Author: Steffen Jakob
Date: 23:21:31 09/03/00
Go up one level in this thread
On September 04, 2000 at 00:37:36, Robert Hyatt wrote: >On September 03, 2000 at 23:06:57, stuart taylor wrote: > >>On September 03, 2000 at 22:44:16, Robert Hyatt wrote: >> >>>On September 03, 2000 at 22:36:41, William Bryant wrote: >>> >>>>I/m working on my time allocation to improve my program's play. >>>> >>>>Given a time increment of Game/Unit_Time (or Game/Unit_Time + Increment per >>>>move), how do you allocate the Unit_Time among individual searches? >>>> >>>>At present I'm doing the following. >>>> >>>>1. In the opening, (when out of book), I'm allocating 1/64th of the time left >>>>for a search (time_left>>6). >>>> >>>>2. When past move 16, I change this to 1/32nd of the time left (time_left>>5). >>>> >>>>3. If I fail low at the root, I will allocate up to an additional 5x the initial >>>>time to resolve the fail low. It may take less, but this is the limit. >>>> >>>>Any thoughts, comments, suggestions or discussion would be appreciated. >>>> >>>>William >>>>wbryant@ix.netcom.com >>> >>> >>>I use a simple apporach for "level X Y" games on ICC. My target time is >>>always X/25+Y (x=time left on clock, y=increment). >> >>Hiarcs 7.00 (at any rate. 7.01 is the same, and I think 7.32 is too). allocates >>very intelligently when losing, I think. The more the evaluation drops, the more >>it searches deeper. Until it is just too much time. But when winning, it is too >>confident. It doesn't use most of its clock time. It can sometimes even fail >>high prematurely, and only on next move, fail low. >>S.Taylor > > >Crafty has always done that, as does most programs nowadays. IE if the score >drops significantly, it won't move in the normal target time limit, but will >continue to search longer to try to solve the problem it has just seen... I don't only search longer when the score drops but also if the score gets BETTER a lot. I do this to avoid falling into traps. I implemented this feature long time ago after a game Hossa vs Diep where Diep gave away two rooks for a mate attack. Hossa was very happy to grab all the material and realized too late that he was in big trouble. The idea behind extending the time when the score jumps high is that you win anyway if this score is correct and that you might escape if it was a trap. IMHO this is a good feature especially for slow searchers like Hossa which get outsearched very often. Another trick I do: if the score is between 0.7 and 2.8 I search longer. Here is the idea that these are often positions where a winning shot can be found if you search deep enough. Again this is especially good for slow searchers. Best wishes, Steffen.
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.