levesque@csli.stanford.edu (Hector Levesque) (06/19/91)
I've never posted to comp.ai before (and may regret it!), so please forgive violations of protocol. But these attacks on logic and truth, though maybe familiar in AI from about 15 years ago, deserve some comment. First, a minor complaint about Minsky's post about logic. I agree completely that universal generalization could end up playing a very minor role in our cognitive life. But I disagree completely that the utility of logic is somehow thereby compromised. One very popular logical theory, sometimes called "quantification theory" is indeed concerned with expressing in logical terms the properties of "for all" and "for some." But even existing logical theories go well beyond this. The theory of generalized quantifiers, for example, examines properties of quantifiers like "many", "most," "almost all" and the like that stand to play a much more important role in expressing what we believe. Then there are the statistical/probabilistic accounts (a la Bacchus/Halpern), the nonmonotinic accounts etc. To say that logic lives or dies with generalization and instantiation is like saying it lives or dies with exclusive-or. Another point: I think it is a simple mistake (of logic!) to conclude that because we can never be *certain* about what we mean when we say something, or what we are agreeing about, or what is true, that somehow the truth of the matter is thereby open to negotiation or interpretation, or that we can decide to act in a way that does not take it into account. If I tell you "there's a truck coming towards you from behind", I may have no way of knowing for sure that my statement is correct, and you may have no way of being sure either of what I'm getting at or (assuming you've figured it out) of whether or not what I am saying is true. But it's a mistake (and a dangerous one) to conclude from this lack of certainty that the truck issue is somehow thereby reduced in importance, or that what ultimately matters is your goals and desires, or our linguistic conventions, or even that one opinion on the issue is as good as another. None of these follow from admitting that we may never know for sure one way or another if there is a truck. A skeptic may choose to focus on what I said, question what I mean by a "truck" (a toy truck?), or just observe the loaded context dependency and unavoidable subjectivism in how I perceive and report things. But if he or she after all this doesn't get it, and does not come to appreciate very clearly the relevant issue, all is for nought, and the world will do the rest. You don't have to *know* what I said, and you don't have to *know* if what I said is true, but for your own safety and comfort, you'd better be able to figure out what it would be like for it to be true. This, I assume, is what logic is for, at least for AI purposes. Focussing on Truth in some abstract, all-or-nothing, eternal, godlike sense, is a bit of a red herring. What matters I think in AI is being able to explore the consequences of things being one way and not another, even while admitting that much of our view of the world is not going to be right (even by our own terms), and that there is no way to achieve certainty about almost all of it. We need to be able to ask ourselves "according to what I now believe, what would things be like if P?" The fact that we first use natural language typically to express a P, and that this language is infinitely rich and open to endless interpretation and uneliminable context dependency and bla-bla-bla should really not fool us into thinking that there is no issue to settle regarding the way the world is. To fall for this, as far as I can see, is to undervalue the difference between being right or wrong about the truck, for example, and to guarantee for oneself a hermeneutically rich but very short life. The fact is, I don't think anyone takes this position too seriously except when assuming a philosophical stance. Show me a philosohper that doesn't fall into realism of the most naive sort when confronted with memos that say "I'm sorry, but your salary will be reduced by 50%." Under the right circumstances (having nothing to do with mathematics or formal artificial domains!) relativism is put on hold, and the ordinary objective truth about what might appear to be hopelessly vague imponderables suddenly becomes very crisp, very precise, and very relevant to action. Hector Levesque
ISSSSM@NUSVM.BITNET (Stephen Smoliar) (06/19/91)
In article <20018@csli.Stanford.EDU> levesque@csli.stanford.edu (Hector Levesque) writes: > I think it is a simple mistake (of logic!) to conclude that >because we can never be *certain* about what we mean when we say something, >or what we are agreeing about, or what is true, that somehow the truth of >the matter is thereby open to negotiation or interpretation, or that we can >decide to act in a way that does not take it into account. If I tell you >"there's a truck coming towards you from behind", I may have no way of >knowing for sure that my statement is correct, and you may have no way of >being sure either of what I'm getting at or (assuming you've figured it >out) of whether or not what I am saying is true. But it's a mistake (and a >dangerous one) to conclude from this lack of certainty that the truck issue >is somehow thereby reduced in importance, or that what ultimately matters >is your goals and desires, or our linguistic conventions, or even that one >opinion on the issue is as good as another. None of these follow from >admitting that we may never know for sure one way or another if there is a >truck. A skeptic may choose to focus on what I said, question what I mean >by a "truck" (a toy truck?), or just observe the loaded context dependency >and unavoidable subjectivism in how I perceive and report things. But if >he or she after all this doesn't get it, and does not come to appreciate >very clearly the relevant issue, all is for nought, and the world will do >the rest. You don't have to *know* what I said, and you don't have to >*know* if what I said is true, but for your own safety and comfort, you'd >better be able to figure out what it would be like for it to be true. > I think the REAL mistake in this argument is the attempt to pile too much on the shoulders of logic. When you are standing out there in the world, the issue is not a matter of truth, certainty, or even "what it would be like for it to be true." The issue is far simpler: What do you do when someone says "there's a truck coming towards you from behind?" At the risk of attaching too much importance to Skinner (who has no more claim to having all the answers than the logicians do) the answer to this question, in its simplest terms, is that you BEHAVE. In a situation as urgent as this one, anything you are likely to call reasoning will not take place until AFTER you have behaved and you are reflecting on what just happened (perhaps while choking on the exhaust fumes). Thus, I think Hector's example is a good illustration of the danger of confusing the EXPLANATORY value of logic with any PREDICTIVE value--a point which I recently raised in comp.ai.philosophy. >This, I assume, is what logic is for, at least for AI purposes. Focussing >on Truth in some abstract, all-or-nothing, eternal, godlike sense, is a bit >of a red herring. What matters I think in AI is being able to explore the >consequences of things being one way and not another, even while admitting >that much of our view of the world is not going to be right (even by our >own terms), and that there is no way to achieve certainty about almost all >of it. We need to be able to ask ourselves "according to what I now >believe, what would things be like if P?" The fact that we first use >natural language typically to express a P, and that this language is >infinitely rich and open to endless interpretation and uneliminable context >dependency and bla-bla-bla should really not fool us into thinking that >there is no issue to settle regarding the way the world is. To fall for >this, as far as I can see, is to undervalue the difference between being >right or wrong about the truck, for example, and to guarantee for oneself a >hermeneutically rich but very short life. > I think it is certainly true that we do "reason" (in that same sense of the word which I was arguing about above) about hypothetical situations. Indeed, our ability to do so is one of the reasons why Skinner does not have all the answers. Nevertheless, there is no reason to believe that any machinery which we engage to ponder hypotheticals (which we tend to be free to do only when any other demands of the situation are relatively low) is the SAME machinery which exercises control over our behavior in the here-and-now. Such uniformity would be architecturally elegant, but elegance cannot hold a candle to more fundamental issues of survival such as those Chris Malcolm recently posed on comp.ai.philosophy. =============================================================================== Stephen W. Smoliar Institute of Systems Science National University of Singapore Heng Mui Keng Terrace, Kent Ridge SINGAPORE 0511 BITNET: ISSSSM@NUSVM "He was of Lord Essex's opinion, 'rather to go an hundred miles to speak with one wise man, than five miles to see a fair town.'"--Boswell on Johnson
thomas@ckgp.UUCP (Michael Thomas) (06/20/91)
In article <9106190527.AA17403@lilac.berkeley.edu>, ISSSSM@NUSVM.BITNET (Stephen Smoliar) writes: > In article <20018@csli.Stanford.EDU> levesque@csli.stanford.edu (Hector > Levesque) writes: > > I think it is a simple mistake (of logic!) to conclude that > >because we can never be *certain* about what we mean when we say something, > >or what we are agreeing about, or what is true, that somehow the truth of > >the matter is thereby open to negotiation or interpretation, or that we can > >decide to act in a way that does not take it into account. If I tell you > >"there's a truck coming towards you from behind", I may have no way of > >knowing for sure that my statement is correct, and you may have no way of Despite what we learned in school, the TRUTH or what the REALITY is is something that you create (or the AI must create) for itself. You must deturmine what is true and what is not. If you said a truck is coming at me then tha brain/mind would instantly become aware or look for the sound sight or other stimulus that would lead to personal truth. > The issue is far simpler: What do you do when someone > says "there's a truck coming towards you from behind?" At the risk of > attaching too much importance to Skinner (who has no more claim to having > all the answers than the logicians do) the answer to this question, in its > simplest terms, is that you BEHAVE. In a situation as urgent as this one, The person would also deturmine the reliable state of the speaker... (you know the story the boy who cried wolf?) Yes teh first two times it works but then died out. this was a story, in REAL life what would happen if the stimulus wasn't there (actually I have never heared of anyone getting hit by a truck or car? Just parents saying that kids will if they go NEAR the street.) try it just say to someone: "LOOK OUT THERE IS A TRUCK COMING!" see what they do? first if your not in the middle of the street it wont work... if you are I bet they will turn around first! > >This, I assume, is what logic is for, at least for AI purposes. Focussing > >on Truth in some abstract, all-or-nothing, eternal, godlike sense, is a bit > >of a red herring. What matters I think in AI is being able to explore the > >consequences of things being one way and not another, even while admitting I'm sure you can see that the AI must use some clues from the world/environment to aid its own views on the TRUTH in the world around us. Isn't this what we do, trust our senses? "Some things have to be believed to be seen." --Ralph Hodgen Thanks for listening... -- Thank you, Michael Thomas (..uunet!ckgp!thomas)
byland@iris.cis.ohio-state.edu (Tom Bylander) (06/26/91)
In article <9106190527.AA17403@lilac.berkeley.edu> ISSSSM@NUSVM.BITNET (Stephen Smoliar) writes: >When you are standing out there in the world, the >issue is not a matter of truth, certainty, or even "what it would be like >for it to be true." The issue is far simpler: What do you do when someone >says "there's a truck coming towards you from behind?" >[T]he answer to this question, in its >simplest terms, is that you BEHAVE. In a situation as urgent as this one, >anything you are likely to call reasoning will not take place until AFTER >you have behaved and you are reflecting on what just happened (perhaps while >choking on the exhaust fumes). I think there is a couple (common) confusions here. First, there is the confusion of equating reasoning with deliberative, conscious behavior. If I accidently touch a hot surface, and then involuntarily flinch, have I made no inferences at all? To the contrary, one answer is that I (or the relevant part of my nervous system) "perceived" that I am touching a hot surface, and that I have inferred that I should move away from it quickly. The fact that the transformation from hotness to flinching is not deliberative/conscious does not imply that no reasoning has occurred, or that no "truths" have been represented [too many negatives, I know]. Second, there is the confusion between languages of analysis and the phenomena being analyzed. Using logic to analyze some reasoning does not imply that the reasoning itself explicitly uses the rules of logic. For example, computational learning theory uses statistics and computational complexity to analyze inductive learning algorithms. However, the algorithms themselves do not apply the rules of statistics or computational complexity. Similarly, an algorithm can be analyzed using logic without any requirement that the algorithm explicitly use resolution, quantifiers, etc. (Note: to avoid logical analysis, you will have to avoid, among other things, doing any programming at all!) The bottom line is that any argument of the sort "logic is bad because we don't explicitly use it" is a non-starter. >Thus, I think Hector's example is a good >illustration of the danger of confusing the EXPLANATORY value of logic with >any PREDICTIVE value--a point which I recently raised in comp.ai.philosophy. It is very mysterious to me how you are going to make any predictions without inferring them from some initial situation, i.e., without doing logic. I should mention that I do not believe that logic is going to solve all the world's problems. As many articles have noted, there are lots of problems with logic. However, just because logic has some problems doesn't mean that logic is dispensable. Whether we like it or not, modus ponens is still something we will have to take into account. Tom Bylander byland@cis.ohio-state.edu