Author: Graham Laight
Date: 07:27:10 05/03/01
Go up one level in this thread
On May 03, 2001 at 08:52:44, Robert Hyatt wrote: >On May 03, 2001 at 06:53:26, Graham Laight wrote: > >> >> >>There are 2 glaringly obvious points to be made in reply here: >> >>1. If the weights of the evaluation components are wrong (or if, as seems more >>likely, the program doesn't modify the values of the weightings according to the >>type of position as well as a human), this still represents a knowledge deficit >> >>2. The research discussed in "Chess Skills In Man And Machine" indicated that a >>human GM has expert knowledge on about 50,000 positional patterns. Crafty (and, >>I'm sure, any other program) has nothing like that number of evaluation >>components (clearly 1 evaluation component <> 1 positional pattern, but there's >>probably a correlation) > >This is probably a bit of "apples and oranges". IE humans recognize a >"fork" by pattern. A program recognizes it with a 3 ply search. Ditto for >overloaded pieces and a host of other things. Some things have been relegated >to the search to handle, others have to be done within the eval. > >I certainly agree that the computers are _way_ behind humans in total knowledge. >But the computers are ahead in some types of tactics (they are behind in some >of course). The problem now seems to be that humans are learning how to guide chess games into the type of position where the computers are weak. This will be a position where the program cannot evaluate the nodes accurately enough because: 1. In terms of the knowledge the computer has, the nodes all look to be of similar value 2. There's an important piece of positional knowledge which the human has, but which the computer doesn't, for which the consequences are over the computer's horizon The remedy would be, of course, to find ways to acquire and manage knowledge. If, for some reason, game tree search had proven to be impracticable for chess, then it may well be that static evaluation would be extremely sophisticated by now. >> >> >>I would reconcile the contradiction by reminding you that I'm not suggesting the >>use of a single technique, but rather a combination of techniques. > >If you look at traditional data mining applications, you generally find that >the humans know what they are looking for. IE one project here was to look Not necessarily. Data mining can show up correlations which the operators weren't expecting. When researchers were trying to find the cause of the increase in lung cancer in the 1950s, they tried correlating all sorts of variables to the disease. However - each one would have had to have been individually requested by a human. They did, of course, discover the good correlation with smoking, which lead to all the research to uncover the links. If they'd had the same statistics today, a data mining system would almost certainly have identified smoking as the highest correlation to lung cancer without a human having to specify that this correlation in particular be calculated. Caveat: data mining is not my field of expertise. I'm still quite confident that the above is correct, though. >at treatment in the ER for bacterial infections and outcomes. This might have >(years ago) uncovered the fact that young kids, asprin, and flu often lead to >a serious interaction. > >But chess seems different. I think that the difference is that in chess, it's not immediately obvious that you're searching for correlations between things. I would argue that this is what you actually are doing, however. Suppose, for example, that having your king near the centre of the board when the queens are removed is good. You could get a database of chase games, and correlate king proximity to the centre when the queens are removed with the final score. If there is a correlation, then you've obtained some knowledge. The weighting you would give this knowledge might be related to the strength of the correlation. Obviously, to make this work, you'd need a method of generating patterns to work with. Maybe genetic alorithms could be used for this purpose. >>In my own work with AI problems (not chess related unfortunately), I've found >>that some problems cannot be solved easily by using a single technique, but that >>by combining a combination of techniques, they can be resolved surprisingly >>well. >> >>For example - maybe an NN could be trained to guide the search for significant >>patterns in chess positions (and self improve on the job when it starts the real >>work). >> >>It is an unfortunate aspect of chess that other techniques have worked >>sufficiently well to have prevented interest in the real intelligence - the >>evaluation of positions - from becoming the major focus. >> >>-g > > >I wouldn't say that at all. if you look at my evaluation code you will >probably conclude that "knowledge" is considered important in many places... >Both general-purpose knowledge _and_ special-case knowledge. I can't speak >for everybody, but _my_ program has gotten "smarter" over the years. For >a reason... I assume you remove some knowledge as increasing NPS renders it irrelevant. But also, as the standard of your opponents becomes higher, you need to cover more of the positional cracks - so you need to add knowledge to stay in the game. This is good - but it's still not enough. A question posed in "Chess Skill In Man And Machine" was, "Why can't a computer be more like a human?". To rephrase that question for Crafty, "Why can a machine that can do both selective search and a good evaluation of half a million positions per second still be beaten by people who can only evaluate 2 or 3 positions per second?". Richard Lang told me at WMCCC 2K (as many other programmers have said) that his biggest frustration was that Genius had reached a certain level, and he found it very hard to get stronger, because if he improved the eval in one area, it would tend to get worse in another. At the same event, Franz Morsch (no less!) also told me that it was becoming very difficult to raise computer chess above the current standard. If this is the general situation, then those who master knowledge acquisition and management (the essence of intelligence) could well be richly rewarded. -g
This page took 0.06 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.