Author: Alessandro Scotti
Date: 01:14:08 09/08/04
Go up one level in this thread
On September 07, 2004 at 20:45:39, David Dahlem wrote: >On September 07, 2004 at 20:39:31, Alessandro Scotti wrote: > >>On September 07, 2004 at 09:24:22, Rafael Andrist wrote: >> >>>There is not just only one position: >>> >>>8/8/7N/8/8/8/8/K1k1B3 w - - dm 33; >>>... >>>8/8/8/6B1/8/5k2/8/2K4N w - - dm 33; >> >>This list did no good for the KBNK algorithm... :-( > >Mine either. It takes mine 42 moves to solve a mate in 18 at Blitz 5. :-( > >I'm using two tables to force the enemy king into the proper corner. It doesn't >take much code to implement this method, i think if i find the proper table >values, it will work much better. Yesterday (or better, tonight, as that's when I can have some time for hobbies) I spent a couple of hours playing with EGTB and KBNK coefficients... my biggest problem is that I could find positions where the engine would not see a way to do real progress, because of the way parameters interact. Changing a parameter, maybe of just a small amount, would solve the position but break another. In most cases the basic algorithm is doing fine, often following the exact EGTB path, so these exceptions are really frustrating. Things that I have now written on my "todo" list: - double check the search algorithm, especially detection of 3-fold repetition and pruning, which in these cases is very dangerous; - staged algorithm; - automatic tuning of parameters using autoplay with EGTB. With the "staged" algorithm I try to set intermediate goals and reach them without worrying about the rest. For example: 1) centralize king, make sure king is closer to center than losing king, minimize distance; 2) drive losing king to the edge of the board; 3) drive losing king towards the bishop angle. When a goal is achieved, it is assigned a bonus score and the next goal kicks in. The idea is that dividing the steps in simpler goals could help the engine make progress, with respect to having a single goal that depends on many interacting factors that are difficult to tune. I hope I can do more experiments in the forthcoming nights, will post results... 8-O
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.