Date: Mon 26 Sep 1988 01:59-EDT From: AIList Moderator Nick Papadakis Reply-To: AIList@AI.AI.MIT.EDU Us-Mail: MIT LCS, 545 Tech Square, Rm# NE43-504, Cambridge MA 02139 Phone: (617) 253-6524 Subject: AIList Digest V8 #92 To: AIList@AI.AI.MIT.EDU Status: RO AIList Digest Monday, 26 Sep 1988 Volume 8 : Issue 92 Philosophy: Theft or Honest Toil, Pinker & Prince, learning rules Common sense knowledge of continuous action (2 messages) I got rhythm Commonsense reasoning ---------------------------------------------------------------------- Date: 2 Sep 88 22:35:01 GMT From: mnetor!utzoo!dciem!dretor!client2!mmt@uunet.uu.net (Martin Taylor) Subject: Theft or Honest Toil, (was Re: Pinker & Prince Reply (long version)) Harnad characterizes learning rules from a rule-provider as "theft", whereas obtaining them by evaluation of the statistics of input data is "honest toil". But the analogy is perhaps better in a different domain: learning by evaluating the statistics of the environment is like building up amino acids and other nutritious things from inorganic molecules through photosynthesis, whereas obtaining rules from rule-providers is like eating already built nutritious things. One ofthe great advantages of language is that we CAN take advantage of the regularities discovered in the data by other people. The rules they tell us may be wrong, but to use them is easier than to discover our own rules. It is hardly to be taken as an analogy to "theft". If we look at early child learning, the "theft" question becomes: Has evolution provided us with a set of rules that we do not have to obtain from the data, so that we can later obtain more rules from people who did themselves learn from data? Obviously in some sense the answer is "yes" there are SOME innate rules regarding how we interpret sensory input, even if those rules are as low-level as to indicate how to put together a learning net. Obviously, also, there are MANY rules that we have to get from the data and/or from people who learned them from the data. The question then becomes whether the "rules" regarding past-tense formation are of the innate kind, of the data-induced kind, or of the passed-on kind. My understanding of the developmental literature is that children pass through three phases: (i) correct past-tense formation for those verbs for which the child uses the past tense frequently; (ii) false regularization, in which non-regular past tenses (went) are replaced by regularized ones (goed); (iii) more-or-less correct past tense formation, in which exceptions are properly used, AND novel or neologized verbs are given regular past tenses (in some sense of regular). This sequence suggests to me that the pattern does not have any innate rule component. Initially, all words are separate, in the sense that "went" is a different word from "go". Later, relations among words are made (I will not say "noticed"), and the notion of "go" becomes part of the notion of "went". Furthermore, the notion of a root meaning with tense modification becomes part of verbs in general. Again, I will not say that this is connected with any kind of symbolic rule. It may be the development of net nodes that are activated for root parts and for modifer parts of words. It would be overly rash to claim either that rules are involved or that they are not. In the final stage, the rule-like way of obtaining past tenses is well established enough that the exceptions can be clearly distinguished (whether statistically or otherwise is again disputable). One thing that seems perfectly clear is that humans are in general capable of inducing rules in the sense that some people can verbalize those rules. When such a person "teaches" a rule to a "student", the student must, initially at least, apply it AS a rule. But even in this case, it is not clear that skilled use of what has been learned involves continuing to use the rule AS a rule. It may have served to induce new node structures in a net. In "The Psychology of Reading" (Academic Press, 1983), my wife and I discussed such a sequence under the heading of "Three-phased Learning", which we took to be a fairly general pattern in the learning of skilled behaviour (such as reading). Phase 1 is the learning of large-scale unique patterns. Phase 2 is the discovery of consistent sub-patterns and consistent ways in which the sub-patterns relate to each other (induction or acquisition of rules). Phase 3 is the incorporation of these sub-elements and relational patterns into newly structured global patterns--the acquisition of true skill. "Theft," in Harnad's terms, can occur only as part of Phase 2. Both Phase 1 and Phase 3 involve "honest toil." My feeling is that current connectionist models are mainly appropriate to Phase 1, and that symbolic approaches are mainly appropriate to Phase 2, though there is necessarily overlap. There should not be a contention among models using one or other approach, if this is so. They are both correct, but under different circumstances. -- Martin Taylor DCIEM, Box 2000, Downsview, Ontario, Canada M3M 3B9 uunet!mnetor!dciem!client1!mmt or mmt@zorac.arpa (416) 635-2048 ------------------------------ Date: 18 Sep 88 1543 PDT From: John McCarthy Subject: common sense knowledge of continuous action If Genesereth and Nilsson didn't give an example to illustrate why differential equations aren't enough, they should have. The example I like to give when I lecture is that of spilling the water glass on the lectern. If the front row is very close, it might get wet, but usually not even that. The Navier-Stokes equations govern the flow of the spilled water but are entirely useless in this common sense situation. No-one can acquire the initial conditions or integrate the equations sufficiently rapidly. Moreover, absorption of water by the materials it flows over is probably a strong enough effect, so that more than the Navier-Stokes equations would be necessary. Thus there is no "scientific theory" involving differential equations, queuing theory, etc. that can be used by a robot to determine what can be expected when a glass of water is spilled, given what information is actually available to an observer. To use the terminology of my 1969 paper with Pat Hayes, the differential equations don't form an epistemologically adequate model of the phenomenon, i.e. a model that uses the information actually available. While some people are interested in modelling human performance as an aspect of psychology, my interest is artificial intelligence. There is no conflict with science. What we need is a scientific theory that can use the information available to a robot with human opportunities to observe and do as well as a human in predicting what will happen. Thus our goal is a scientific common sense. The Navier-Stokes equations are important in (1) the design of airplane wings, (2) in the derivation of general inequalities, some of which might even be translatable into terms common sense can use. For example, the Bernoulli effect, once a person has (usually with difficulty) integrated it into his common sense knowledge can be useful for qualitatively predicting the effects of winds flowing over a house. Finally, the Navier Stokes equations are imbedded in a framework of common sense knowledge and reasoning that determine the conditions under which they are applied to the design of airplane wings, etc. ------------------------------ Date: 19 Sep 88 01:18:29 GMT From: garth!smryan@unix.sri.com (Steven Ryan) Subject: Re: state and change/continuous actions >Foundations of Artificial Intelligence," I find it interesting to >compare and contrast the concepts described in Chapter 11 - "State >and Change" with state/change concepts defined within systems >theory and simulation modeling. The authors make the following statement: >"Insufficient attention has been paid to the problem of continuous >actions." Now, a question that immediately comes to mind is "What problem?" Presumably, they are referring to that formal systems are strictly discrete and finite. This has to do to with `effective computation.' Discrete systems can be explained in such simple terms that is always clear exactly what is being done. Continuous systems are computably using calculus, but is this `effective computation?' Calculus uses a number of existent theorems which prove some point or set exists, but provide no method to effectively compute the value. Or is knowing the value exists sufficient because, after all, we can map the real line into a bounded interval which can be traversed in finite time? It is not clear that all natural phenomon can be modelled on the discrete and finite digital computer. If not, what computer could we use? >Any thoughts? ------------------------------ Date: 19 Sep 88 01:18:45 GMT From: dscatl!mgresham@gatech.edu (Mark Gresham) Subject: I got rhythm In a recent article PGOETZ@LOYVAX.BITNET writes: >Here's a question for anybody: Why do we have rhythm? > >Picture yourself tapping your foot to the tune of the latest Top 40 trash hit. >While you do this, your brain is busy processing sensory inputs, controlling >the muscles in your foot, and thinking about whatever you think about when >you listen to Top 40 music. >[...text deleted...] >It comes down to this: Different actions require different processing >overhead. So why, no matter what we do, do we perceive time as a constant? The fact is, we *don't*. (Take it from a musician!) Generally people have a quite erratic perception of time. Th perception (the top 40 example) is one of constancy in relationship to some other perceived event be believe to be constant (or assume is so). Hence, the "beats" in the music (which we deem to be regular) are giving us fresh input which we use to "correct" our foot tapping. >Why do we, in fact, have rhythm? Do we have an internal clock, or a >"main loop" which takes a constant time to run? Or do we have an inadequate >view of consciousness when we see it as a program? > >Phil Goetz >PGOETZ@LOYVAX.bitnet Try this experiment. Or several of you try it. Take a stopwatch (digital is preferable because silent). Don't look at it or any other clock, and don't count; press the start button. Then, when you think five minutes are up, stop it. Look at the watch and see how you did. I know of one percussionist who is said to be quite accurate. If you are really concentrating on "the passage of time" --genuinely trying to be aware of it--my guess is that you'll start to sweat (or otherwise become uncomfortable) after about 40 seconds or so. It takes quite a bit of discipline to empty your mind enough to successfully do that. Try it. Invent other similar experiments. Let me know what you discover. --Mark Gresham (please e-mail or post to rec.music.classical) ++++++++++++++++++++++++++++++++++++++++++ Mark Gresham Atlanta, GA, USA UUCP: ...!gatech!dscatl!mgresham INTERNET: mgresham@dscatl.UUCP ++++++++++++++++++++++++++++++++++++++++++ ------------------------------ Date: 19 Sep 88 15:20:13 GMT From: fishwick@bikini.cis.ufl.edu (Paul Fishwick) Subject: commonsense reasoning I very much appreciate Prof. McCarthy's response and would like to comment. The "water glass on the lectern" example is a good one for commonsense reasoning; however, let's further examine this scenario. First, if we wanted a highly accurate model of water flow then we would probably use flow equations (such as the NS equations) possibly combined with projectile modeling. Note also that a lumped model of the detailed math model may reduce complexity and provide an answer for us. We have not seen specific work in this area since spilt water in a room is of little scientific value to most researchers. Please note that I am not trying to be facetious -- I am just trying to point out that *if* the goal is "to solve the problem of predicting the result of continuous actions" then math models (and not commonsense models) are the method of choice. Note that the math model need not be limited to a single set of PDE's. Also, the math model can be an abstract "lumped model" with less complexity. The general method of simulation incorporates combined continuous and discrete methods to solve all kinds of physical problems. For instance, one needs to use notions of probability (that a water will make it to the front row), simplified flow equations, and projectile motion. Also, solving of the "problem of what happens to the water" need not involve flow equations. Witness, for instance, the work of Toffoli and Wolfram where cellular automata may be used "as an alternative to" differential equations. Also, the problem may be solved using visual pattern matching - it is quite likely that humans "reason" about "what will happen" to spilt liquids using associative database methods (the neural netlanders might like this approach) based on a huge library of partial images from previous experience (note Kosslyn's work). I still haven't mentioned anything about artificial intelligence yet - just methods of problem solving. I agree that differential equations by themselves do not comprise an epistemologically adequate model. But note that no complex problem is solved using only one model language (such as DE's). The use of simulation is a nice example since, in simulating a complex system, one might use many "languages" to solve the problem. Therefore, I'm not sure that epistemological adequacy is the issue. The issue is, instead, to solve the problem by whatever methods available. Now, back to AI. I agree that "there is no theory involving DE's (etc.) that can be used by a robot to determine what can be expected when a glass of water is spilled." I would like to take the stronger position that searching for such a singular theory seems futile. Certainly, robots of the future will need to reason about the world and about moving liquids; however, we can program robots to use pattern matching and whatever else is necesssary to "solve the problem." I supposed that I am predisposed to an engineering philosophy that would suggest research into a method to allow robots to perform pattern recognition and equation solving to answer questions about the real world. I see no evidence of a specific theory that will represent the "intelligence" of the robot. I see only a plethora of problem solving tools that can be used to make future robots more and more adaptive to their environments. If commonsense theories are to be useful then they must be validated. Against what? Well, these theories could be used to build programs that can be placed inside working robots. Those robots that performed better (according to some statistical criterion) would validate respective theories used to program them. One must either 1) validate against real world data [the cornerstone to the method of computer simulation] , or 2) improved performance. Do commonsense theories have anything to say about these two "yardsticks?" Note that there are many AI research efforts that have addressed validation - expert systems such as MYCIN correctly answered "more and more" diagnoses as the program was improved. The yardstick for MYCIN is therefore a statistical measure of validity. My hat is off to the MYCIN team for proving the efficacy of their methods. Expert systems are indeed a success. Chess programs have a simple yardstick - their USCF or FIDE rating. This concentration of yardsticks and method of validation is not only helpful, it is essential to demonstrate the an AI method is useful. -paul +------------------------------------------------------------------------+ | Prof. Paul A. Fishwick.... INTERNET: fishwick@bikini.cis.ufl.edu | | Dept. of Computer Science. UUCP: gatech!uflorida!fishwick | | Univ. of Florida.......... PHONE: (904)-335-8036 | | Bldg. CSE, Room 301....... FAX is available | | Gainesville, FL 32611..... | +------------------------------------------------------------------------+ ------------------------------ End of AIList Digest ********************