Author: Jay Scott
Date: 16:19:35 12/29/97
Go up one level in this thread
On December 27, 1997 at 14:04:12, Don Dailey wrote: >I felt the same way when I read this. I don't believe you can change >a programs playing "style" very much with only search changes. I'm not so sure. I think drastic search changes can have big effects on a program's style. If you've only tried mostly-full-width search with the usual kinds of pruning and extensions, then you've never tried a drastically different search algorithm. Several years ago I wrote a chess program named Kon to look into automated learning of search control. I hacked up a crude fixed evaluation function by hand and gave the program a method to learn from experience which moves to search. Kon used a breadth-first search algorithm which kept the entire tree in memory, so it was quite different from the usual alpha-beta stuff. Kon played on what was then the Internet Chess Server. Before learning, I think the ICS rating of the final version was something like 1300. Search control learning was able to boost this to 1700 or so, which was respectable if you figure that the search was slow and the evaluator was extremely simpleminded. I was struck by the stylistic difference between Kon and the typical alpha-beta program. Kon was much less consistent: sometimes it would find a clever, deep tactic that was beyond the skill of a human B-player or a B-strength alpha-beta program, and sometimes it would fall for two-movers. The search control learning had picked up how to spot some tactical moves, but it wasn't able to figure them all out, even some shallow ones. In a losing position Kon would usually search broadly and shallowly, struggling to find a move that stopped all the threats. In a tactical position it would typically search deeply and narrowly, trying to fathom the tricks, and sometimes completely ignore a whole set of what it considered boring moves at the first ply. A full-width program would have noticed that some of those "boring" moves were actually tricky a few plies down the line. Kon was popular with lower-rated players on ICS. I think that was partly because of its interesting style (and partly because there weren't any other programs weak enough to be fun for them to play). The evaluation function sets the program's goals, but the search controls how well the program can meet them. Changing the search can reposition a program on the defensive<->aggressive spectrum (a la Genius, if I can believe what I'm reading) and on the solid<->speculative spectrum (which you could also call consistent<->inconsistent). Kon was pretty much equally defensive and aggressive, but definitely on the speculative side of solid. Jay
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.