Author: Bob Durrett
Date: 18:43:17 11/17/00
How do humans get high performance ratings? They win most of the time and do it against strong opposition. This must be true for chess-playing programs too. High probability of getting the move right results in higher ratings, since ratings are merely measures of average performance. Ratings are everything. It seems that the best approach to engine design surely must be a probabilistic approach, more than anything else. Add whatever features increase the probabilities and discard the rest. One adds "knowledge" until the performance drops. One adds "speed" till the performance drops. And many innovations are added, but thrown out if the performance drops. But performance must be measured in probabilistic terms inasmuch as very high probability of winning against strong opposition is "good enough." So, in the final analysis, isn't that really the way the top programs are produced? They are tested, tested again, and then tested ten more times. Why? So that they will get the right answer most of the time, at least more often than the competition. Any computations which do not "earn their salt" are discarded. If they take up too much computer resources too often then they are bad. Developing the very best programs must be like playing a game of poker. You try to win most of the time, and if you do, you go home rich. I doubt that the top programs are intentionally designed to meet any idealistic goals, like "being knowledge based" or "being deep searchers." The only idealism seems to be to up the rating. Incidentally, maybe "going home rich" is the really best, practical, criteria for engine design. If you can sell a million copies of a poorly-performing program, then you "win."
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.