John, Background Note: My first full-time job was doing CAD for the first Boeing 737s - the first planes to be substantially designed by CAD. There was little CAD software back then, so we had to create it ourselves. Starting from wind tunnel data, we had to figure out how to place supports into wings and bodies to be able to achieve the desired shape by bending flat sheet metal over the supports. Note that cars utilize stamped metal, but airplanes are entirely covered by bent sheets of metal.
Aircraft designers call this "the spline fit problem", which is an analog (vs. digital) version of the interpolation/extrapolation problem. The problem appears (to me) to boil down to much the same as the pattern problem. There is some profound physics behind those long list of coordinates from the wind tunnel, along with some noise due to the finite amount of testing and optimization occurring there, measurement errors, etc. We had to find simultaneous solutions between the "understanding" of shape from the windtunnel and the "understanding" of what it took to achieve a shape with bent sheets of metal. I came in late in this process, which started with a "cooks tour" of past failed efforts by the mathematicians who preceded me. They had (not unexpectedly) run into the multiple solution problem, and had to strike a balance between the presumed noise and how radical allowed solutions were. Sure they could exactly fit the wind tunnel points, but it would take a lot of supports and the skin would have some nasty ripples at points between the coordinates, just as paper would have to be wrinkled to fit over misaligned supports. Just like in the rest of the real world, no one knew how much of the "data" was real and how much was noise until you tried to model with it, and even then some of the presumed noise could in fact be imperfections in your model, e.g. relativistic errors in Newtonian models. Reading your comments somewhat through this lens... On 5/16/08, John G. Rose <[EMAIL PROTECTED]> wrote: > > It's not just – oh now I recognize a pattern and can predict the next few > items. It's oh I see and understand numerous patterns all intermeshed and > have a rich knowledge base of raw and processed pattern data...and my > operators are patterns, etc.. > This is ONLY tractable when you understand substantially everything (as in aircraft design). Where there remains a significant source of "noise" (read that: More than one unrecognized phenomena), this goes nowhere. Further, looking at the difference between what you do understand and what you do observe only helps if the entire remainder is SO simple that you can learn from this observation. Most stuff in the universe is patterns, but all intertwined and mixed > together. > Agreed. Note: pattern is an overloaded term- there are several definitions. > Which is probably where most of our communications problems arise. I am obviously new to this discussion. Any help to bring me up to speed would be GREATLY appreciated. Thanks. Steve Richfield > *From:* Steve Richfield [mailto:[EMAIL PROTECTED] > *Sent:* Friday, May 16, 2008 9:38 PM > *To:* [email protected] > *Subject:* [agi] Pattern extrapolation as a method requiring limited > intelligence > > > > Hi y'all, > > > > I can still remember an incident when I was just 13 years old. I was called > into the Principal's office and asked about my poor score on an IQ test. You > see, I was the school genius with off-scale IQ, yet I had barely tested > average=100. > > > > The test had been a hundred or so patterns of numbers with me having to > pick the next number from a multiple-choice list. However, I could often see > reasonable explanations to support more than one choice, and sometimes > different explanations to support ALL choices. I had asked about this at the > time of the test, and was told to "pick the answer with the simplest > explanation". I protested that they ALL had simple explanations, to which I > was instructed to "just do the best that you can". > > > > I suggested that if the principal would just produce some of the test > questions, that I would gladly show him how there were many simple > answers. I suggested that the IQ test was really more of a measure of the > person creating the test rather than of the people taking the test. The > principal just shook his head and sent me back to class. > > > > Now, decades later, come the present discussions about patterns, apparently > advanced along with the same lines of "thought" that was behind that IQ test > so many years ago. Pattern recognition without underlying supporting theory > is WORTHLESS (or perhaps worse) except perhaps to suggest possible > underlying theory. Any good AGI would see a limitless list of possible next > items from any real-world list. > > > > Take a sequence of alternating musical notes. Does anyone here REALLY think > that someone is going to sit in Carnegie Hall and continuously play two > alternating notes just because he started out that way? Note that Fur Elise > does indeed start out that way. No, at some point, the odds of continuing > with the same two notes drops quite low. A good music composition program > might theorize the best sounding subsequent notes, and THOSE would be much > more likely than continuing with the same mindless repeating pattern. > > > > No, I am NOT saying that pattern identification is worthless, just that > when used to predict rather than understand an underlying process, that it > may well be of negative value. In any case, I really can't see any value to > AGI or other efforts from success here. > > > > Note the parallels with unsupervised neural networks, which seek to do much > the same. However, unsupervised NNs typically look for VARIATIONS from past > patterns rather extrapolating from them. I think that this "small" > difference is crucial. > > > > Have I missed something important here? > > > > Steve Richfield > > > > > ------------------------------ > > *agi* | Archives <http://www.listbox.com/member/archive/303/=now> > <http://www.listbox.com/member/archive/rss/303/>| > Modify<http://www.listbox.com/member/?&;>Your Subscription > > <http://www.listbox.com/> > > > ------------------------------ > *agi* | Archives <http://www.listbox.com/member/archive/303/=now> > <http://www.listbox.com/member/archive/rss/303/> | > Modify<http://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=8660244&id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
