My experiment showed that the average difference in positions that qsearch considers quiet is 1.5 pawns or .75 pawns stm bonus. I would not call this small. There are many factors which can affect the value of a tempo, such as open lines, king safety, and pawn structure. Perhaps the simplest of these to understand is pawn structure. If the pawns are blocked, the value of a tempo is often very low. Another example is gambit positions. A tempo could easily be the difference between wining and losing, and I would not expect qsearch to be able to see this.wgarvin wrote:I don't understand what is position dependent about it. Q-search should take care of that, leaving you to evaluate a position where you can probably only make a small positional gain. Shouldn't the value of being able to make one move in a quiet position, almost always be small?jwes wrote:The value of the side to move is highly position dependent and I believe that if you set the the side to move bonus to the average value, it would drive the search to positions where the actual value is much lower, which would cause these positions to be significantly misevaluated.
Side to Move Bonus---does it help?
Moderators: hgm, Rebel, chrisw
-
- Posts: 778
- Joined: Sat Jul 01, 2006 7:11 am
Re: Side to Move Bonus---does it help?
-
- Posts: 10282
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: Side to Move Bonus---does it help?
If you want to check the real value of the side to move
then it may be better not to have side to move bonus unless the
position is quiet from both sides.
In other words result of qsearch is exactly the same as static evaluation in both cases when the side to move is white and when the side to move is black.
The problem is that checking it may be expensive.
then it may be better not to have side to move bonus unless the
position is quiet from both sides.
In other words result of qsearch is exactly the same as static evaluation in both cases when the side to move is white and when the side to move is black.
The problem is that checking it may be expensive.
Re: Side to Move Bonus---does it help?
With my engine Troy, a direct descendant of Hector, which competed in the nineties in a few CSVN NK tournaments and Aegon tournaments, I've done some testing with the side to move bonus.
I've implemented it as an evaluation term called right to move bonus. Current setting is 10 centipawns for normal positions throughout the game. I do know it helps, I can't tell how much, not having a test environment like Bob Hyatt.
I did not test Bob's approach of having a different bonus for middle games and end game. In endgame positions there is an evaluation term called opposition, which is a kind of right to move bonus as well.
For the opening position one could estimate the RTM bonus as follows:
- white wins approx 55%
- win% = 0.15 * eval + 0.5 for eval in [-3,3], approximation derived from data of http://www.ascotti.org/programming/chess/mat_stats.html
- so bonus in opening position is 33 centipawns.
If the RTM bonus is set at that level you'll see odd behaviour of the engine as describe by Bob, drooping pawns left and right.
regards, Maarten
I've implemented it as an evaluation term called right to move bonus. Current setting is 10 centipawns for normal positions throughout the game. I do know it helps, I can't tell how much, not having a test environment like Bob Hyatt.
I did not test Bob's approach of having a different bonus for middle games and end game. In endgame positions there is an evaluation term called opposition, which is a kind of right to move bonus as well.
For the opening position one could estimate the RTM bonus as follows:
- white wins approx 55%
- win% = 0.15 * eval + 0.5 for eval in [-3,3], approximation derived from data of http://www.ascotti.org/programming/chess/mat_stats.html
- so bonus in opening position is 33 centipawns.
If the RTM bonus is set at that level you'll see odd behaviour of the engine as describe by Bob, drooping pawns left and right.
regards, Maarten
-
- Posts: 20943
- Joined: Mon Feb 27, 2006 7:30 pm
- Location: Birmingham, AL
Re: Side to Move Bonus---does it help?
That is my thinking. The only reason I believe it works is that you effectively say "Here is the positional evaluation, but the side on move gets another chance to change things, which might swing the evaluation in his favor a bit. In endgames this is more important since that last king move can be very important for centralization/etc.wgarvin wrote:I don't understand what is position dependent about it. Q-search should take care of that, leaving you to evaluate a position where you can probably only make a small positional gain. Shouldn't the value of being able to make one move in a quiet position, almost always be small?jwes wrote:The value of the side to move is highly position dependent and I believe that if you set the the side to move bonus to the average value, it would drive the search to positions where the actual value is much lower, which would cause these positions to be significantly misevaluated.