Moderator: Andres Valverde
Tom Likens wrote:Afternoon,
I've been experimenting lately with various move ordering schemes to reduce the size of the search tree. I've mainly been tracking two components,
1) the percentage of fail-highs at CUT nodes for the 1st, 2nd and 3rd moves
2) the percentage of times the 1st-10th moves at a PV node is the best move in the position.
As Gian-Carlo pointed out recently (1) isn't such a great indicator of tree size (it averages around 90-93% for the first move), but (2) is much more interesting, it averages around 50-65%, also for the first move. To really understand this, I break my search up into the following phases:
1. Hash move
2. Killer mates
3. Winning/Even captures
4. Killers (non-mate)
5. Queen promotions
6. Castling moves (kingside/queenside)
7. Winning even mundanes
8. Losing captures
9. Losing mundanes
10. Non-queen promotions
I verify of course, that I don't search any move more than once. The interesting thing about this is that if I track the percentage of each phase where the 1st, 2nd, 3rd etc. move caused a cutoff or was the best move--the move ordering is significantly different for CUT nodes versus PV nodes (I don't check ALL nodes, since they shouldn't produce a cutoff and order is unimportant).
Has anyone else seen a similar phenomenon? It may be advantageous to order the moves differently depending on the type of node we are searching. It's interesting, although the number of PV nodes is small compared to the number of CUT/ALL nodes, their overall effect on the size of the tree is enormous. Spending extra time ordering these nodes could potentially produce a huge benefit.
regards,
--tom
mridul wrote:Promotions come after killers for you ?
For me , they are at par - and sometimes higher than good captures.
for depth > some_value_based_on_initial_depth , I do detailed analysis to determine good moves/captures and order moves accordingly (like using hung piece info , etc).
Mridul
Uri Blass wrote:Hello Tom,
I think that it is a mistake generally to generate promotion after killers.
In most of the cases you can detect fast that there is no pawn in the 7th rank so there is no promotion so the overhead is not significant and when there is a pawn in the 7th rank it is important to check promotions.
Uri
Tom Likens wrote:Uri Blass wrote:Hello Tom,
I think that it is a mistake generally to generate promotion after killers.
In most of the cases you can detect fast that there is no pawn in the 7th rank so there is no promotion so the overhead is not significant and when there is a pawn in the 7th rank it is important to check promotions.
Uri
Hey Uri,
I'll give the reordering of promotions another try. It's been a long time since I measured its effect and a lot has changed in my program since then, so it may be better now.
OTOH, have you tried different reordering schemes based on the type of node you're searching in movei? I'm curious about other programmer's results.
regards,
--tom
Tom Likens wrote:mridul wrote:Promotions come after killers for you ?
For me , they are at par - and sometimes higher than good captures.
for depth > some_value_based_on_initial_depth , I do detailed analysis to determine good moves/captures and order moves accordingly (like using hung piece info , etc).
Mridul
Hello Mridul,
Yes, the various phases are generated with separate move generators which has a (small) amount of overhead. In 99.9+% of the cases there are no promotion moves, so the overhead of calling the queen promotion move generator is simply wasted. In endgames where this matters, there is some disadvantage, but nothing significant in my testing and overall the benefit of not checking it outweighs the advantage seen in the late endgame.
I've toyed, but never seriously pursued, the idea of using the evaluation function to order the winning/even mundanes in the first few (upper) plies of the search. I do an extensive move ordering using the eval at the root, but not below the root, instead relying on the move ordering I previously posted.
regards,
--tom
Tom Likens wrote:Afternoon,
I've been experimenting lately with various move ordering schemes to reduce the size of the search tree. I've mainly been tracking two components,
1) the percentage of fail-highs at CUT nodes for the 1st, 2nd and 3rd moves
2) the percentage of times the 1st-10th moves at a PV node is the best move in the position.
As Gian-Carlo pointed out recently (1) isn't such a great indicator of tree size (it averages around 90-93% for the first move), but (2) is much more interesting, it averages around 50-65%, also for the first move. To really understand this, I break my search up into the following phases:
1. Hash move
2. Killer mates
3. Winning/Even captures
4. Killers (non-mate)
5. Queen promotions
6. Castling moves (kingside/queenside)
7. Winning even mundanes
8. Losing captures
9. Losing mundanes
10. Non-queen promotions
I verify of course, that I don't search any move more than once. The interesting thing about this is that if I track the percentage of each phase where the 1st, 2nd, 3rd etc. move caused a cutoff or was the best move--the move ordering is significantly different for CUT nodes versus PV nodes (I don't check ALL nodes, since they shouldn't produce a cutoff and order is unimportant).
Has anyone else seen a similar phenomenon? It may be advantageous to order the moves differently depending on the type of node we are searching. It's interesting, although the number of PV nodes is small compared to the number of CUT/ALL nodes, their overall effect on the size of the tree is enormous. Spending extra time ordering these nodes could potentially produce a huge benefit.
regards,
--tom
pv cut all flip
0.13 77.57 22.3 4.5
0.06 78.5 21.44 2.46
0.09 72.15 27.76 8.2
0.11 77.22 22.68 5.5
0.07 75.04 24.89 0.77
0.09 76.55 23.36 4.55
0.03 80.79 19.18 2.13
0.11 84.71 15.18 2.84
0.07 84.51 15.41 2.67
0.17 71.19 28.64 8.71
0.08 80.99 18.92 2.58
0.05 69.38 30.58 9.31
0.05 76.62 23.33 3.95
0.07 74.03 25.9 4.56
0.08 77.47 22.45 5.11
0.06 82.07 17.86 2.51
0.08 72.54 27.38 12.08
0.23 78.46 21.32 3.12
0.08 75.76 24.16 6.95
0.17 78.98 20.85 4.52
0.07 67.03 32.9 9.15
0.13 82.99 16.89 3.82
0.04 80.56 19.4 0.82
0.07 78.73 21.2 5.53
0.11 79.83 20.06 3.38
0.06 85.54 14.4 1.89
0.11 73.9 25.99 4.71
0.09 74.28 25.63 7.14
0.05 80.98 18.97 3.13
0.11 82.84 17.05 3.16
average
0.09 77.70 22.20 4.66
pv cut all flip fh fh_pv fh_cut fh_all fh_3 fh_pv3 fh_cut3 fh_all3
0.13 77.47 22.40 4.53 32.33 0.11 87.06 12.83 31.39 0.10 89.66 10.24
0.06 78.21 21.73 2.40 23.62 0.05 90.73 9.22 23.33 0.05 91.87 8.08
0.08 71.92 28.00 8.56 39.66 0.05 79.33 20.61 37.66 0.05 83.55 16.40
0.09 77.17 22.74 5.50 37.62 0.06 86.82 13.12 36.20 0.06 90.23 9.72
0.06 73.67 26.27 0.67 11.13 0.12 90.70 9.18 11.05 0.12 91.34 8.55
0.08 76.94 22.98 4.49 28.92 0.06 86.26 13.67 28.16 0.06 88.60 11.34
0.03 80.50 19.47 2.20 21.72 0.03 89.30 10.67 21.23 0.03 91.36 8.61
0.09 84.46 15.45 2.91 22.94 0.08 87.39 12.53 22.28 0.08 89.96 9.96
0.07 84.58 15.35 2.57 17.57 0.09 85.41 14.50 17.13 0.09 87.59 12.33
0.15 70.36 29.49 9.29 38.08 0.12 78.03 21.85 36.54 0.10 81.30 18.59
0.09 81.36 18.55 2.46 22.87 0.10 89.82 10.07 22.47 0.10 91.42 8.49
0.05 69.46 30.50 9.48 39.62 0.03 80.02 19.95 37.92 0.03 83.61 16.36
0.05 76.57 23.38 3.95 39.35 0.03 90.61 9.36 38.51 0.03 92.58 7.39
0.06 74.69 25.25 4.27 22.76 0.05 79.88 20.07 21.94 0.05 82.88 17.07
0.08 77.19 22.73 5.30 35.71 0.06 86.42 13.52 34.64 0.05 89.09 10.86
0.05 82.56 17.39 2.28 28.55 0.04 92.16 7.80 28.06 0.04 93.76 6.20
0.07 72.25 27.68 11.62 36.13 0.05 69.00 30.95 33.82 0.05 73.70 26.25
0.20 78.59 21.21 3.10 27.80 0.17 90.20 9.63 27.29 0.17 91.88 7.95
0.07 75.49 24.44 6.91 36.48 0.05 81.68 18.26 35.26 0.05 84.51 15.44
0.14 78.63 21.23 4.52 27.87 0.11 83.85 16.04 26.95 0.10 86.70 13.20
0.06 66.76 33.18 9.23 41.91 0.03 81.08 18.89 39.77 0.03 85.43 14.54
0.12 83.05 16.83 3.81 25.87 0.12 85.73 14.15 25.13 0.10 88.27 11.62
0.03 80.46 19.51 0.75 30.65 0.02 97.58 2.40 30.56 0.02 97.88 2.10
0.07 79.03 20.90 5.22 28.04 0.05 84.10 15.85 26.95 0.05 87.49 12.46
0.10 79.65 20.25 3.35 23.99 0.12 86.59 13.29 23.42 0.11 88.71 11.18
0.05 85.55 14.40 1.95 19.29 0.07 89.96 9.98 18.89 0.06 91.84 8.10
0.09 74.16 25.76 4.52 28.43 0.06 85.06 14.87 27.59 0.06 87.66 12.28
0.08 74.64 25.28 7.06 38.53 0.05 83.11 16.84 36.80 0.04 87.01 12.95
0.05 80.50 19.45 3.30 29.33 0.03 89.47 10.50 28.78 0.03 91.17 8.81
0.11 82.86 17.02 3.07 24.88 0.10 87.99 11.91 24.24 0.09 90.32 9.58
0.11 78.78 21.11 4.91 31.29 0.07 85.51 14.41 30.32 0.07 88.27 11.67
0.13 82.98 16.88 2.70 26.31 0.11 89.55 10.34 25.86 0.10 91.11 8.79
0.10 82.50 17.40 2.82 26.77 0.07 89.53 10.40 26.00 0.07 92.19 7.75
0.15 85.06 14.79 3.48 25.77 0.08 87.81 12.10 25.07 0.07 90.25 9.67
0.19 87.17 12.64 2.94 22.27 0.19 86.70 13.12 21.57 0.16 89.54 10.31
0.07 82.41 17.52 2.74 23.86 0.05 88.01 11.93 23.36 0.05 89.91 10.05
0.09 86.27 13.65 1.60 20.14 0.09 90.93 8.99 19.80 0.08 92.50 7.42
0.10 87.15 12.75 2.18 19.69 0.09 87.82 12.09 19.24 0.08 89.86 10.06
0.16 78.69 21.15 4.63 34.48 0.09 87.37 12.53 33.22 0.08 90.70 9.22
0.07 82.14 17.79 4.14 29.08 0.06 86.35 13.59 28.19 0.05 89.07 10.89
0.09 84.33 15.58 3.86 25.56 0.08 85.36 14.56 24.89 0.07 87.67 12.26
0.18 83.52 16.30 3.47 28.41 0.14 87.99 11.87 27.52 0.12 90.83 9.04
0.07 83.47 16.46 2.83 26.75 0.07 89.19 10.75 26.03 0.06 91.65 8.29
0.11 84.44 15.45 3.66 28.06 0.09 87.23 12.68 27.36 0.08 89.48 10.44
0.06 77.58 22.36 3.99 25.38 0.06 83.87 16.07 24.65 0.05 86.35 13.60
0.03 83.39 16.58 0.88 20.21 0.02 95.09 4.89 20.06 0.02 95.83 4.15
0.05 81.60 18.35 1.18 13.35 0.08 90.17 9.74 13.11 0.08 91.83 8.09
0.10 80.95 18.94 2.88 27.75 0.11 90.17 9.72 27.24 0.10 91.85 8.05
0.26 79.97 19.77 3.32 28.85 0.12 89.20 10.68 28.15 0.10 91.43 8.47
0.08 80.58 19.35 3.48 30.98 0.05 89.04 10.91 30.13 0.04 91.55 8.40
0.08 79.79 20.13 3.44 27.24 0.08 87.58 12.33 26.66 0.08 89.49 10.43
0.19 75.87 23.94 6.41 35.47 0.13 82.89 16.98 34.26 0.11 85.80 14.08
0.07 77.32 22.61 3.82 29.38 0.05 87.64 12.31 28.50 0.05 90.33 9.63
0.11 85.32 14.57 3.23 25.83 0.08 87.55 12.37 24.97 0.07 90.54 9.39
0.08 83.90 16.03 4.14 26.19 0.06 84.86 15.09 25.42 0.05 87.43 12.52
0.10 83.35 16.56 3.31 26.83 0.08 88.23 11.69 26.07 0.07 90.80 9.13
0.06 78.32 21.61 2.73 30.85 0.04 92.26 7.70 30.40 0.04 93.62 6.35
0.12 81.16 18.72 3.55 23.92 0.12 85.51 14.37 23.18 0.12 88.24 11.65
0.26 78.01 21.73 3.42 20.91 0.32 82.88 16.80 20.37 0.31 85.06 14.63
0.05 76.43 23.52 6.58 30.60 0.04 81.16 18.80 29.30 0.04 84.77 15.19
0.07 80.59 19.35 3.79 26.31 0.06 86.67 13.28 25.42 0.05 89.69 10.26
0.17 79.98 19.85 3.98 22.64 0.18 82.69 17.13 21.94 0.17 85.33 14.50
0.05 81.71 18.24 3.40 22.85 0.04 87.57 12.39 22.52 0.04 88.85 11.11
0.08 74.92 25.00 3.21 23.06 0.08 86.24 13.69 22.44 0.07 88.62 11.31
0.08 77.58 22.34 2.41 16.92 0.13 83.05 16.82 16.59 0.13 84.69 15.19
pv cut all flip fh fh_pv fh_cut fh_all fh_3 fh_pv3 fh_cut3 fh_all3
0.09 79.48 20.42 4.02 27.61 0.08 86.54 13.37 26.80 0.08 88.98 10.95
Daniel Shawul wrote:Ok i run the test on lct2 and bt2630.
Average for both is:
- Code: Select all
pv cut all flip fh fh_pv fh_cut fh_all fh_3 fh_pv3 fh_cut3 fh_all3
0.09 79.48 20.42 4.02 27.61 0.08 86.54 13.37 26.80 0.08 88.98 10.95
fh_3 is fail high in the first 3 moves, and flip is change from expected CUT node to an ALL node.
ps: how to post a spread sheat?
where did the test forum go?
Daniel Shawul wrote:I made a quick experiment on bt2630 and measured the flip rate (from cut to all node depending on the result of search of the first move only).
I will measure percentage of fail high on pv nodes (your criteria)later.
- Code: Select all
pv cut all flip
0.13 77.57 22.3 4.5
0.06 78.5 21.44 2.46
0.09 72.15 27.76 8.2
0.11 77.22 22.68 5.5
0.07 75.04 24.89 0.77
0.09 76.55 23.36 4.55
0.03 80.79 19.18 2.13
0.11 84.71 15.18 2.84
0.07 84.51 15.41 2.67
0.17 71.19 28.64 8.71
0.08 80.99 18.92 2.58
0.05 69.38 30.58 9.31
0.05 76.62 23.33 3.95
0.07 74.03 25.9 4.56
0.08 77.47 22.45 5.11
0.06 82.07 17.86 2.51
0.08 72.54 27.38 12.08
0.23 78.46 21.32 3.12
0.08 75.76 24.16 6.95
0.17 78.98 20.85 4.52
0.07 67.03 32.9 9.15
0.13 82.99 16.89 3.82
0.04 80.56 19.4 0.82
0.07 78.73 21.2 5.53
0.11 79.83 20.06 3.38
0.06 85.54 14.4 1.89
0.11 73.9 25.99 4.71
0.09 74.28 25.63 7.14
0.05 80.98 18.97 3.13
0.11 82.84 17.05 3.16
average
0.09 77.70 22.20 4.66
BT2630 is a tactical testsuite so it might not be a good represenative.
What is interesting for me is the flip rate is small in endgame position like 5 , 23. And very high in tactical postions like pos 17 (this one is similar to the famous WAC141).
Daniel Shawul wrote:Ok i run the test on lct2 and bt2630.
Average for both is:
- Code: Select all
pv cut all flip fh fh_pv fh_cut fh_all fh_3 fh_pv3 fh_cut3 fh_all3
0.09 79.48 20.42 4.02 27.61 0.08 86.54 13.37 26.80 0.08 88.98 10.95
fh_3 is fail high in the first 3 moves, and flip is change from expected CUT node to an ALL node.
Actually, my critiera was a bit different than just the fail-high percentage at PV nodes. I was (and am) interested in the percentage of times the first move searched at a PV node returns the *best* score at that node, since improving this number should help reduce the tree, (although your idea about using the information to make better late-move reductions is clever).
Thinking about your posted numbers I'm wondering if the percentage of fail-highs seen at PV nodes isn't a direct measure of how appropriately the aspiration window has been set at the root. It would be interesting to run an experiment where the aspiration window was varied to determine if the fail-highs at PV nodes also varied in the opposite direction, (this would also be affected, I believe by the type of search you have--e.g. either fail-soft or fail-hard).
Return to Programming and Technical Discussions
Users browsing this forum: No registered users and 36 guests