You can find out all you want at Bruce Mroeland's site: http://www.seanet.com/~brucemo/chess.htm
He explains a number of the different algorithms and techniques used today, and also provides a simple program, Gerbil, with commented source-code available for you to learn from.
Your reply dealt mostly with Opening Book learning which is only a part of the equation. I think the original poster had in mind something more along the lines of neural network programs such as used in backgammon and fathered by Gerald Tesauro (see his paper at http://www.research.ibm.com/massive/tdl.html). There have been efforts done to have programs adjust their own evaluation weights (values for each piece of knowledge), hoping to arrive at some perfect balance. The success of this has varied a great deal. One of the problems could very well be that the quantity of games was insufficient, since as can be seen in Tesauro's results, his first version of TD-Gammon was the result of 300,000 games, and his last from 1.5 million. Another problem is simply in what knowledge is included. You can have the program play itself forever, but if it doesn't know what a doubled pawn is, or take special cases into consideration, the weights will never work. Today's programs try to achieve a fine balance between the most knowledge that is *practical* with speed and depths achieved. The reason (I realize you know this Mig) is that too much knowledge can slow the program down so much that this knowledge cannot compensate the shallow depth reached. In other words, understanding the position like a World Champion won't make up for the fact that it can't see beyond its nose. Of course seeing 20 moves ahead but clueless is no better. Deep Blue, which has often been touted as being a very ignorant program, really wasn't and most likely suffered from a severe imbalance in its weights as Hsu (the author) said it had no less than 6000(!) individual elements of chess knowledge, which no matter how you look at it is loads more than any program today. DB could afford it, since the knowledge didn't slow it down any. In comparison, Crafty, a top freeware program written by Dr. Robert Hyatt, and not considered ignorant by any means had a little over 150 pieces of knowledge a couple of versions ago. Those interested can download the source-code at Hyatt's FTP site at ftp.cis.uab.edu/pub/hyatt Hsu used some automatic learning process to try to finish fine-tuning the evaluation weights, but considering that there were reports off debugging DURING the match, it seems obvious that it was very far from sufficient. It's not unlikely that had the DB team spent one year JUST fine-tuning that eval, without making any other changes, the program would have been enormously stronger, but to be fair, that's only speculation.
Note that Fritz7 has slowed down over the years. This is because it now has a lot of general chess knowledge built in. But you cannot measure knowledge by lack of speed. Fritz7 has more knowledge than Hiarcs 7, which is a number of years old. In fact it has more chess knowledge than any other top program available today.
How can you affirm that last phrase? Can you truly affirm it has more knowledge than all of the following programs (for example)?:
Shredder6
Junior7
Tiger14
RebelCentury4
Almost all the commercials are participating. The only ones that aren't are Chessmaster (no doubt because the company just changed hands), Tiger 14, and Rebel. AFAIK, every other one is there. Not to mention experimental versions (such as Hiarcs 8) and programs that can compete with commercials (Ferret and Yace for example).
Slashdot Mirror
You can find out all you want at Bruce Mroeland's site: http://www.seanet.com/~brucemo/chess.htm He explains a number of the different algorithms and techniques used today, and also provides a simple program, Gerbil, with commented source-code available for you to learn from.
Your reply dealt mostly with Opening Book learning which is only a part of the equation. I think the original poster had in mind something more along the lines of neural network programs such as used in backgammon and fathered by Gerald Tesauro (see his paper at http://www.research.ibm.com/massive/tdl.html). There have been efforts done to have programs adjust their own evaluation weights (values for each piece of knowledge), hoping to arrive at some perfect balance. The success of this has varied a great deal. One of the problems could very well be that the quantity of games was insufficient, since as can be seen in Tesauro's results, his first version of TD-Gammon was the result of 300,000 games, and his last from 1.5 million. Another problem is simply in what knowledge is included. You can have the program play itself forever, but if it doesn't know what a doubled pawn is, or take special cases into consideration, the weights will never work. Today's programs try to achieve a fine balance between the most knowledge that is *practical* with speed and depths achieved. The reason (I realize you know this Mig) is that too much knowledge can slow the program down so much that this knowledge cannot compensate the shallow depth reached. In other words, understanding the position like a World Champion won't make up for the fact that it can't see beyond its nose. Of course seeing 20 moves ahead but clueless is no better. Deep Blue, which has often been touted as being a very ignorant program, really wasn't and most likely suffered from a severe imbalance in its weights as Hsu (the author) said it had no less than 6000(!) individual elements of chess knowledge, which no matter how you look at it is loads more than any program today. DB could afford it, since the knowledge didn't slow it down any. In comparison, Crafty, a top freeware program written by Dr. Robert Hyatt, and not considered ignorant by any means had a little over 150 pieces of knowledge a couple of versions ago. Those interested can download the source-code at Hyatt's FTP site at ftp.cis.uab.edu/pub/hyatt Hsu used some automatic learning process to try to finish fine-tuning the evaluation weights, but considering that there were reports off debugging DURING the match, it seems obvious that it was very far from sufficient. It's not unlikely that had the DB team spent one year JUST fine-tuning that eval, without making any other changes, the program would have been enormously stronger, but to be fair, that's only speculation.
Albert Silver
Note that Fritz7 has slowed down over the years. This is because it now has a lot of general chess knowledge built in. But you cannot measure knowledge by lack of speed. Fritz7 has more knowledge than Hiarcs 7, which is a number of years old. In fact it has more chess knowledge than any other top program available today. How can you affirm that last phrase? Can you truly affirm it has more knowledge than all of the following programs (for example)?: Shredder6 Junior7 Tiger14 RebelCentury4
Almost all the commercials are participating. The only ones that aren't are Chessmaster (no doubt because the company just changed hands), Tiger 14, and Rebel. AFAIK, every other one is there. Not to mention experimental versions (such as Hiarcs 8) and programs that can compete with commercials (Ferret and Yace for example).
You are mistaken. Kasparov played Deep Thought in a 2-game match in 1989 and clobbered it.