knowledge, paradoxically, about how we think and communicate with and understand others.The paradoxical part is that a lot of this will be common sense about we use words themselves. Hence it is that experts have immense difficulties describing how they think about problems. They don't have the words for how they use their words. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/18/08, Mike Tintner [EMAIL PROTECTED] wrote: I believe I offered the beginning of a v. useful way to conceive of this whole area in an earlier post. The key concept is inventory of the world. First of all, what is actually being talked about here is only a VERBAL/SYMBOLIC KB. One of the grand illusions of a literature culture is that words/symbols refer to everything. The reality is that we have a v. limited verbal inventory of the world. Words do not describe most parts of your body, for example, only certain key divisions. Check over your hand for a start and see how many bits you can name - minute bit by bit. When it comes to the movements of objects, our vocabulary is breathtakingly limited. In fact, our verbal/symbolic inventory of the world (as provided for by our existing cultural vocabulary - for all its millions of words) is, I suggest, only a tiny fraction of our COMMON SENSE KB/ inventory of the world - i.e. that knowledge we hold purely in sensory image form - and indeed in common-sense form (since as Tye points out, we never actually experience/operate one sense in isolation - even though we have the intellectual illusion that we do). When we learn to respect the extent of our true common sense knowledge of the world as distinct from our formal, verbal knowledge of the world, we will realise another major reason why CYC like projects are doomed. They have nothing to do with common sense. Of course they will never be able to work out, pace Minsky, whether you can whistle and eat at the same time, or whether you can push or pull an object with a string. This is true common sense knowledge. I can give labels to every tiny sub-section of my hand, thus increasing the resolution of the symbolic description. If I give labels to each very small visual features of my hand, then the distinction between visual representation and symbolic representation disappears. Therefore, I think symbolic KBs like Cyc's is not doomed -- the symbolic KB can merge with perceptual grounding in a continuum fashion. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
This raises another v. interesting dimension of KB's and why they are limited. The social dimension. You might, purely for argument's sake, be able to name a vast amount of unnamed parts of the world. But you would then have to secure social agreement for them to become practically useful. Not realistic - if you were say to add even scores let alone thousands of names for each bit of your hand. And even when there are a set of agreed words - and this is a problem that absolutely plagues all of us on this board - there may still not be an agreed terminology. For example, we are having massive problems as a community, along with our society, with what words like intelligence, AGI, symbol, image image schema, etc etc. mean. We may agree broadly on the words that are relevant in a given area, but we have no agreed terminology as to which of those words should be used when, and what they mean.. And actually, now that I think of it, the more carefully intellectuals define their words, the MORE disagreements and misunderstandings you often get. Words like free and determined for philosophers and scientists (and all of us here) are absolute minefields. MT: I believe I offered the beginning of a v. useful way to conceive of this whole area in an earlier post. The key concept is inventory of the world. First of all, what is actually being talked about here is only a VERBAL/SYMBOLIC KB. One of the grand illusions of a literature culture is that words/symbols refer to everything. The reality is that we have a v. limited verbal inventory of the world. Words do not describe most parts of your body, for example, only certain key divisions. Check over your hand for a start and see how many bits you can name - minute bit by bit. When it comes to the movements of objects, our vocabulary is breathtakingly limited. In fact, our verbal/symbolic inventory of the world (as provided for by our existing cultural vocabulary - for all its millions of words) is, I suggest, only a tiny fraction of our COMMON SENSE KB/ inventory of the world - i.e. that knowledge we hold purely in sensory image form - and indeed in common-sense form (since as Tye points out, we never actually experience/operate one sense in isolation - even though we have the intellectual illusion that we do). When we learn to respect the extent of our true common sense knowledge of the world as distinct from our formal, verbal knowledge of the world, we will realise another major reason why CYC like projects are doomed. They have nothing to do with common sense. Of course they will never be able to work out, pace Minsky, whether you can whistle and eat at the same time, or whether you can push or pull an object with a string. This is true common sense knowledge. I can give labels to every tiny sub-section of my hand, thus increasing the resolution of the symbolic description. If I give labels to each very small visual features of my hand, then the distinction between visual representation and symbolic representation disappears. Therefore, I think symbolic KBs like Cyc's is not doomed -- the symbolic KB can merge with perceptual grounding in a continuum fashion. YKY -- agi | Archives | Modify Your Subscription -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.516 / Virus Database: 269.20.7/1285 - Release Date: 2/18/2008 5:50 AM --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Pei: Resolution-based FOL on a huge KB is intractable. Agreed. However Cycorp spend a great deal of programming effort (i.e. many man-years) finding deep inference paths for common queries. The strategies were: prune the rule set according to the contextsubstitute procedural code for modus ponens in common query paths (e.g. isa-links inferred via graph traversal)structure the inference engine as a nested set of iterators so that easy answers are returned immediately, and harder-to-find answers trickle out later.establish a battery of inference engine controls (e.g. time bounds, speed vs. completeness - whether to employ expensive inference strategies for greater coverage of answers) and have the inference engine automatically apply the optimal control configuration for queriesdetermine rule utility via machine learning and apply prioritized inference modules within the given time constraints My last in-house talk at Cycorp, in the summer of 2006, described a notion of mine that Cyc's deductive inference engine behaves as an interpreter, and that for a certain set of queries, a dramatic speed improvement (e.g. four orders of magnitude) could be achieved by compiling the query, and possibly preprocessing incoming facts to suit expected queries. The queries that interested me were those embedded in an intelligent application, and which could be viewed as a query template with parameters. The compilation process I described would explore the parameter space with programmer-chosen query examples. Then the resulting proof trees would be compiled into executable code - avoiding entirely the time consuming candidate rule search and their application when the query executes. My notion for Cyc's deductive inference engine optimization is analogous to SQL query optimization technology. I expect to use this technique in the Texai project at the point when I need a deductive inference engine. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: Pei Wang [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Monday, February 18, 2008 6:17:59 AM Subject: Re: [agi] would anyone want to use a commonsense KB? On Feb 17, 2008 9:42 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: So far I've been using resolution-based FOL, so there's only 1 inference rule and this is not a big issue. If you're using nonstandard inference rules, perhaps even approximate ones, I can see that this distinction is important. Resolution-based FOL on a huge KB is intractable. Pei --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com Looking for last minute shopping deals? Find them fast with Yahoo! Search. http://tools.search.yahoo.com/newsearch/category.php?category=shopping --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Steve, I also agree with what you said, and what Cyc uses is no longer pure resolution-based FOL. A purely resolution-based inference engine is mathematically elegant, but completely impractical, because after all the knowledge are transformed into the clause form required by resolution, most of the semantic information in the knowledge structure is gone, and the result is equivalent to the original knowledge in truth-value only. It is hard to control the direction of the inference without semantic information. Pei On Feb 18, 2008 11:13 AM, Stephen Reed [EMAIL PROTECTED] wrote: Pei: Resolution-based FOL on a huge KB is intractable. Agreed. However Cycorp spend a great deal of programming effort (i.e. many man-years) finding deep inference paths for common queries. The strategies were: prune the rule set according to the context substitute procedural code for modus ponens in common query paths (e.g. isa-links inferred via graph traversal) structure the inference engine as a nested set of iterators so that easy answers are returned immediately, and harder-to-find answers trickle out later. establish a battery of inference engine controls (e.g. time bounds, speed vs. completeness - whether to employ expensive inference strategies for greater coverage of answers) and have the inference engine automatically apply the optimal control configuration for queries determine rule utility via machine learning and apply prioritized inference modules within the given time constraints My last in-house talk at Cycorp, in the summer of 2006, described a notion of mine that Cyc's deductive inference engine behaves as an interpreter, and that for a certain set of queries, a dramatic speed improvement (e.g. four orders of magnitude) could be achieved by compiling the query, and possibly preprocessing incoming facts to suit expected queries. The queries that interested me were those embedded in an intelligent application, and which could be viewed as a query template with parameters. The compilation process I described would explore the parameter space with programmer-chosen query examples. Then the resulting proof trees would be compiled into executable code - avoiding entirely the time consuming candidate rule search and their application when the query executes. My notion for Cyc's deductive inference engine optimization is analogous to SQL query optimization technology. I expect to use this technique in the Texai project at the point when I need a deductive inference engine. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: Pei Wang [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Monday, February 18, 2008 6:17:59 AM Subject: Re: [agi] would anyone want to use a commonsense KB? On Feb 17, 2008 9:42 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: So far I've been using resolution-based FOL, so there's only 1 inference rule and this is not a big issue. If you're using nonstandard inference rules, perhaps even approximate ones, I can see that this distinction is important. Resolution-based FOL on a huge KB is intractable. Pei --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try it now. agi | Archives | Modify Your Subscription --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
--- YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 2/18/08, Matt Mahoney [EMAIL PROTECTED] wrote: Heh... I think you could give away read-only access and charge people to update it. Information has negative value, you know. Well, the idea is to ask lots of people to contribute to the KB, and pay them with virtual credits. (I expect such people to have a little knowledge in logic or Prolog, so they can enter complex rules. Also, they can be assisted by inductive learning algorithms.) The income of the KB will be given back to them. I'll take a bit of administrative fees =) Why would this approach succeed where Cyc failed? Cyc paid people to build the knowledge base. Then when they couldn't sell it, the tried giving it away. Still, nobody used it. For an AGI to be useful, people have to be able to communicate with it in natural language. It is easy to manipulate formulas like if P then Q. It is much harder to explain how this knowledge is represented and learned in a language model. Cyc did not solve this problem, and we see the result. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Pei, Another issue with a KB inference engine as contrasted with a FOL theorem prover is that the former seeks answers to queries, and the latter often seeks to disprove the negation of the theorem by finding a contradiction. Cycorp therefore could not reuse much of the research from the automatic theorem proving community. And on the other hand the database community commonly did not investigate deep inference. As the Semantic Web community continues to develop new deductive inference engines tuned to inference (ie. query answering) over large RDF KBs , I expect to see open-source forward-chaining, and backward-chaining inference engines that can be optimized in the same way that I described for Cyc. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: Pei Wang [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Monday, February 18, 2008 10:47:43 AM Subject: Re: [agi] would anyone want to use a commonsense KB? Steve, I also agree with what you said, and what Cyc uses is no longer pure resolution-based FOL. A purely resolution-based inference engine is mathematically elegant, but completely impractical, because after all the knowledge are transformed into the clause form required by resolution, most of the semantic information in the knowledge structure is gone, and the result is equivalent to the original knowledge in truth-value only. It is hard to control the direction of the inference without semantic information. Pei On Feb 18, 2008 11:13 AM, Stephen Reed [EMAIL PROTECTED] wrote: Pei: Resolution-based FOL on a huge KB is intractable. Agreed. However Cycorp spend a great deal of programming effort (i.e. many man-years) finding deep inference paths for common queries. The strategies were: prune the rule set according to the context substitute procedural code for modus ponens in common query paths (e.g. isa-links inferred via graph traversal) structure the inference engine as a nested set of iterators so that easy answers are returned immediately, and harder-to-find answers trickle out later. establish a battery of inference engine controls (e.g. time bounds, speed vs. completeness - whether to employ expensive inference strategies for greater coverage of answers) and have the inference engine automatically apply the optimal control configuration for queries determine rule utility via machine learning and apply prioritized inference modules within the given time constraints My last in-house talk at Cycorp, in the summer of 2006, described a notion of mine that Cyc's deductive inference engine behaves as an interpreter, and that for a certain set of queries, a dramatic speed improvement (e.g. four orders of magnitude) could be achieved by compiling the query, and possibly preprocessing incoming facts to suit expected queries. The queries that interested me were those embedded in an intelligent application, and which could be viewed as a query template with parameters. The compilation process I described would explore the parameter space with programmer-chosen query examples. Then the resulting proof trees would be compiled into executable code - avoiding entirely the time consuming candidate rule search and their application when the query executes. My notion for Cyc's deductive inference engine optimization is analogous to SQL query optimization technology. I expect to use this technique in the Texai project at the point when I need a deductive inference engine. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: Pei Wang [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Monday, February 18, 2008 6:17:59 AM Subject: Re: [agi] would anyone want to use a commonsense KB? On Feb 17, 2008 9:42 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: So far I've been using resolution-based FOL, so there's only 1 inference rule and this is not a big issue. If you're using nonstandard inference rules, perhaps even approximate ones, I can see that this distinction is important. Resolution-based FOL on a huge KB is intractable. Pei --- agi Archives: http://www.listbox.com/member/archive/303
Re: [agi] would anyone want to use a commonsense KB?
On Feb 18, 2008 12:37 PM, Stephen Reed [EMAIL PROTECTED] wrote: Pei, Another issue with a KB inference engine as contrasted with a FOL theorem prover is that the former seeks answers to queries, and the latter often seeks to disprove the negation of the theorem by finding a contradiction. Cycorp therefore could not reuse much of the research from the automatic theorem proving community. And on the other hand the database community commonly did not investigate deep inference. The automatic theorem proving community does that because resolution by itself is not complete, while resolution-refutation is complete. Pei As the Semantic Web community continues to develop new deductive inference engines tuned to inference (ie. query answering) over large RDF KBs , I expect to see open-source forward-chaining, and backward-chaining inference engines that can be optimized in the same way that I described for Cyc. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: Pei Wang [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Monday, February 18, 2008 10:47:43 AM Subject: Re: [agi] would anyone want to use a commonsense KB? Steve, I also agree with what you said, and what Cyc uses is no longer pure resolution-based FOL. A purely resolution-based inference engine is mathematically elegant, but completely impractical, because after all the knowledge are transformed into the clause form required by resolution, most of the semantic information in the knowledge structure is gone, and the result is equivalent to the original knowledge in truth-value only. It is hard to control the direction of the inference without semantic information. Pei On Feb 18, 2008 11:13 AM, Stephen Reed [EMAIL PROTECTED] wrote: Pei: Resolution-based FOL on a huge KB is intractable. Agreed. However Cycorp spend a great deal of programming effort (i.e. many man-years) finding deep inference paths for common queries. The strategies were: prune the rule set according to the context substitute procedural code for modus ponens in common query paths (e.g. isa-links inferred via graph traversal) structure the inference engine as a nested set of iterators so that easy answers are returned immediately, and harder-to-find answers trickle out later. establish a battery of inference engine controls (e.g. time bounds, speed vs. completeness - whether to employ expensive inference strategies for greater coverage of answers) and have the inference engine automatically apply the optimal control configuration for queries determine rule utility via machine learning and apply prioritized inference modules within the given time constraints My last in-house talk at Cycorp, in the summer of 2006, described a notion of mine that Cyc's deductive inference engine behaves as an interpreter, and that for a certain set of queries, a dramatic speed improvement (e.g. four orders of magnitude) could be achieved by compiling the query, and possibly preprocessing incoming facts to suit expected queries. The queries that interested me were those embedded in an intelligent application, and which could be viewed as a query template with parameters. The compilation process I described would explore the parameter space with programmer-chosen query examples. Then the resulting proof trees would be compiled into executable code - avoiding entirely the time consuming candidate rule search and their application when the query executes. My notion for Cyc's deductive inference engine optimization is analogous to SQL query optimization technology. I expect to use this technique in the Texai project at the point when I need a deductive inference engine. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: Pei Wang [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Monday, February 18, 2008 6:17:59 AM Subject: Re: [agi] would anyone want to use a commonsense KB? On Feb 17, 2008 9:42 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: So far I've been using resolution-based FOL, so there's only 1 inference rule and this is not a big issue. If you're using nonstandard inference rules, perhaps even approximate ones, I can see that this distinction is important. Resolution-based FOL on a huge KB is intractable. Pei --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 9:56 AM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: I'm planning to collect commonsense knowledge into a large KB, in the form of first-order logic, probably very close to CycL. Before you embark on such a project, it might be worth first looking closely at the question of why Cyc hasn't been useful, so that you don't end up making the same mistakes. There's a school of thought, to which I subscribe, that it's because Cyc's knowledge base isn't grounded. Are you instead taking the view that Cyc's fundamental approach is correct, and it just needs a somewhat different choice of logical axioms or whatnot? --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 4:11 PM, Russell Wallace [EMAIL PROTECTED] wrote: On Feb 17, 2008 9:56 AM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: I'm planning to collect commonsense knowledge into a large KB, in the form of first-order logic, probably very close to CycL. Before you embark on such a project, it might be worth first looking closely at the question of why Cyc hasn't been useful, so that you don't end up making the same mistakes. There's a school of thought, to which I subscribe, that it's because Cyc's knowledge base isn't grounded. Are you instead taking the view that Cyc's fundamental approach is correct, and it just needs a somewhat different choice of logical axioms or whatnot? It might be considered 'grounded' in some sense, but the problem is that it isn't used to derive other statements that are grounded in the same sense. Semantics for which Cyc database is grounded (human knowledge of word usage) is different from semantics that is used for inference, so in semantics used for inference it's ungrounded. But it might be possible to make inference engine that will use Cyc database in grounded way. Another question is if Cyc database will be useful for such inference engine. -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 2:11 PM, Russell Wallace [EMAIL PROTECTED] wrote: On Feb 17, 2008 9:56 AM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: I'm planning to collect commonsense knowledge into a large KB, in the form of first-order logic, probably very close to CycL. Before you embark on such a project, it might be worth first looking closely at the question of why Cyc hasn't been useful, so that you don't end up making the same mistakes. This is perhaps a good opportunity to poll you on why do you think Cyc KB hasn't been useful / successful, I'm interested in grounded opinions (Stephen?), and not about Cyc as an AGI but about Cyc KB as what it was supposed to be (e.g. a universal backbone so that expert systems didn't fall off the knowledge cliff). --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 1:49 PM, Vladimir Nesov [EMAIL PROTECTED] wrote: It might be considered 'grounded' in some sense Well, the intent of my statement is this: Maybe somewhere in the Cyc knowledge base there's the assertion Eat(Cats, Mice) or equivalent, but if you show Cyc a picture of a cat, a picture of a mouse, then two candidate successor pictures, one of a mouse stuffed into a cat's mouth and the other of a cat and a mouse talking philosophy over a pint of beer, and ask Cyc which is the more likely successor state, it won't have a clue. That's what I mean by grounded, and it's the reason Cyc isn't useful. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
As Lukasz just pointed out, there are two topics: 1. Cyc as an AGI project 2. Cyc as a knowledge base useful for AGI systems. The grounding problem you raised maybe an issue for 1 (even that depending on what intelligence is understood, and Lenat will argue otherwise), but it is much less an issue for 2, because that function, if it exists in the system, is usually not carried out mainly by the knowledge base. There have been many criticisms to Cyc on 1, and I agree with Lukasz that it may be more fruitful to discuss 2, which is also more relevant to the original question of YKY. My own opinion is: Cyc is too close to first-order predicate logic, which is a good formal language/logic for mathematical knowledge, but not for commonsense knowledge, and minor revisions are not enough. A more detailed discussion is in http://nars.wang.googlepages.com/wang.cognitive_mathematical.pdf Pei On Feb 17, 2008 10:02 AM, Russell Wallace [EMAIL PROTECTED] wrote: On Feb 17, 2008 1:49 PM, Vladimir Nesov [EMAIL PROTECTED] wrote: It might be considered 'grounded' in some sense Well, the intent of my statement is this: Maybe somewhere in the Cyc knowledge base there's the assertion Eat(Cats, Mice) or equivalent, but if you show Cyc a picture of a cat, a picture of a mouse, then two candidate successor pictures, one of a mouse stuffed into a cat's mouth and the other of a cat and a mouse talking philosophy over a pint of beer, and ask Cyc which is the more likely successor state, it won't have a clue. That's what I mean by grounded, and it's the reason Cyc isn't useful. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/17/08, Lukasz Stafiniak [EMAIL PROTECTED] wrote: On Feb 17, 2008 2:11 PM, Russell Wallace [EMAIL PROTECTED] wrote: Before you embark on such a project, it might be worth first looking closely at the question of why Cyc hasn't been useful, so that you don't end up making the same mistakes. This is perhaps a good opportunity to poll you on why do you think Cyc KB hasn't been useful / successful, I'm interested in grounded opinions (Stephen?), and not about Cyc as an AGI but about Cyc KB as what it was supposed to be (e.g. a universal backbone so that expert systems didn't fall off the knowledge cliff). Yes, I'd like to hear others' opinion on Cyc. Personally I don't think it's the perceptual grounding issue -- grounding can be added incrementally later. I think Cyc (the KB) is on the right track, but it doesn't have enough rules. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
In that sense, is a Wikipedia article grounded, if it doesn't contain a photo? Can it be useful for us? I mean, symbol grounding is indeed an important issue, but it doesn't show up everywhere. A Cyc-like KB can be useful, if it use a proper formal language, which allows its concepts to be related to each other, as well as to other items outside the KB, such as sensorimotor mechanism. It would be nice to have a public KB in which the concepts are already linked to images and operations, but since sensorimotor tends to be highly system-dependent, I cannot expect that in the near future. On the other hand, a commonly accepted formal language is much more plausible to get, even though it doesn't provide all necessary knowledge for an AGI. What is in your mind as a grounded KB? Pei On Feb 17, 2008 11:30 AM, Russell Wallace [EMAIL PROTECTED] wrote: On Feb 17, 2008 3:34 PM, Pei Wang [EMAIL PROTECTED] wrote: As Lukasz just pointed out, there are two topics: 1. Cyc as an AGI project 2. Cyc as a knowledge base useful for AGI systems. Well, I'm talking about Cyc (and similar systems) as useful for anything at all (other than experience to tell us what doesn't work and why not). But if it's proposed that such a system might be a useful knowledge base for something, then the something will have to have solved the grounding problem, right? And what I'm saying is, I wouldn't start off building a Cyc-like knowledge base and assume the grounding problem will be solved later. I'd start off with the grounding problem. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Yes, I'd like to hear others' opinion on Cyc. Personally I don't think it's the perceptual grounding issue -- grounding can be added incrementally later. I think Cyc (the KB) is on the right track, but it doesn't have enough rules. I do think it's possible a Cyc approach could work if one had a few billion rules in there -- but so what? (Work meaning: together with a logic engine, serve as the seed for an AGI that really learns and understands) It's clear that the mere millions of rules in their KB now are VASTLY inadequate in terms of scale... Similarly, AIXItl or related approaches could work for AGI if one had an insanely powerful computer -- but so what AGI approaches that could work, in principle if certain wildly infeasible conditions were met, are not hard to come by ... ;=) -- Ben G --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
1) While in my own AI projects I am currently gravitating toward an approach involving virtual-worlds grounding, as a general rule I don't think it's obvious that sensorimotor grounding is needed for AGI. Certainly it's very useful, but there is no strong argument that it's required. The human path to AGI is not the only one. 2) I think that, potentially, building a KB could be part of an approach to solving the grounding problem. Encode some simple knowledge, instruct the system in how to ground it in its sensorimotor experience ... then encode some more (slightly more complex) knowledge ... etc. I'm not saying this is the best way but it seems a viable approach. Thus, even if you want to take a grounding-focused approach, it doesn't follow that fully solving the grounding problem must precede the creation and utilization of a KB. Rather, there could be a solution to the grounding problem that couples a KB with other aspects. In the NM approach, we could proceed with or without a KB, and with or without sensorimotor grounding; and I believe NARS has that same property... My feeling is that sensorimotor grounding is an Extremely Nice to Have whereas a KB is just a Sort of Nice to Have, but I don't have a rigorous demonstration of that -- Ben G On Feb 17, 2008 11:30 AM, Russell Wallace [EMAIL PROTECTED] wrote: On Feb 17, 2008 3:34 PM, Pei Wang [EMAIL PROTECTED] wrote: As Lukasz just pointed out, there are two topics: 1. Cyc as an AGI project 2. Cyc as a knowledge base useful for AGI systems. Well, I'm talking about Cyc (and similar systems) as useful for anything at all (other than experience to tell us what doesn't work and why not). But if it's proposed that such a system might be a useful knowledge base for something, then the something will have to have solved the grounding problem, right? And what I'm saying is, I wouldn't start off building a Cyc-like knowledge base and assume the grounding problem will be solved later. I'd start off with the grounding problem. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] If men cease to believe that they will one day become gods then they will surely become worms. -- Henry Miller --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 4:48 PM, Ben Goertzel [EMAIL PROTECTED] wrote: 1) While in my own AI projects I am currently gravitating toward an approach involving virtual-worlds grounding And I think that's a very good idea. as a general rule I don't think it's obvious that sensorimotor grounding is needed for AGI. Well I wouldn't say it's obvious - it took me a good while to figure it out, after all :) Just true. Then again, AI is a hard problem; very few true things about it are obvious. The human path to AGI is not the only one. Oh indeed - as I said before, I'm not expecting anything like human-equivalent AGI in the foreseeable future. But I still think grounding is central for making useful AI programs. It's an example of the heuristic that applies to software in general: Internal computation is easy. Interfaces are most of the difficulty and most of the value. I don't _want_ to believe that, mind you. Internal computation is much more fun. But reality's rubbed my nose in itself on that one too many times for me to ignore. 2) I think that, potentially, building a KB could be part of an approach to solving the grounding problem. Encode some simple knowledge, instruct the system in how to ground it in its sensorimotor experience ... then encode some more (slightly more complex) knowledge ... etc. I'm not saying this is the best way but it seems a viable approach. Thus, even if you want to take a grounding-focused approach, it doesn't follow that fully solving the grounding problem must precede the creation and utilization of a KB. Rather, there could be a solution to the grounding problem that couples a KB with other aspects. I agree, that might be a viable approach. But the key phrase is Encode some simple knowledge, instruct the system in how to ground it in its sensorimotor experience - i.e. you're _not_ spending a decade writing a million assertions and _then_ looking for the first time at the grounding problem. Instead grounding is addressed, if not as step 1, then at least as step 1.001. My feeling is that sensorimotor grounding is an Extremely Nice to Have whereas a KB is just a Sort of Nice to Have, but I don't have a rigorous demonstration of that Heck, I don't have a rigorous demonstration of any nontrivial fact about any program longer than ten lines, except that any working program provides a rigorous existence proof that the methods it used _can_ solve the problem it solves. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
I agree, that might be a viable approach. But the key phrase is Encode some simple knowledge, instruct the system in how to ground it in its sensorimotor experience - i.e. you're _not_ spending a decade writing a million assertions and _then_ looking for the first time at the grounding problem. Instead grounding is addressed, if not as step 1, then at least as step 1.001. Well, I find that grounding-based AGI is the kind I can think about most easily, since that's how human intelligence works... But I'm less confident that it's the only possible kind of AGI... I've got to wonder if the masses of text on the Internet could, in themselves, display a sufficient richness of patterns to obviate the need for grounding in another domain like a physical or virtual world, or mathematics. In other words, maybe what you think needs to be gotten from grounding in a nonlinguistic domain, could somehow be gotten indirectly via grounding in masses of text? I am not confident this is feasible, nor that it isn't ... and it's not the approach I'm following ... but I'm uncomfortable dismissing it out of hand... -- Ben G --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 5:27 PM, Ben Goertzel [EMAIL PROTECTED] wrote: I've got to wonder if the masses of text on the Internet could, in themselves, display a sufficient richness of patterns to obviate the need for grounding in another domain like a physical or virtual world, or mathematics. In other words, maybe what you think needs to be gotten from grounding in a nonlinguistic domain, could somehow be gotten indirectly via grounding in masses of text? I am not confident this is feasible, nor that it isn't ... and it's not the approach I'm following ... but I'm uncomfortable dismissing it out of hand... *nods* I'm comfortable dismissing it out of hand, for several reasons, not least of which is that we humans do not and cannot do anything remotely resembling what you're proposing. It's been described as the equivalent of trying to learn Chinese equipped only with a Chinese-Chinese dictionary - something already hopelessly impossible for humans. It's actually much worse than that, because we'd start off knowing our own native language, and that Chinese is spoken by humans who have mostly the same concepts as we have. It's actually much worse even than a newborn baby trying to learn Chinese as his first language from a Chinese-Chinese dictionary without ever hearing any form of speech, because the baby starts off with a lot of cognitive machinery about language, the real world and connections between the two, that an AI program doesn't. At the end of the day, the Internet just doesn't contain most of the needed information. Consider the question of whether it's possible to learn about water flowing downhill, from Internet text alone. From Google (example not original to me, though I forget who first ran this test): Results 1 - 10 of about 864 for water flowing downhill Results 1 - 10 of about 2,130 for water flowing uphill The prosecution rests :) --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/17/08, Pei Wang [EMAIL PROTECTED] wrote: There is no similar plan for OpenNARS. When the time comes, it probably will get its knowledge, in a mixed manner, (1) from various existing sources of formatted knowledge, including Cyc, (2) from the Internet, using information retrieval/extraction, data mining, etc., (3) through a natural language interface, (4) through a sensorimotor interface, (5) by human tutoring. The last approach will require manually coded knowledge (commonsense or not), but in a much smaller scale. See http://nars.wang.googlepages.com/wang.roadmap.pdf Thanks, I'll read it and give you some feedback later. But I'm interested in how AGI can be achieved collaboratively, and sharing KBs is one possiblity, and may be very important. Sure, you can go it alone, but that may not be the best choice. I raised this issue before: by logical rules, do you mean inference rules (like Derive conclusion C from premises A and B), or implication statements (like If A and B are true, then C is true)? These two are very often confused with each other, and that confusion has serious consequences. AGI needs plenty of the latter, but just a relatively small number of the former. Sorry... I can't see the distinction. Maybe you mean causation vs implication? For example, eating sweets may cause cavities, but it is not an implication because P(cavities|sweets) != 1? What I mean by rule is any formula that has variables in it. The kind of rules I have in mind... let me give an example. One day I opened the microwave and saw a dish of raw fish inside. I abductively conclude that my mom has put a frozen fish inside to defrost it but was too lazy to wait till it finished to take it out. In order to do this reasoning I need the following facts: 1. the microwave is normally empty when not in use 2. humans can move things around 3. defrosting takes time 4. waiting for the fish to defrost is boring 5. putting the fish inside and forgeting to press the cook button is unlikely because the 2 actions occur closely 6. forgetfulness usually require a substantial time interval 7. etc etc... Obviously the current Cyc KB do not have these facts. That's why I say more facts are needed. Secondly, I suspect that some implicit rules are needed for an inference engine to string these facts together to form a linear proof -- if you get my drift. But I find it hard to explain... YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 12:56 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: I raised this issue before: by logical rules, do you mean inference rules (like Derive conclusion C from premises A and B), or implication statements (like If A and B are true, then C is true)? These two are very often confused with each other, and that confusion has serious consequences. AGI needs plenty of the latter, but just a relatively small number of the former. Sorry... I can't see the distinction. Maybe you mean causation vs implication? For example, eating sweets may cause cavities, but it is not an implication because P(cavities|sweets) != 1? The best example of this difference is Carroll's Paradox --- see http://www.ditext.com/carroll/tortoise.html What I mean by rule is any formula that has variables in it. Both of them can have variables in them. The kind of rules I have in mind... let me give an example. One day I opened the microwave and saw a dish of raw fish inside. I abductively conclude that my mom has put a frozen fish inside to defrost it but was too lazy to wait till it finished to take it out. In order to do this reasoning I need the following facts: 1. the microwave is normally empty when not in use 2. humans can move things around 3. defrosting takes time 4. waiting for the fish to defrost is boring 5. putting the fish inside and forgeting to press the cook button is unlikely because the 2 actions occur closely 6. forgetfulness usually require a substantial time interval 7. etc etc... Then by rules you mean implication statements, not inference rules. Obviously the current Cyc KB do not have these facts. That's why I say more facts are needed. Sure. No KB can be complete in this sense. However I'm not sure if you can do better than Cyc. If you just want to add more knowledge, why not build on the top of Cyc? Secondly, I suspect that some implicit rules are needed for an inference engine to string these facts together to form a linear proof -- if you get my drift. But I find it hard to explain... That will be control rules, which is part of the control mechanism. With commonsense knowledge you cannot really have a proof that settles the truth-value of a conclusion once for all. You can only have arguments, which are much less conclusive. Pei --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
In other words, maybe what you think needs to be gotten from grounding in a nonlinguistic domain, could somehow be gotten indirectly via grounding in masses of text? I am not confident this is feasible, nor that it isn't ... and it's not the approach I'm following ... but I'm uncomfortable dismissing it out of hand... *nods* I'm comfortable dismissing it out of hand, for several reasons, not least of which is that we humans do not and cannot do anything remotely resembling what you're proposing. I don't assume that all successful AGI's must be humanlike... At the end of the day, the Internet just doesn't contain most of the needed information. Consider the question of whether it's possible to learn about water flowing downhill, from Internet text alone. From Google (example not original to me, though I forget who first ran this test): Results 1 - 10 of about 864 for water flowing downhill Results 1 - 10 of about 2,130 for water flowing uphill The prosecution rests :) Google is not an AGI, so I have no idea why you think this proves anything about AGI ... I strongly suspect there is enough information in the text online for an AGI to learn that water flows downhill in most circumstances, without having explicit grounding... -- Ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 6:32 PM, Ben Goertzel [EMAIL PROTECTED] wrote: I don't assume that all successful AGI's must be humanlike... Neither do I - on the contrary, I think a humanlike AGI isn't going to happen, in the same way that we never did achieve birdlike flight. But the only reason we have for believing ill-posed problems (i.e. nearly all the problems presented by the real world) to be solvable at all is that humans (in some cases) provide an existence proof. Where a problem is ill-posed, and humans can't come close to solving it, and we can't point to a specific human limit that would enable us to solve it if overcome, then the reasonable default conclusion is that it's not solvable. Google is not an AGI, so I have no idea why you think this proves anything about AGI ... It doesn't. It does, however, prove something about the contents of the Web, and constitutes a reason... I strongly suspect there is enough information in the text online for an AGI to learn that water flows downhill in most circumstances, without having explicit grounding... ...for disagreeing with you on this point. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On Feb 17, 2008 7:37 PM, Ben Goertzel [EMAIL PROTECTED] wrote: water flows downhill site:wikipedia.org -- 4 results water flows uphill site:wikipedia.org -- 2 results BUT, both of the latter 2 results are within user talk pages, not regular wikipedia entries... whereas all of the former 4 results are on regular wikipedia entries I'm not saying one can use wikipedia as the knowledge base for an AGI, it's clearly not big enough, but I think this certainly defuses your example a bit obviously, defusing more complex examples would require more work -- many useful pieces of info exist only implicitly among various Web pages rather than on a single Web page... Of course no matter what example I give, you can defuse it with an appropriate amount of work. But the key word in that isn't defuse, it's you. Yes, _you_, being a general intelligence with the requisite real-world knowledge, can know what's relevant and what isn't, and therefore ignore the documents that aren't relevant, keep searching until you find one that is, and regard it as the answer. For an AI program to do that, it would have to start off with precisely the sort of real-world knowledge that I've been talking about. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/18/08, Pei Wang [EMAIL PROTECTED] wrote: I raised this issue before: by logical rules, do you mean inference rules (like Derive conclusion C from premises A and B), or implication statements (like If A and B are true, then C is true)? These two are very often confused with each other, and that confusion has serious consequences. AGI needs plenty of the latter, but just a relatively small number of the former. Sorry... I can't see the distinction. Maybe you mean causation vs implication? For example, eating sweets may cause cavities, but it is not an implication because P(cavities|sweets) != 1? The best example of this difference is Carroll's Paradox --- see http://www.ditext.com/carroll/tortoise.html I'm reading this: http://en.wikipedia.org/wiki/What_the_Tortoise_Said_to_Achilles which is easier to understand, but I still don't get it. Achilles grants to the tortoise that the second kind of reader may exist, but I think this second kind of reader is absurd. If A and B are true, then a sane person MUST admit that Z is true. I don't see why not? YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Only statements containing in a KB as content have truth-value, or need acceptance. An inference rule is part of the system, which just applies, and does not need acceptance within the system. An inference rule has no truth-value. If it is still unclear, try this: http://www.mathacademy.com/pr/prime/articles/carroll/index.asp Of course, the two are related (see http://en.wikipedia.org/wiki/Deduction_theorem), but if you have the two confused when designing an inference system, you'll run into trouble. Pei On Feb 17, 2008 2:53 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 2/18/08, Pei Wang [EMAIL PROTECTED] wrote: I raised this issue before: by logical rules, do you mean inference rules (like Derive conclusion C from premises A and B), or implication statements (like If A and B are true, then C is true)? These two are very often confused with each other, and that confusion has serious consequences. AGI needs plenty of the latter, but just a relatively small number of the former. Sorry... I can't see the distinction. Maybe you mean causation vs implication? For example, eating sweets may cause cavities, but it is not an implication because P(cavities|sweets) != 1? The best example of this difference is Carroll's Paradox --- see http://www.ditext.com/carroll/tortoise.html I'm reading this: http://en.wikipedia.org/wiki/What_the_Tortoise_Said_to_Achilles which is easier to understand, but I still don't get it. Achilles grants to the tortoise that the second kind of reader may exist, but I think this second kind of reader is absurd. If A and B are true, then a sane person MUST admit that Z is true. I don't see why not? YKY agi | Archives | Modify Your Subscription --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Mark Waser wrote: I've got to wonder if the masses of text on the Internet could, in themselves, display a sufficient richness of patterns to obviate the need for grounding in another domain like a physical or virtual world, or mathematics. A system is grounded if it's internal representations are internally consistent and map accurately and completely to the experiences possible in a (physical or virtual domain). Expert systems are not grounded because they do not map completely. There is always some additional factor that they do not experience or account for. Most typical proto-AGI systems pretend to ground because they use English words that are grounded for the observer but which are not grounded for the system because they have a meaning which is enforced upon the system without being understood by the system. A system which can only experience text still could be grounded in the physical world provided that there is enough text to describe the physical world well enough for the system to be grounded. Couldn't any of us be said to still be grounded in the physical world even if we were removed from it except for a text interface? The real trick is to get a system to state where entirely internally consistent (in terms of definitions, etc., not predictions) and large enough to be useful. I applaud your attempt to bring some sense to this discussion. It won't work, of course. There is just no way to stop people having meaningless discussions about grounding, in which the thing they mean by the word has only a distant relationship to the real meaning. Pity, because the real thing is indeed worth discussing. Richard Loosemore --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/18/08, Ben Goertzel [EMAIL PROTECTED] wrote: I strongly suspect there is enough information in the text online for an AGI to learn that water flows downhill in most circumstances, without having explicit grounding... I strongly suspect the contrary =) for the simple reason that adults don't talk about things that all 3-year-olds know -- such knowledge is often implicit in conversations / writings. That's why I believe that some manual method of teaching is *necessary*, hence we need commonsense collection from web users. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Yes, I would be very glad to incorporate any content that I can then republish using a Wikipedia-compatible license, e.g. GNU Free Documentation License. Any weaker license, such as Apache, BSD would be OK too. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: YKY (Yan King Yin) [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Sunday, February 17, 2008 3:56:40 AM Subject: [agi] would anyone want to use a commonsense KB? I'm planning to collect commonsense knowledge into a large KB, in the form of first-order logic, probably very close to CycL. Would current AGI projects (OpenNARS, OpenCog, Texai, etc) find it useful? Or would you prefer to collect commonsense on your own? It seems that the Cyc KB is focused on building an ontology (hence a lot of is_a relations), and there's not enough emphasis on other logical rules. I anticipate that AGIs will need plenty of such rules. YKY agi | Archives | Modify Your Subscription Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try it now. http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Briefly, I think that Cyc indeed has solved the brittleness problem observed with 1980's style narrow-domain expert systems. During the Halo project, Cyc was merely extended in a principled fashion to answer a battery of word questions in the chemistry domain. In my opinion the chief drawback of the Cycorp approach to commonsense knowledge is their overwhelming emphasis on what Cyc knows, as contrasted with what skills has Cyc learned and can demonstrate. My own work is the construction of a bootstrap English dialog system for the purpose of linguistic knowledge *and* skill acquisition. Also by using a robotics-style hierarchical control system, I hope to later connect high level symbolic concepts with low-level perceptions - as somewhat illustrated by current driverless cars. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 - Original Message From: YKY (Yan King Yin) [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Sunday, February 17, 2008 10:51:12 AM Subject: Re: [agi] would anyone want to use a commonsense KB? On 2/17/08, Lukasz Stafiniak [EMAIL PROTECTED] wrote: On Feb 17, 2008 2:11 PM, Russell Wallace [EMAIL PROTECTED] wrote: Before you embark on such a project, it might be worth first looking closely at the question of why Cyc hasn't been useful, so that you don't end up making the same mistakes. This is perhaps a good opportunity to poll you on why do you think Cyc KB hasn't been useful / successful, I'm interested in grounded opinions (Stephen?), and not about Cyc as an AGI but about Cyc KB as what it was supposed to be (e.g. a universal backbone so that expert systems didn't fall off the knowledge cliff). Yes, I'd like to hear others' opinion on Cyc. Personally I don't think it's the perceptual grounding issue -- grounding can be added incrementally later. I think Cyc (the KB) is on the right track, but it doesn't have enough rules. YKY agi | Archives | Modify Your Subscription Looking for last minute shopping deals? Find them fast with Yahoo! Search. http://tools.search.yahoo.com/newsearch/category.php?category=shopping --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
IMHO Cyc was doomed by the lack of a natural language interface. It cannot map between Eats(cats, mice) and cats eat mice, or recognize their equivalence. In Cyc, cats and Eats are just labels used to help human programmers enter facts. Without a natural language interface, it is very expensive to verify and update the knowledge base. More importantly, there is no human interface. It's not that Cycorp isn't aware of the problem. Last year some people at Cycorp were interested in entering the Hutter text compression contest, but they wanted us to change the rules to not count the size of the database (we declined). Text compression or prediction is AI-complete, but it would require a natural language model to predict the next word in cats eat. The example rule I gave seems trivial to solve, but anyone who has worked with NLP knows it is not, of course. I believe the fundamental design error was to insert knowledge at the wrong end. Children learn lexical rules first (segmenting continuous speech at 7-10 months), then semantics (starting at 12 months), then grammar (2-3 years), then logical rules. Structured rules take a lot less computing power to implement than language statistics, so they had to skip the earlier steps, especially in the 1980's when Cyc was launched. As a result, Cyc has no theory that explains how people learn and apply language and facts or how they communicate. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com