AW: [agi] If your AGI can't learn to play chess it is no AGI
The goal of chess is well defined: Avoid being checkmate and try to checkmate your opponent. What checkmate means can be specified formally. Humans mainly learn chess from playing chess. Obviously their knowledge about other domains are not sufficient for most beginners to be a good chess player at once. This can be proven empirically. Thus an AGI would not learn chess completely different from all what we now. It would learn from experience which is one of the most common kinds of learning. I am sure that everyone who learns chess by playing against chess computers and is able to learn good chess playing (which is not sure as also not everyone can learn to be a good mathematician) will be able to be a good chess player against humans. My first posting in this thread shows the very weak point in the argumentation of those people who say that social and other experiences are needed to play chess. You suppose knowledge must be available from another domain to solve problems of the domain of chess. But everything of chess in on the chessboard itself. If you are not able to solve chess problems from chess alone then you are not able to solve certain solvable problems. And thus you cannot call your AI AGI. If you give an AGI all facts which are sufficient to solve a problem then your AGI must be able to solve the problem using nothing else than these facts. If you do not agree with this, then how should an AGI know which experiences in which other domains are necessary to solve the problem? The magic you use is the overestimation of real-world experiences. It sounds as if the ability to solve arbitrary problems in arbitrary domains depend essentially on that your AGI plays in virtual gardens and speaks often with other people. But this is completely nonsense. No one can play good chess by those experiences. Thus such experiences are not sufficient. On the other hand there are programs which definitely do not have such experiences and outperform humans in chess. Thus those experiences are neither sufficient nor necessary to play good chess and emphasizing on such experiences is mystifying AGI, similar as it is done by the doubters of AGI who always argue with Goedel or quantum physics which in fact has no relevance for practical AGI at all. - Matthias Trent Waddington [mailto:[EMAIL PROTECTED] wrote Gesendet: Donnerstag, 23. Oktober 2008 07:42 An: agi@v2.listbox.com Betreff: Re: [agi] If your AGI can't learn to play chess it is no AGI On Thu, Oct 23, 2008 at 3:19 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I do not think that it is essential for the quality of my chess who had taught me to play chess. I could have learned the rules from a book alone. Of course these rules are written in a language. But this is not important for the quality of my chess. If a system is in state x then it is not essential for the future how x was generated. Thus a programmer can hardcode the rules of chess in his AGI and then, concerning chess the AGI would be in the same state as if someone teaches the AGI the chess rules via language. The social aspect of learning chess is of no relevance. Sigh. Ok, let's say I grant you the stipulation that you can hard code the rules of chess some how. My next question is, in a goal-based AGI system, what goal are you going to set and how are you going to set it? You've ruled out language, so you're going to have to hard code the goal too, so excuse my use of language: Play good chess Oh.. that sounds implementable. Maybe you'll give it a copy of GNUChess and let it go at it.. but I've known *humans* who learnt to play chess that way and they get trounced by the first human they play against. How are you going to go about making an AGI that can learn chess in a complete different way to all the known ways of learning chess? Or is the AGI supposed to figure that out? I don't understand why so many of the people on this list seem to think AGI = magic. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Thu, Oct 23, 2008 at 4:13 AM, Trent Waddington [EMAIL PROTECTED] wrote: On Thu, Oct 23, 2008 at 8:39 AM, Vladimir Nesov [EMAIL PROTECTED] wrote: If you consider programming an AI social activity, you very unnaturally generalized this term, confusing other people. Chess programs do learn (certainly some of them, and I guess most of them), not everything is hardcoded. They may learn tactics or even how to prune their tree better, but I know of no chess AI that learns how to play the same way you would say a person learns how to play. Of course. And that's the whole point of this general AI thing we're trying to get across.. learning how to do a task given appropriate instruction and feedback by a teacher is the golden goose here.. Not necessarily. The ultimate teacher is our real environment in general. -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: AW: [agi] Language learning (was Re: Defining AGI)
On Thu, Oct 23, 2008 at 12:55 AM, Matt Mahoney wrote: I suppose you are right. Instead of encoding mathematical rules as a grammar, with enough training data you can just code all possible instances that are likely to be encountered. For example, instead of a grammar rule to encode the commutative law of addition, 5 + 3 = a + b = b + a = 3 + 5 a model with a much larger training data set could just encode instances with no generalization: 12 + 7 = 7 + 12 92 + 0.5 = 0.5 + 92 etc. I believe this is how Google gets away with brute force n-gram statistics instead of more sophisticated grammars. It's language model is probably 10^5 times larger than a human model (10^14 bits vs 10^9 bits). Shannon observed in 1949 that random strings generated by n-gram models of English (where n is the number of either letters or words) look like natural language up to length 2n. For a typical human sized model (1 GB text), n is about 3 words. To model strings longer than 6 words we would need more sophisticated grammar rules. Google can model 5-grams (see http://googleresearch.blogspot.com/2006/08/all-our-n-gram-are-belong-to-you.html ), so it is able to generate and recognize (thus appear to understand) sentences up to about 10 words. Gigantic databases are indeed Google's secret sauce. See: http://googleresearch.blogspot.com/2008/09/doubling-up.html Quote: Monday, September 29, 2008 Posted by Franz Josef Och Machine translation is hard. Natural languages are so complex and have so many ambiguities and exceptions that teaching a computer to translate between them turned out to be a much harder problem than people thought when the field of machine translation was born over 50 years ago. At Google Research, our approach is to have the machines learn to translate by using learning algorithms on gigantic amounts of monolingual and translated data. Another knowledge source is user suggestions. This approach allows us to constantly improve the quality of machine translations as we mine more data and get more and more feedback from users. A nice property of the learning algorithms that we use is that they are largely language independent -- we use the same set of core algorithms for all languages. So this means if we find a lot of translated data for a new language, we can just run our algorithms and build a new translation system for that language. As a result, we were recently able to significantly increase the number of languages on translate.google.com. Last week, we launched eleven new languages: Catalan, Filipino, Hebrew, Indonesian, Latvian, Lithuanian, Serbian, Slovak, Slovenian, Ukrainian, Vietnamese. This increases the total number of languages from 23 to 34. Since we offer translation between any of those languages this increases the number of language pairs from 506 to 1122 (well, depending on how you count simplified and traditional Chinese you might get even larger numbers). - BillK --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Understanding and Problem Solving
Once again, there is a depth to understanding - it's not simply a binary proposition. Don't you agree that a grandmaster understands chess better than you do, even if his moves are understandable to you in hindsight? If I'm not good at math, I might not be able to solve y=3x+4 for x, but I might understand that y equals 3 times x plus four. My understanding is superficial compared to someone who can solve for x. Finally, don't you agree that understanding natural language requires solving problems? If not, how would you account for an AI's ability to understand novel metaphor? Terren --- On Thu, 10/23/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: From: Dr. Matthias Heger [EMAIL PROTECTED] Subject: [agi] Understanding and Problem Solving To: agi@v2.listbox.com Date: Thursday, October 23, 2008, 1:47 AM Terren Suydam wrote: Understanding goes far beyond mere knowledge - understanding *is* the ability to solve problems. One's understanding of a situation or problem is only as deep as one's (theoretical) ability to act in such a way as to achieve a desired outcome. I disagree. A grandmaster of chess can explain his decisions and I will understand them. Einstein could explain his theory to other physicist(at least a subset) and they could understand it. I can read a proof in mathematics and I will understand it – because I only have to understand (= check) every step of the proof. Problem solving is much much more than only understanding. Problem solving is the ability to *create* a sequence of actions to change a system’s state from A to a desired state B. For example: Problem Find a path from A to B within a graph. An algorithm which can check a solution and can answer details about the solution is not necessarily able to find a solution. -Matthias agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: AW: [agi] Language learning (was Re: Defining AGI)
I have already proved something stronger What would you consider your best reference/paper outlining your arguments? Thanks in advance. - Original Message - From: Matt Mahoney [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 8:55 PM Subject: Re: AW: AW: [agi] Language learning (was Re: Defining AGI) --- On Wed, 10/22/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: You make the implicit assumption that a natural language understanding system will pass the turing test. Can you prove this? If you accept that a language model is a probability distribution over text, then I have already proved something stronger. A language model exactly duplicates the distribution of answers that a human would give. The output is indistinguishable by any test. In fact a judge would have some uncertainty about other people's language models. A judge could be expected to attribute some errors in the model to normal human variation. Furthermore, it is just an assumption that the ability to have and to apply the rules are really necessary to pass the turing test. For these two reasons, you still haven't shown 3a and 3b. I suppose you are right. Instead of encoding mathematical rules as a grammar, with enough training data you can just code all possible instances that are likely to be encountered. For example, instead of a grammar rule to encode the commutative law of addition, 5 + 3 = a + b = b + a = 3 + 5 a model with a much larger training data set could just encode instances with no generalization: 12 + 7 = 7 + 12 92 + 0.5 = 0.5 + 92 etc. I believe this is how Google gets away with brute force n-gram statistics instead of more sophisticated grammars. It's language model is probably 10^5 times larger than a human model (10^14 bits vs 10^9 bits). Shannon observed in 1949 that random strings generated by n-gram models of English (where n is the number of either letters or words) look like natural language up to length 2n. For a typical human sized model (1 GB text), n is about 3 words. To model strings longer than 6 words we would need more sophisticated grammar rules. Google can model 5-grams (see http://googleresearch.blogspot.com/2006/08/all-our-n-gram-are-belong-to-you.html ) , so it is able to generate and recognize (thus appear to understand) sentences up to about 10 words. By the way: The turing test must convince 30% of the people. Today there is a system which can already convince 25% http://www.sciencedaily.com/releases/2008/10/081013112148.htm It would be interesting to see a version of the Turing test where the human confederate, machine, and judge all have access to a computer with an internet connection. I wonder if this intelligence augmentation would make the test easier or harder to pass? -Matthias 3) you apply rules such as 5 * 7 = 35 - 35 / 7 = 5 but you have not shown that 3a) that a language understanding system necessarily(!) has this rules 3b) that a language understanding system necessarily(!) can apply such rules It must have the rules and apply them to pass the Turing test. -- Matt Mahoney, [EMAIL PROTECTED] -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] constructivist issues
But, I still do not agree with the way you are using the incompleteness theorem. Um. OK. Could you point to a specific example where you disagree? I'm a little at a loss here . . . . It is important to distinguish between two different types of incompleteness. 1. Normal Incompleteness-- a logical theory fails to completely specify something. 2. Godelian Incompleteness-- a logical theory fails to completely specify something, even though we want it to. I'm also not getting this. If I read the words, it looks like the difference between Normal and Godelian incompleteness is based upon our desires. I think I'm having a complete disconnect with your intended meaning. However, it seems like all you need is type 1 completeness for what you are saying. So, Godel's theorem is way overkill here in my opinion. Um. OK. So I used a bazooka on a fly? If it was a really pesky fly and I didn't destroy anything else, is that wrong? :-) It seems as if you're not arguing with my conclusion but saying that my arguments were way better than they needed to be (like I'm being over-efficient?) . . . . = = = = = Seriously though, I having a complete disconnect here. Maybe I'm just having a bad morning but . . . huh? :-) If I read the words, all I'm getting is that you disagree with the way that I am using the theory because the theory is overkill for what is necessary. - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 9:05 PM Subject: Re: [agi] constructivist issues Mark, I own and have read the book-- but my first introduction to Godel's Theorem was Douglas Hofstadter's earlier work, Godel Escher Bach. Since I had already been guided through the details of the proof (and grappled with the consequences), to be honest chapter 10 you refer to was a little boring :). But, I still do not agree with the way you are using the incompleteness theorem. It is important to distinguish between two different types of incompleteness. 1. Normal Incompleteness-- a logical theory fails to completely specify something. 2. Godelian Incompleteness-- a logical theory fails to completely specify something, even though we want it to. Logicians always mean type 2 incompleteness when they use the term. To formalize the difference between the two, the measuring stick of semantics is used. If a logic's provably-true statements don't match up to its semantically-true statements, it is incomplete. However, it seems like all you need is type 1 completeness for what you are saying. Nobody claims that there is a complete, well-defined semantics for natural language against which we could measure the provably-true (whatever THAT would mean). So, Godel's theorem is way overkill here in my opinion. --Abram On Wed, Oct 22, 2008 at 7:48 PM, Mark Waser [EMAIL PROTECTED] wrote: Most of what I was thinking of and referring to is in Chapter 10. Gödel's Quintessential Strange Loop (pages 125-145 in my version) but I would suggest that you really need to read the shorter Chapter 9. Pattern and Provability (pages 113-122) first. I actually had them conflated into a single chapter in my memory. I think that you'll enjoy them tremendously. - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 4:19 PM Subject: Re: [agi] constructivist issues Mark, Chapter number please? --Abram On Wed, Oct 22, 2008 at 1:16 PM, Mark Waser [EMAIL PROTECTED] wrote: Douglas Hofstadter's newest book I Am A Strange Loop (currently available from Amazon for $7.99 - http://www.amazon.com/Am-Strange-Loop-Douglas-Hofstadter/dp/B001FA23HM) has an excellent chapter showing Godel in syntax and semantics. I highly recommend it. The upshot is that while it is easily possible to define a complete formal system of syntax, that formal system can always be used to convey something (some semantics) that is (are) outside/beyond the system -- OR, to paraphrase -- meaning is always incomplete because it can always be added to even inside a formal system of syntax. This is why I contend that language translation ends up being AGI-complete (although bounded subsets clearly don't need to be -- the question is whether you get a usable/useful subset more easily with or without first creating a seed AGI). - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 12:38 PM Subject: Re: [agi] constructivist issues Mark, The way you invoke Godel's Theorem is strange to me... perhaps you have explained your argument more fully elsewhere, but as it stands I do not see your reasoning. --Abram On Wed, Oct 22, 2008 at 12:20 PM, Mark Waser [EMAIL PROTECTED] wrote: It looks like all this disambiguation by moving to a more formal language is about sweeping the problem under the rug, removing the need for uncertain reasoning from surface
Re: Lojban (was Re: [agi] constructivist issues)
Hi. I don't understand the following statements. Could you explain it some more? - Natural language can be learned from examples. Formal language can not. I think that you're basing this upon the methods that *you* would apply to each of the types of language. It makes sense to me that because of the regularities of a formal language that you would be able to use more effective methods -- but it doesn't mean that the methods used on natural language wouldn't work (just that they would be as inefficient as they are on natural languages. I would also argue that the same argument applies to the first statement of following the following two. - Formal language must be parsed before it can be understood. Natural language must be understood before it can be parsed. - Original Message - From: Matt Mahoney To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 9:23 PM Subject: Lojban (was Re: [agi] constructivist issues) Why would anyone use a simplified or formalized English (with regular grammar and no ambiguities) as a path to natural language understanding? Formal language processing has nothing to do with natural language processing other than sharing a common lexicon that make them appear superficially similar. - Natural language can be learned from examples. Formal language can not. - Formal language has an exact grammar and semantics. Natural language does not. - Formal language must be parsed before it can be understood. Natural language must be understood before it can be parsed. - Formal language is designed to be processed efficiently on a fast, reliable, sequential computer that neither makes nor tolerates errors, between systems that have identical, fixed language models. Natural language evolved to be processed efficiently by a slow, unreliable, massively parallel computer with enormous memory in a noisy environment between systems that have different but adaptive language models. So how does yet another formal language processing system help us understand natural language? This route has been a dead end for 50 years, in spite of the ability to always make some initial progress before getting stuck. -- Matt Mahoney, [EMAIL PROTECTED] --- On Wed, 10/22/08, Ben Goertzel [EMAIL PROTECTED] wrote: From: Ben Goertzel [EMAIL PROTECTED] Subject: Re: [agi] constructivist issues To: agi@v2.listbox.com Cc: [EMAIL PROTECTED] Date: Wednesday, October 22, 2008, 12:27 PM This is the standard Lojban dictionary http://jbovlaste.lojban.org/ I am not so worried about word meanings, they can always be handled via reference to WordNet via usages like run_1, run_2, etc. ... or as you say by using rarer, less ambiguous words Prepositions are more worrisome, however, I suppose they can be handled in a similar way, e.g. by defining an ontology of preposition meanings like with_1, with_2, with_3, etc. In fact we had someone spend a couple months integrating existing resources into a preposition-meaning ontology like this a while back ... the so-called PrepositionWordNet ... or as it eventually came to be called the LARDict or LogicalArgumentRelationshipDictionary ... I think it would be feasible to tweak RelEx to recognize these sorts of subscripts, and in this way to recognize a highly controlled English that would be unproblematic to map semantically... We would then say e.g. I ate dinner with_2 my fork I live in_2 Maryland I have lived_6 for_3 41 years (where I suppress all _1's, so that e.g. ate means ate_1) Because, RelEx already happily parses the syntax of all simple sentences, so the only real hassle to deal with is disambiguation. We could use similar hacking for reference resolution, temporal sequencing, etc. The terrorists_v1 robbed_v2 my house. After that_v2, the jerks_v1 urinated in_3 my yard. I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. -- Ben G On Wed, Oct 22, 2008 at 12:00 PM, Mark Waser [EMAIL PROTECTED] wrote: IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity Actually, I've been making pretty good progress. You just always use big words and never use small words and/or you use a specific phrase as a word. Ambiguous prepositions just disambiguate to one of three/four/five/more possible unambiguous words/phrases. The problem is that most previous subsets (Simplified English, Basic
Re: Lojban (was Re: [agi] constructivist issues)
--- On Thu, 10/23/08, Mark Waser [EMAIL PROTECTED] wrote: Hi. I don't understand the following statements. Could you explain it some more? - Natural language can be learned from examples. Formal language can not. I really mean that formal languages like C++ and HTML are not designed to be learned by the machines that implement them. We write a formal specification of their syntax and semantics. Obviously they are learnable by humans in the same way that humans learn natural languages -- by generalizing from lots of examples. Formal languages serve as a bridge between humans and machines. As such, a language is designed as a compromise between ease of machine specification and ease of human learnability. - Formal language must be parsed before it can be understood. Natural language must be understood before it can be parsed. In formal languages, the meaning of sentence depends heavily on its parse, for example: a = b - c; // a comment b = c - a; // a comment // a - b = c; a comment In natural language, a parse depends greatly on the meanings of the words. For example: - I ate pizza with chopsticks. - I ate pizza with pepperoni. - I ate pizza with Bob. But word order has only a small effect on meaning: - With Bob I ate pizza. - I with Bob ate pizza. - Pizza Bob I ate with. This is my objection to using formal languages to train AGI in a childhood development model like OpenCog (artificial toddler, child, adult, scientist). A child would be trained on single words with semantic content like pizza. Then an adult would learn increasingly complex grammatical structures. Only at the scientist level would an AGI be capable of learning formal languages. There really isn't any stage where a clean language like Lojban or Esperanto seems to help much with knowledge acquisition. If it did, then we would be teaching it in our schools. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: [agi] Understanding and Problem Solving
Guys, A slightly weird conversation. *Everything* cognitive involves problem-solving. Perception (is it a bird or a plane?) involves problem-solving. Perhaps what you really mean is ...involves *deliberate/conscious* problem-solving as opposed to *automatic/unconscious* problem-solving ? Matthias, I say understanding natural language requires the ability to solve problems. Do you disagree? If so, then you must have an explanation for how an AI that could understand language would be able to understand novel metaphors or analogies without doing any active problem-solving. What is your explanation for that? If on the other hand you agree that NLU entails problem-solving, then that is a start. From there we can argue whether the problem-solving abilities necessary for NLU are sufficient to allow problem-solving to occur in any domain (as I have argued). Terren --- On Thu, 10/23/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: From: Dr. Matthias Heger [EMAIL PROTECTED] Subject: AW: [agi] Understanding and Problem Solving To: agi@v2.listbox.com Date: Thursday, October 23, 2008, 10:12 AM I do not agree. Understanding a domain does not imply the ability to solve problems in that domain. And the ability to solve problems in a domain even does not imply to have a generally a deeper understanding of that domain. Once again my example of the problem to find a path within a graph from node A to node B: Program p1 (= problem solver) can find a path. Program p2 (= expert in understanding) can verify and analyze paths. For instance, p2 could be able compare the length of the path for the first half of the nodes with the length of the path for the second half of the nodes. It is not necessary that P1 can do this as well. P2 can not necessarily find a path. But p1 can not necessarily analyze its solution. Understanding and problem solving are different things which might have a common subset but it is wrong that the one implies the other one or vice versa. And that’s the main reason why natural language understanding is not necessarily AGI-complete. -Matthias Terren Suydam [mailto:[EMAIL PROTECTED] wrote: Once again, there is a depth to understanding - it's not simply a binary proposition. Don't you agree that a grandmaster understands chess better than you do, even if his moves are understandable to you in hindsight? If I'm not good at math, I might not be able to solve y=3x+4 for x, but I might understand that y equals 3 times x plus four. My understanding is superficial compared to someone who can solve for x. Finally, don't you agree that understanding natural language requires solving problems? If not, how would you account for an AI's ability to understand novel metaphor? Terren --- On Thu, 10/23/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: From: Dr. Matthias Heger [EMAIL PROTECTED] Subject: [agi] Understanding and Problem Solving To: agi@v2.listbox.com Date: Thursday, October 23, 2008, 1:47 AM Terren Suydam wrote: Understanding goes far beyond mere knowledge - understanding *is* the ability to solve problems. One's understanding of a situation or problem is only as deep as one's (theoretical) ability to act in such a way as to achieve a desired outcome. I disagree. A grandmaster of chess can explain his decisions and I will understand them. Einstein could explain his theory to other physicist(at least a subset) and they could understand it. I can read a proof in mathematics and I will understand it – because I only have to understand (= check) every step of the proof. Problem solving is much much more than only understanding. Problem solving is the ability to *create* a sequence of actions to change a system’s state from A to a desired state B. For example: Problem Find a path from A to B within a graph. An algorithm which can check a solution and can answer details about the solution is not necessarily able to find a solution. -Matthias agi | Archives | Modify Your Subscription -- agi | Archives | Modify Your Subscription --
AW: AW: [agi] Understanding and Problem Solving
Natural language understanding is a problem. And a system with the ability to understand natural language is obviously able to solve *this* problem. But the ability to talk about certain domains does not imply the ability to solve the problems in this domain. I have argued this point with my example of the two programs for the domain of graphs. As Ben has said, it essentially depends on definitions. Probably, you have a different understanding of the meaning of understanding ;-) But for me there is a difference between understanding a domain and the ability to solve problems in a domain. I can understand a car but this does not imply that I can drive a car. I can understand a proof but this does not imply that I can create it. My computer understands my programs because it executes every step correctly but it cannot create a single statement in the language it understands. Did you never experienced a situation where you could not solve a problem but when another person has shown you the solution you understood it at once? You could not create it but you needed not to learn to understand it. Of course, often when you see a solution for a problem then you learn to solve it at the same time. But this is exactly the reason why you have the illusion that understanding and problem solving are the same. Think about a very difficult proof. You can understand every step. But when you get just an empty piece of paper to write it down again then you cannot remember the whole proof and thus you cannot create it. But you can understand it, if you read it. Obviously there is a difference between understanding and problem solving. . I am sure, you want to define understanding differently. But I do not agree because then the term understanding would be overloaded and too much mystified. And we already have too much terms which are unnecessarily mystified in AI. - Matthias Terren Suydam [mailto:[EMAIL PROTECTED] wrote Matthias, I say understanding natural language requires the ability to solve problems. Do you disagree? If so, then you must have an explanation for how an AI that could understand language would be able to understand novel metaphors or analogies without doing any active problem-solving. What is your explanation for that? If on the other hand you agree that NLU entails problem-solving, then that is a start. From there we can argue whether the problem-solving abilities necessary for NLU are sufficient to allow problem-solving to occur in any domain (as I have argued). Terren --- On Thu, 10/23/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: From: Dr. Matthias Heger [EMAIL PROTECTED] Subject: AW: [agi] Understanding and Problem Solving To: agi@v2.listbox.com Date: Thursday, October 23, 2008, 10:12 AM I do not agree. Understanding a domain does not imply the ability to solve problems in that domain. And the ability to solve problems in a domain even does not imply to have a generally a deeper understanding of that domain. Once again my example of the problem to find a path within a graph from node A to node B: Program p1 (= problem solver) can find a path. Program p2 (= expert in understanding) can verify and analyze paths. For instance, p2 could be able compare the length of the path for the first half of the nodes with the length of the path for the second half of the nodes. It is not necessary that P1 can do this as well. P2 can not necessarily find a path. But p1 can not necessarily analyze its solution. Understanding and problem solving are different things which might have a common subset but it is wrong that the one implies the other one or vice versa. And that's the main reason why natural language understanding is not necessarily AGI-complete. -Matthias Terren Suydam [mailto:[EMAIL PROTECTED] wrote: Once again, there is a depth to understanding - it's not simply a binary proposition. Don't you agree that a grandmaster understands chess better than you do, even if his moves are understandable to you in hindsight? If I'm not good at math, I might not be able to solve y=3x+4 for x, but I might understand that y equals 3 times x plus four. My understanding is superficial compared to someone who can solve for x. Finally, don't you agree that understanding natural language requires solving problems? If not, how would you account for an AI's ability to understand novel metaphor? Terren --- On Thu, 10/23/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: From: Dr. Matthias Heger [EMAIL PROTECTED] Subject: [agi] Understanding and Problem Solving To: agi@v2.listbox.com Date: Thursday, October 23, 2008, 1:47 AM Terren Suydam wrote: Understanding goes far beyond mere knowledge - understanding *is* the ability to solve problems. One's understanding of a situation or problem is only as deep as one's (theoretical) ability to act in such a way as to achieve a desired outcome. I disagree. A
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Fri, Oct 24, 2008 at 8:41 AM, Ben Goertzel [EMAIL PROTECTED] wrote: Yes ... at the moment the styles of human and computer chess players are different enough that doing well against computer players does not imply doing nearly equally well against human players ... though it certainly helps a lot ... Does it? I've heard many chess instructors say that playing against a computer hinders young players more than it helps. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Thu, Oct 23, 2008 at 6:46 PM, Trent Waddington [EMAIL PROTECTED] wrote: On Fri, Oct 24, 2008 at 8:41 AM, Ben Goertzel [EMAIL PROTECTED] wrote: Yes ... at the moment the styles of human and computer chess players are different enough that doing well against computer players does not imply doing nearly equally well against human players ... though it certainly helps a lot ... Does it? I've heard many chess instructors say that playing against a computer hinders young players more than it helps. I suspect that's a half-truth... --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Fri, Oct 24, 2008 at 8:48 AM, Ben Goertzel [EMAIL PROTECTED] wrote: I suspect that's a half-truth... Well as a somewhat good chess instructor myself, I have to say I completely agree with it. People who play well against computers rarely rank above first time players.. in fact, most of them tend to not even know the rules of the game.. having had the computer there to coddle them at every move. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: [agi] If your AGI can't learn to play chess it is no AGI
Just now there is a world championship in chess. My chess programs (e.g. Fritz 11) can give a ranking for all moves given an arbitrary chess position. The program agrees with the grandmasters which moves are in the top 5. In most situations it even agrees which move is the best one. Thus, human style chess of top grandmasters and computer chess are quite the same today. - Matthias Von: Ben Goertzel [mailto:[EMAIL PROTECTED] Gesendet: Freitag, 24. Oktober 2008 00:41 An: agi@v2.listbox.com Betreff: Re: [agi] If your AGI can't learn to play chess it is no AGI On Thu, Oct 23, 2008 at 5:38 PM, Trent Waddington [EMAIL PROTECTED] wrote: On Thu, Oct 23, 2008 at 6:11 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I am sure that everyone who learns chess by playing against chess computers and is able to learn good chess playing (which is not sure as also not everyone can learn to be a good mathematician) will be able to be a good chess player against humans. And you're wrong. Trent Yes ... at the moment the styles of human and computer chess players are different enough that doing well against computer players does not imply doing nearly equally well against human players ... though it certainly helps a lot ... ben g _ agi | https://www.listbox.com/member/archive/303/=now Archives https://www.listbox.com/member/archive/rss/303/ | https://www.listbox.com/member/?; 7 Modify Your Subscription http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
Yeah, but these programs did not learn to play via playing other computer players or studying the rules of the game ... they use alpha-beta pruning combined with heuristic evaluation functions carefully crafted by human chess experts ... i.e. they are created based on human knowledge about playing human players... I do think that a sufficiently clever AGI should be able to learn to play chess very well based on just studying the rules. However, it's notable that **either no, or almost no, humans have ever done this** ... so it would require a quite high level of intelligence in this domain... ben g On Thu, Oct 23, 2008 at 7:25 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: Just now there is a world championship in chess. My chess programs (e.g. Fritz 11) can give a ranking for all moves given an arbitrary chess position. The program agrees with the grandmasters which moves are in the top 5. In most situations it even agrees which move is the best one. Thus, human style chess of top grandmasters and computer chess are quite the same today. - Matthias *Von:* Ben Goertzel [mailto:[EMAIL PROTECTED] *Gesendet:* Freitag, 24. Oktober 2008 00:41 *An:* agi@v2.listbox.com *Betreff:* Re: [agi] If your AGI can't learn to play chess it is no AGI On Thu, Oct 23, 2008 at 5:38 PM, Trent Waddington [EMAIL PROTECTED] wrote: On Thu, Oct 23, 2008 at 6:11 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I am sure that everyone who learns chess by playing against chess computers and is able to learn good chess playing (which is not sure as also not everyone can learn to be a good mathematician) will be able to be a good chess player against humans. And you're wrong. Trent Yes ... at the moment the styles of human and computer chess players are different enough that doing well against computer players does not imply doing nearly equally well against human players ... though it certainly helps a lot ... ben g -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/| Modifyhttps://www.listbox.com/member/?;Your Subscription http://www.listbox.com -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/ | Modifyhttps://www.listbox.com/member/?;Your Subscription http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: [agi] If your AGI can't learn to play chess it is no AGI
I am very impressed about the performance of humans in chess compared to computer chess. The computer steps through millions(!) of positions per second. And even if the best chess players say they only evaluate max 3 positions per second I am sure that this cannot be true because there are so many traps in chess which must be considered. I think humans represent chess by a huge number of *visual* patterns. The chessboard is 8x8 squares. Probably, a human considers all 2x2, 3x3 4x4 and even more subsets of the chessboard at once beside the possible moves. We see if a pawn is alone or if a knight is at the edge of the board. We see if the pawns are in a diagonal and much more. I would guess that the human brain observes many thousands of visual patterns in a single position. This is the only explanation for me why the best chess players still have a little chance to win against computers. Even a beginner who never has played chess would see some patterns in the initial position. All pieces with the same color are together at different sides. All pawns of the same color are in the same raw and so on. The interesting question is why the beginner can already see regularities. I think the human has a lot of visual bias which is also useful to see patterns in chess. On the other hand visual embodied experience is of course important too. In my opinion, sophisticated vision is much more important for an artificial human than natural language understanding -Matthias Von: Ben Goertzel [mailto:[EMAIL PROTECTED] Gesendet: Freitag, 24. Oktober 2008 01:53 An: agi@v2.listbox.com Betreff: Re: [agi] If your AGI can't learn to play chess it is no AGI Yeah, but these programs did not learn to play via playing other computer players or studying the rules of the game ... they use alpha-beta pruning combined with heuristic evaluation functions carefully crafted by human chess experts ... i.e. they are created based on human knowledge about playing human players... I do think that a sufficiently clever AGI should be able to learn to play chess very well based on just studying the rules. However, it's notable that **either no, or almost no, humans have ever done this** ... so it would require a quite high level of intelligence in this domain... ben g --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
RE: [agi] If your AGI can't learn to play chess it is no AGI
Within the domain of chess there is everything to know about chess. So if it comes up to be a good chess player learning chess from playing chess must be sufficient. Thus, an AGI which is not able to enhance its abilities in chess from playing chess alone is no AGI. I'm jumping into this conversation a little late, but I think chess is something that should be avoided in the context of AGI. I have three reasons for this, but as far as I can see, it has primarily been the first of these that has been discussed. 1. Intelligence is too easy to fake in chess 2. Chess is too hard to learn from scratch 3. Chess is tainted with the failed ambitions of early GOFAI. The success of Deep Blue and more modern chess playing systems like Fritz and Rybka, with hand-coded search algorithms and heuristics demonstrates how easy it is to fake real intelligence. I'm not a good chess player, but it wouldn't be too hard for me to implement a search algorithm over a simple heuristic evaluation function, resulting in a chess system that can outplay me. In contrast, the seemingly less intelligent problem of walking, is very hard to get right by hand-coding my own knowledge (the nicest walks are almost always discovered by machine learning). That is, in chess it is too easy to fake intelligence by hand coding your own knowledge into the heuristics. Less structured problems, in which expert knowledge is little assistance, are a better challenge for early AGI systems. Chess is also too difficult a problem to learn in a general way from zero knowledge. The difficulty of chess in the general game playing competitions confirms this (last time I heard from one of the teams, even though the best systems do pretty well on simpler games, in chess they can't do much more than simply play legal moves). When playing chess, we have draw on knowledge of space and time and concepts like control and domination. We quickly realize by ourselves that it is good to control the centre of the board, and that the queen is often worth defending, and that even though you can win with just two pieces it is generally bad to lose pieces. But a real AGI would have to discover concepts like center and more powerful by itself (center is a difficult concept to express if you only know about 64 squares and which ones are next to each other). The chess board itself is too large, the moves are too complicated, and the rewards come far far too late to expect a system to automatically discover how to play good chess with no prior knowledge of simpler games or of the larger world. I suspect that the complexity of the problem is such that a system that learns chess without prior knowledge would discover quirky rules that provide local maxima: for example, it might unintentionally learn to sacrifice many of its own pieces because doing so makes the search space smaller, so that the system can think more moves ahead (rather than, say, developing heuristics to simply disregard some of its pieces). And finally, while I did not personally experience the early days of AI, there seems to be an implication in some of the early literature that if only we could create a chess playing robot, then we'll have solved the problem of AI. I think AI has moved on from this simple attitude, and even mentioning chess seems to - at least in my mind - sound like forgetting all the mistakes and lessons from the past. Even with a good argument for resurrecting chess, and an explanation for the past failure of the game to generate real progress in strong AI, I still suspect that mentioning chess is bad marketing for a young field. -Ben --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Fri, Oct 24, 2008 at 10:38 AM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I think humans represent chess by a huge number of *visual* patterns. http://www.eyeway.org/inform/sp-chess.htm Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
Trent: On Fri, Oct 24, 2008 at 10:38 AM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I think humans represent chess by a huge number of *visual* patterns. http://www.eyeway.org/inform/sp-chess.htm We've been over this one several times in the past (perhaps you haven't been here). Blind people can see - they can draw the shapes of objects. . They create their visual shapes out of touch.Touch comes prior to vision in evolution All living creatures are common sense intelligences. IOW the senses are integrated and information is shared between them. It is only at an intellectual level that we can think we can function with only one sense in isolation. It's actually impossible in practice. {See Michael Tye]. - (And there is much, much food for thought in that reality). So yes, Matthias, is correct. How other than visually (and common-sensically) do you think people play chess? P.S. Matthias seems to be cheerfully cutting his own throat here. The idea of a single domain AGI or pre-AGI is a contradiction in terms every which way - not just in terms of domains/subjects or fields, but also sensory domains. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Fri, Oct 24, 2008 at 1:04 PM, Mike Tintner [EMAIL PROTECTED] wrote: We've been over this one several times in the past (perhaps you haven't been here). Blind people can see - they can draw the shapes of objects. . They create their visual shapes out of touch.Touch comes prior to vision in evolution Just cause you've repeated yourself several times doesn't mean you've convinced anyone. If you redefine visual to mean adjacency then maybe you've got a nice workable theory there.. Objects exist in the world and the brain has to have a good model of this.. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] constructivist issues
Mark, I'm saying Godelian completeness/incompleteness can't be easily defined in the context of natural language, so it shouldn't be applied there without providing justification for that application (specifically, unambiguous definitions of provably true and semantically true for natural language). Does that make sense, or am I still confusing? Matthias, I agree with your point in this context, but I think you also mean to imply that Godel's incompleteness theorem isn't of any importance for artificial intelligence, which (probably pretty obviously) I wouldn't agree with. Godel's incompleteness theorem tells us important limitations of the logical approach to AI (and, indeed, any approach that can be implemented on normal computers). It *has* however been overused and abused throughout the years... which is one reason I jumped on Mark... --Abram On Thu, Oct 23, 2008 at 4:07 PM, Mark Waser [EMAIL PROTECTED] wrote: So to sum up, while you think linguistic vagueness comes from Godelian incompleteness, I think Godelian incompleteness can't even be defined in this context, due to linguistic vagueness. OK. Personally, I think that you did a good job of defining Godelian Incompleteness this time but arguably you did it by reference and by building a new semantic structure as you went along. On the other hand, you now seem to be arguing that my thinking that linguistic vagueness comes from Godelian incompleteness is wrong because Godelian incompleteness can't be defined . . . . I'm sort of at a loss as to how to proceed from here. If Godelian Incompleteness can't be defined, then by definition I can't prove anything but you can't disprove anything. This is nicely Escheresque and very Hofstadterian but . . . . - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Thursday, October 23, 2008 11:54 AM Subject: Re: [agi] constructivist issues --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com