Author: Bob Durrett
Date: 08:34:11 11/03/02
Go up one level in this thread
On November 03, 2002 at 09:14:09, andrew tanner wrote: >On November 02, 2002 at 22:45:21, Bob Durrett wrote: > >>On November 02, 2002 at 22:25:37, andrew tanner wrote: >> >>>On November 02, 2002 at 00:06:08, Robert Hyatt wrote: >>> >>>>On November 01, 2002 at 22:52:14, Bob Durrett wrote: >>>> >>>>>On October 31, 2002 at 20:01:10, Robert Hyatt wrote: >>>>> >>>>>>On October 31, 2002 at 17:00:19, Bob Durrett wrote: >>>>>> >>>>>>> >>>>>>>Solving the general problem of emulating the chess play of "humanity" might be a >>>>>>>prohibitively difficult task. >>>>>>> >>>>>> >>>>>>This has been the "holy grail" of AI since its early days. But the problem is, >>>>>>in 25 words or less "we have no idea how a person does what he does when playing >>>>>>chess (or anything else for that matter), which makes it _impossible_ to emulate >>>>>>what we don't understand." >>>>> >>>>>Well, Bob H., emulating the chess play of a human is not exactly what the AI >>>>>people want to do, is it. They wish to make a carbon copy of a human in all >>>>>it's gory details. >>>>> >>>>>Many orders of magnitude different, I would say. >>>>> >>>>>Bob D. >>>> >>>>They really want to emulate human thought processes related to chess, >>>>at least for the computer chess/AI purists. But until we know how the >>>>human does what he does, emulation is futile, to paraphrase the borg. >>>> >>> >>>Wouldn't it be possible to emulate human thought by having the program learn how >>>to play chess from scratch, just as people do? >> >>What would make it learn how to be human rather than something else entirely? >> >>Bob D. > > > > It would learn to play "human-like" from calculations (analyses) of the >stored games of the past Excluding stored computer games of the past, I presume. >combined with it's own active learning function. This >is human-like rather than machine-like in that it is a complete learning >approach from start to finish without short-cuts (hash-tables) (opening books) >(tablebase code)(pre-defined evaluation parameters) (null move search). Your idea prompts another thought: "Would your machine learn to make human mistakes and othewise play imperfect chess?" For example, if you were to feed it only stored games of the past played by ELO 1800 players, would it eventually learn to play at the ELO 1800 level, and to consistently make the types of errors it found in the game collection? Bob D. > emulate: To strive to be equal or excel(webster dictionary) > > >I think mimic would be a better term to use here. I really don't think that >researchers are trying to emulate human thought-that just seems to me to be a >ridiculous notion. Rather, they are trying to mimic a learning process. I guess you would have to be more specific as to which researchers you are referring to. One might consider the collection of all serious chess programmers to be a "research group" of sorts. They certainly do seem to be striving to produce chess engines which excel above the human level. Upon reflection about what you're saying, I now wonder whether or not there exist effective learning processes which are inherently non-human. Perhaps it is possible for a computer program to "learn" but not do it in a manner similar to human learning processes? Bob D. >I >realise my earlier post was completely vague. I hope this clears the air. : ) > This research is just an attempt to utilize processing power differently, as >it is believed that todays chess programs waste it endlessly searching >variations deeper and deeper without "knowing" what they find You have a GREAT sense of humor! : ) I wish there was more of that here at CCC. Bob D. > >> >> >>> >>>I just happened upon this neat little program written in 1993 that seems to do >>>just that. It's very weak, but it's a beginning. >>> >>> >>>http://satirist.org/learn-game/systems/sal.html >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>>>>>Perhaps a lesser accomplishment would be "good enough." For example, one could >>>>>>>select a dozen or more specific humans and then emulate them individually. If >>>>>>>the group of humans selected for emulation were chosen wisely, maybe they would >>>>>>>represent [or "cover"] the entire population reasonably well. The larger the >>>>>>>group, the better they could represent the entire population of human >>>>>>>chessplayers. >>>>>>> >>>>>>>Suppose someone with a 2800 rating were selected and called opponent #1. >>>>>>>Then someone with a 2700 rating might be selected and called opponent #2. >>>>>>>This could be continued until the rating was so low that there would be no need >>>>>>>for more. >>>>>>> >>>>>>>#1 = 2800, #2 = 2700, #3 = 2600, #4 = 2500, etc. >>>>>>> >>>>>>>The next step might be to expand the list by having several individuals at each >>>>>>>level but with different playing styles. >>>>>>> >>>>>>>There should be quite a few distinct emulated humans at the amateur levels, >>>>>>>since that's where most of the people using the program would be. >>>>>>> >>>>>>>The intent maybe should be to emulate these people primarily in the middlegame >>>>>>>and maybe endgame. The opening repertoires of the individuals might also be >>>>>>>copied but that might not be such a good idea unless the repertoires were large. >>>>>>> A typical amateur might not have a complete opening repertoire at all. In this >>>>>>>case, one might be provided for him. >>>>>>> >>>>>>>To select a specific individual to be emulated, it would be necessary to have a >>>>>>>fairly large collection of his/her games played at the desired performance >>>>>>>level. Such a collection might be hard to find for amateurs. >>>>>>> >>>>>>>How could the chess play of a specific individual be emulated without such a >>>>>>>collection of games? Maybe general well-known traits of chessplayers at the >>>>>>>level being considered could be used to synthesize an emulation in that case. >>>>>>> >>>>>>>The emulations could then be used in a chess-playing program designed to serve >>>>>>>as a training tool. People would train against the emulated individuals at >>>>>>>their level in preparation for future contests with humans. >>>>>>> >>>>>>>Clearly, this would be inferior to a full-scale emulation of all of the >>>>>>>chessplayers in the World. For example, if the number of individuals emulated >>>>>>>were too small, one might "learn" the individual traits of the specific >>>>>>>individual emulated. After that, playing against that specific emulated >>>>>>>individual might become boring. >>>>>>> >>>>>>>Care would be required to assure that the games would have some variability. >>>>>>>This could be done in the opening, especially. Maybe a random number generator >>>>>>>would be used to randomly select the openings. This is surely done by most or >>>>>>>all programs using an opening book anyway. >>>>>>> >>>>>>>Bob D.
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.