Author: Dann Corbit
Date: 19:01:30 10/16/03
Go up one level in this thread
On October 16, 2003 at 20:48:23, Anthony Cozzie wrote: >On October 16, 2003 at 19:11:20, Dann Corbit wrote: > >>On October 16, 2003 at 18:49:55, Anthony Cozzie wrote: >> >>>On October 16, 2003 at 18:07:08, Dann Corbit wrote: >>> >>>>On October 16, 2003 at 15:25:43, Steven Edwards wrote: >>>> >>>>>On October 16, 2003 at 09:20:20, Robert Hyatt wrote: >>>>>>On October 16, 2003 at 09:06:17, swaminathan natarajan wrote: >>>>> >>>>>>>about 900 n/s >>>>>> >>>>>>It had better be faster. IE a single xeon runs over 1M nodes >>>>>>per second. >>>>> >>>>>How far we have come! >>>>> >>>>>I seem to recall Slate and Atkin reporting that their program Chess 4.5 ranged >>>>>between 250 and 600 Hz on a CDC 6400 (roughly equivalent to an Intel 33 HMz >>>>>80386+80387), and this was enough to give some humans a decent challenge (back >>>>>in the mid 1970s) along with winning the world CC championship. >>>>> >>>>>Processing speed has increased by a factor of forty or so in the past three >>>>>decades. Are the programs/platfrom combinations of 2003 much more than forty >>>>>times "better" than that of 1973? How much of the "better" ratio is due to >>>>>improvements in algorithms? >>>>> >>>>>More specifically, if one were to take Crafty or a similar program that has the >>>>>NWU Chess 4.x as a great grand uncle and run it on a 33 HMz 80386+80387 class >>>>>machine, how would it fare against Chess 4.x running on a true clock speed >>>>>emulation of CDC 6400 hardware? (The last real CDC 6400 was powered off long >>>>>ago, perhaps in the mid 1980s if I remember correctly.) >>>> >>>>I suspect that in a 100 game match, Crafty would win 100 to zero. We could >>>>reverse the question. Take the program of long ago and compile it with modern >>>>compilers. Now try the experiment on really fast hardware. That is a more >>>>important question to me. I don't care how crafty would perform on a 386 >>>>because I have no intention of running it on a 386 at any time or for any >>>>reason. >>>> >>>>>I assume that the more modern program would win most of the time, but it >>>>>wouldn't be that much of a performance mismatch. If today's programs on today's >>>>>hardware are 1000 Elo stronger than the 1973 CC champ, how much of that is due >>>>>to better algorithms vs better hardware? I'll take a guess and say that thirty >>>>>years of advances in software is responsible for no more than 200 Elo >>>>>improvement and perhaps only 150 Elo points. And most of the software >>>>>improvement is due to only a few new ideas: >>>>> >>>>> 1. PVS/zero width search >>>>> 2. Null move subtree reduction >>>>> 3. History move ordering heuristics >>>>Insignificant >>>> >>>>> 4. Tablebase access during search >>>>Insignificant >>>> >>>>> 5. Automated tuning of evaluation coefficients >>>>Less than insignificant. Nobody has ever exceeded the hand tuned values. Right >>>>now, if you do this, it will make your program play badly. I also suspect that >>>>the Deep Blue team harmed their chess engine with this approach. >>>> >>>>This one is the most important: >>>>#0. Hash tables and move ordering >>>> >>>>Without this, you won't achieve #0: >>>>#1. Better evaluation >>>> >>>>>Computer chess was supposed to be the Drosephilia of AI. If so, CC theory is >>>>>still in the larval stage and I don't see wing buds popping out any time soon. >>>>>Where are the CC planning engines? Where are any general pattern recognition >>>>>algorithms in use? >>>> >>>>Because the hand-tuned algorithms are superior. >>>> >>>>>What program has real machine learning? >>>> >>>>Lots of them. Unless you mean genetic style evolution of strength or neural >>>>nets. Both of those have been tried and are flops (as of this date and for >>>>those attempts that have been published). >>>> >>>>>Which programs are >>>>>adaptive and can re-write better versions of themselves? >>>> >>>>Octavius springs to mind. It's a wimp. >>>> >>>>> How many programs can >>>>>converse in natural language and answer the simplest of questions as to why a >>>>>particular move was made? >>>> >>>>That is 10 years off in the future. >>>> >>>>> Where are the programs that can improve based on >>>>>taking advice vs coding patches to the Evaluate() function? >>>> >>>>There are none of those. Nimzo's programming approach could be considered >>>>similar to this, except that the language is typed and not spoken. He uses a >>>>metalanguage that describes chess (IIRC). >>>> >>>>>And the big question: What has CC done for AI in the past thirty years, and what >>>>>can it do for AI in the next thirty years? >>>> >>>>The Deep Blue chess match is the most famous chess match of all time. The >>>>strongest human player was beaten in a game of exponential complexity. >>>> >>>>It is not a good idea to try to predict the future. Even mathematically >>>>speaking and when you have a lot of data points, extrapolations are very >>>>dangerous. >>>> >>>>>Hint: Any remotely correct answer does not include the phrase "nodes per >>>>>second". >>>> >>>>I disagree. Hans Moravec's book shows that in 30 years, our CPU's will be >>>>smarter than we are. And why is that? Not due to superior algorithms, but >>>>strictly due to Moore's law. >>>>http://www.frc.ri.cmu.edu/~hpm/talks/revo.slides/2030.html >>> >>> >>>I have some serious problems with that slide. >>> >>>1. Moore's law is NOT A LAW. Its going to come to an end by 2020, if not >>>earlier. >> >>Not a chance. It will continue to accelerate. Of course, I could be wrong. > >OK, this is simply wrong. Moore's law postulates continued exponential growth >in integrated circuit transistor density. The "generic" version simply says that compute power doubles every 18 months. >[http://www.intel.com/research/silicon/mooreslaw.htm] Clearly it is impossible >to make a transistor with less than 3 atoms, So it will require some other form of technology. I also do not admit that it is impossible to create a transistor with less than three atoms. Perhaps it is possible to create a quintillion transistors with a single electron. Just because we do not know how to do it does not render it impossible. Perhaps a single electron has a structural complexity greater than how we would currently desribe the known universe. >so it cannot continue forever. I do not accept your first position. Even if I did accept that position (that there is a limiting size to computational density) then I still would not accept your second position (that speed cannot continue to double). >If >I remember my quantum theory correctly, an atom several A in size, so we are >getting close to the end of Moore's law. Already CMOS is on its last legs. >Maxwell's laws have caught up to it - the only thing intel or amd can do with >all the transistors their process guys have been giving them is build bigger >caches. Perhaps we won't compute with transistors any more. Or perhaps we will build 3-dimentional IC's instead of planar ones. That would be a stupendous increase in density. Or perhaps it will be a different technology altogether. > There is a lot of research going on here, but CMOS is still not going >to take us past 2020 (in terms of continuing to shrink). Unless there is some discovery that will enable it to do so. >>>2. According to his graph, a 486/DX2 is equal in intelligence to a spider. I >>>think not. Even the simplest biological organisms have motor control that is >>>better than anything we can do today. >> >>Check out Asimo. >> >>There was also a show I saw where a German autonomous helicopter flew to a scene >>where mock-up accidents occurred. It correctly identified all of the problems. > >Could you post a link? Don't know if there are any links to it. I saw it on the Science Channel. >>>Its pattern recognition is far ahead of >>>the best we can do. >> >>That's because it uses a neural net. Neural nets are designed for pattern >>recognition. > >We don't know what it uses, really :) Not to mention that noone really >understands neural nets. If you build a neural net that recognizes a pattern, >you really have no idea what is going on (other than that the neural net works: >you have no idea *how* or *why* it works). But my point is that, while >computers are very good at certain things, there is much work to be done in >other areas. > >>>we have a long way to go in terms of signal processing before we can >>>even do simple things, let alone reason abstractly as a human does. >> >>I think they are already accomplishing this. >>http://www.ifi.ntnu.no/grupper/ai/eval/robot_links.html >> >>>Will computers ever achieve human like intelligence? I'm certainly not going to >>>state that they aren't. >> >>I am quite sure that they will exceed it. In 1000 years, human intelligence >>will look like a spider compared to the computer. > >Well, I'm not even going to touch that one. I have no idea if humans or >computers will even exist in 1000 years, and realistically neither does anyone >else. I'm more concerned with my lifetime. > >>> Quantum computers in particular are _very_ exciting. >>>But 2020 (as his slide states) is far to soon. >> >>The slide says in HUGE LETTERS 2030. 2020 is on the graph about 'monkey' level. > >Relax, no need to whip out the caps-lock ;) The graph only goes up to 2020 so I >rashly assumed that was his intersection. > >>>I think even 2030 is too soon. >>>If ever computers surpass humans, they definitely won't be Von Neuman machines. >> >>I think it is unlikely to predict what kind of machines they will or won't be. > >We humans are remarkably bad at predicting the future. Perhaps machines will do >it better :) Maybe instead of predicting it, it will be a case of generation.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.