You search drops in QS? Whenever I tried that in Shogi it led to total search explosion. Even if it was just safe contact-check drops. Of course I extended the evasion in that case. Or do you mean you allow drops as check evasion?
Large depth doesn't always mean anything. E.g. in mini-Shogi Nebiyu usually reaches nearly twice the depth Shokidoki reaches, and it still loses most games. You can also prune/reduce too much...
Crazyhouse tournaments and rating list
Moderators: hgm, Rebel, chrisw
-
- Posts: 27808
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
-
- Posts: 2929
- Joined: Sat Jan 22, 2011 12:42 am
- Location: NL
Re: CCVA CrazyWa 0.0.4 Gauntlet Tour Nr 21
I thought I did. Turns out I just generate them, but they are pruned except as check evasions. It might be worth searching safe contact check drops though, but I'll need to look into that.hgm wrote:You search drops in QS? Whenever I tried that in Shogi it led to total search explosion. Even if it was just safe contact-check drops. Of course I extended the evasion in that case. Or do you mean you allow drops as check evasion?
True, Nebiyu seems to prune too aggressively. Sunsetter seems to play very well though.Large depth doesn't always mean anything. E.g. in mini-Shogi Nebiyu usually reaches nearly twice the depth Shokidoki reaches, and it still loses most games. You can also prune/reduce too much...
-
- Posts: 2929
- Joined: Sat Jan 22, 2011 12:42 am
- Location: NL
Re: Crazyhouse tournaments and rating list
This is a very interesting addition; it seems to help SjaakII quite a bit as well. Not enough games yet to say anything definitive though, so I'll wait and see (everything else I tried recently ended up being no better than what I had before).hgm wrote: The improvements from 0.0.0 to 0.0.1 were all search; probably null move contributed the most. Evaluation can do a lot too, though: I noticed that CrazyWa was very passive, so I added two very simple evaluation terms to cure it, by encouraging it to develop its pieces. These were penalties for a Rook trapped in a corner (e.g. Ra1,with own pieces on a2 and b1), and a penalty for a Bishop trapped on its original square (e.g. Bc1 with own pieces on b2 and d2). Those terms alone increase the result against Sjaak from 45% to 63%! (measured in 360 games 40 moves/min).
Looking into this revealed some fundamental flaws in the way I evaluates mobility. For one thing, it fails to clear away occupied squares (so a Rook an a1 in the opening gets counted as having 2 moves available for mobility), but it also fails to make a distinction between squares that can be moved to and squares that can be captured on (not that it matters for most pieces). Conceptually, it should probably consider
- Actual mobility, in the sense of squares the piece can move to.
- Square control, squares on the board where an enemy piece can be captured
- Defending own pieces
- Attacking enemy pieces
-
- Posts: 4833
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Re: CCVA CrazyWa 0.0.4 Gauntlet Tour Nr 21
Average depth comparison for winning and losing score window.Ferdy wrote:Average depth on this blitz (TC 3'+2") tour nr 21, when scores are within the window of [-10.0, +10.0] pawn unit.
Code: Select all
nr player aveDepth games 1 Sunsetter 9 32bit 12.92 24 2 NebiyuAlien 1.45a 32bit 11.40 24 3 TJchess 1.3 64bit 9.60 24 4 Imortal 2.9001 32bit 9.35 24 5 CrazyWa 0.0.4 32bit 9.18 216 6 Sjaak II 1.3.1a 64bit 8.85 24 7 KKFChess 2.6.7 beta 32bit 8.20 24 8 Sjeng 11.2 32bit 7.70 24 9 TSCP zh 1.1 32bit 7.42 24 10 Pulsar 2009 9b 32bit 7.07 24
Everybody gains depth when winning, Sunsetter loses some depths when faced with losing position. Nebiyu is somewhat neutral.
A. Lossing score window:
Code: Select all
file : ccva_tour_nr_21.pgn
event : CCVA CrazyWa 0.0.4 Gauntlet Blitz Tour Nr 21
score window : [-6.0, -1.0]
nr player aveDepth cntDepth games
1 NebiyuAlien 1.45a 32bit 11.07 162 24
2 Sunsetter 9 32bit 9.24 114 24
3 TJchess 1.3 64bit 8.47 105 24
4 Imortal 2.9001 32bit 8.29 129 24
5 Sjaak II 1.3.1a 64bit 8.18 194 24
6 CrazyWa 0.0.4 32bit 7.87 1394 216
7 KKFChess 2.6.7 beta 32bit 7.37 171 24
8 Sjeng 11.2 32bit 6.99 205 24
9 TSCP zh 1.1 32bit 6.30 175 24
10 Pulsar 2009 9b 32bit 5.60 139 24
Code: Select all
file : ccva_tour_nr_21.pgn
event : CCVA CrazyWa 0.0.4 Gauntlet Blitz Tour Nr 21
score window : [+1.0, +6.0]
nr player aveDepth cntDepth games
1 Sunsetter 9 32bit 12.71 329 24
2 NebiyuAlien 1.45a 32bit 11.63 472 24
3 Imortal 2.9001 32bit 9.80 220 24
4 TJchess 1.3 64bit 9.61 293 24
5 CrazyWa 0.0.4 32bit 9.33 2882 216
6 Sjaak II 1.3.1a 64bit 8.64 303 24
7 KKFChess 2.6.7 beta 32bit 7.98 300 24
8 Sjeng 11.2 32bit 7.65 390 24
9 TSCP zh 1.1 32bit 7.58 266 24
10 Pulsar 2009 9b 32bit 7.22 309 24
Code: Select all
aveDepth = totalDepth/cntDepth
-
- Posts: 4833
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Re: CCVA CrazyWa 0.0.4 Gauntlet Tour Nr 21
There was a bug in my previous calculation of average depth, this one is now correct.
A. Win/Loss
B. Losing
C. Winning
A. Win/Loss
Code: Select all
file : ccva_tour_nr_21.pgn
event : CCVA CrazyWa 0.0.4 Gauntlet Blitz Tour Nr 21
score window : [-6.0, +6.0]
nr player aveDepth cntDepth games
1 Sunsetter 9 32bit 17.03 463 24
2 NebiyuAlien 1.45a 32bit 14.57 477 24
3 Imortal 2.9001 32bit 10.70 215 24
4 TJchess 1.3 64bit 10.43 375 24
5 Sjaak II 1.3.1a 64bit 9.45 430 24
6 CrazyWa 0.0.4 32bit 9.01 3892 216
7 KKFChess 2.6.7 beta 32bit 7.45 471 24
8 Sjeng 11.2 32bit 6.93 514 24
9 TSCP zh 1.1 32bit 5.72 440 24
10 Pulsar 2009 9b 32bit 5.33 456 24
Code: Select all
file : ccva_tour_nr_21.pgn
event : CCVA CrazyWa 0.0.4 Gauntlet Blitz Tour Nr 21
score window : [-6.0, -1.0]
nr player aveDepth cntDepth games
1 Sunsetter 9 32bit 15.73 15 24
2 NebiyuAlien 1.45a 32bit 14.09 81 24
3 TJchess 1.3 64bit 11.08 13 24
4 Imortal 2.9001 32bit 11.00 14 24
5 Sjaak II 1.3.1a 64bit 8.84 69 24
6 CrazyWa 0.0.4 32bit 8.02 675 216
7 KKFChess 2.6.7 beta 32bit 7.26 146 24
8 Sjeng 11.2 32bit 6.66 140 24
9 TSCP zh 1.1 32bit 5.50 126 24
10 Pulsar 2009 9b 32bit 5.04 113 24
Code: Select all
file : ccva_tour_nr_21.pgn
event : CCVA CrazyWa 0.0.4 Gauntlet Blitz Tour Nr 21
score window : [+1.0, +6.0]
nr player aveDepth cntDepth games
1 Sunsetter 9 32bit 15.88 183 24
2 NebiyuAlien 1.45a 32bit 14.55 258 24
3 Imortal 2.9001 32bit 10.52 130 24
4 TJchess 1.3 64bit 10.23 183 24
5 Sjaak II 1.3.1a 64bit 8.87 167 24
6 CrazyWa 0.0.4 32bit 8.62 1336 216
7 KKFChess 2.6.7 beta 32bit 7.38 173 24
8 Sjeng 11.2 32bit 6.80 205 24
9 TSCP zh 1.1 32bit 5.51 112 24
10 Pulsar 2009 9b 32bit 5.12 128 24
-
- Posts: 4833
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Average depth from all blitz games
Code: Select all
file : blitz.pgn
score window : [-6.0, +6.0]
nr player aveDepth games
1 Sunsetter 9 32bit 17.44 915
2 NebiyuAlien 1.45 32bit 15.11 204
3 NebiyuAlien 1.45a 32bit 14.89 794
4 Imortal 2.9001 32bit 11.20 1075
5 Sunsetter 7g 64bit 10.84 296
6 TJchess 1.3 64bit 10.56 1105
7 TJchess 1.1 64bit 10.31 152
8 Sjaak II 1.3.1a 64bit 9.62 899
9 Imortal 2.0 32bit 9.49 295
10 CrazyWa 0.0.1 32bit 9.17 129
11 CrazyWa 0.0.4 32bit 8.89 336
12 CrazyWa 0.0.0 32bit 8.01 96
13 Imortal 1.0 32bit 7.88 94
14 KKFChess 2.6.7 beta 32bit 7.64 452
15 Sjeng 11.2 32bit 7.09 832
16 TSCP zh 1.1 32bit 5.58 431
17 Pulsar 2009 9b 32bit 5.31 451
-
- Posts: 4833
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Win probability at winning score
Win probability when the engine score is already winning [+6.0, +10.0] pawn units.
Code: Select all
file : ccva_tour_nr_19.pgn
event : CCVA Swiss Blitz Tour Nr 19
score window : [+6.0, +10.0]
nr player aveDepth games pts(%) winProb numPos
1 NebiyuAlien 1.45a 32bit 14.64 36 47.22 0.95 39
2 Sunsetter 9 32bit 14.21 35 71.43 0.98 42
3 TJchess 1.3 64bit 9.95 35 77.14 1.00 75
4 Imortal 2.9001 32bit 9.71 35 85.71 1.00 76
5 Sjaak II 1.3.1a 64bit 8.31 35 48.57 0.72 71
6 CrazyWa 0.0.1 32bit 7.81 33 51.52 0.86 42
7 KKFChess 2.6.7 beta 32bit 7.36 36 19.44 0.50 14
8 Sjeng 11.2 32bit 6.35 34 44.12 0.97 65
9 TSCP zh 1.1 32bit 5.28 39 25.64 0.26 94
10 Pulsar 2009 9b 32bit 4.51 38 34.21 0.57 53
-
- Posts: 27808
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: CCVA CrazyWa 0.0.4 Gauntlet Tour Nr 21
I again uploaded a new version (0.06). I finally found a King-Safety term that worked. I had to include a measure for the material in hand for that, something that earlier versions did not have available. But in this version it is kept track of incrementally. This new King Safety makes it much more aggressive.
I tried a breadth-first QS to prevent search explosion, with totally disappointing results. Even in cases were the search explosion was extreme, it did not do anything to avoid it. On the contrary, the search just took far more nodes. I had expected that making sure short refutation would always be seen before long ones would narrow the window on the lines that drag on forever so much that it cold be searched much faster (perhaps even closing the window completely, so that it would not have to be searched at all). But either my implementation was totally wrong, or this is not the case, and the long lines are really score-determining.
So I resorted to a hack, and limited QS to 6 levels. After which I just give a hefty bonus to the side to move. This did improve the results against Sjaak. In difficult positions the depth doesn't drop below 4 anymore (before it could even drop to 2, and then usually a blunder would be played that cost it the game).
This change also reduced the number of crashes. I suppose most of these were asserts triggering because the maximum depth of 100 ply was exceeded, which now cannot happen anymore. Strange enough I still had a number of crashes in the 1000 games I was runing tonight, and they all occurred when CrazyWa had black in the position after 1.d4 ! Unfortunately that does not reproduce easily, because CrazyWa randomizes in the first 5 moves. I guess I should make it print the value of its initial randomizer key (based on starting time), and a way to enter that key inorder to reproduce things. As a new instance of CrazyWa is started for every game it is really strange it already crashes on its first move!
BTW, will CrazyWa have to play games with ponder on, in the upcoming championship? I have not implemented pondering yet, and if it is not needed I can better spend the time on other stuff, as there still is much to improve.
I tried a breadth-first QS to prevent search explosion, with totally disappointing results. Even in cases were the search explosion was extreme, it did not do anything to avoid it. On the contrary, the search just took far more nodes. I had expected that making sure short refutation would always be seen before long ones would narrow the window on the lines that drag on forever so much that it cold be searched much faster (perhaps even closing the window completely, so that it would not have to be searched at all). But either my implementation was totally wrong, or this is not the case, and the long lines are really score-determining.
So I resorted to a hack, and limited QS to 6 levels. After which I just give a hefty bonus to the side to move. This did improve the results against Sjaak. In difficult positions the depth doesn't drop below 4 anymore (before it could even drop to 2, and then usually a blunder would be played that cost it the game).
This change also reduced the number of crashes. I suppose most of these were asserts triggering because the maximum depth of 100 ply was exceeded, which now cannot happen anymore. Strange enough I still had a number of crashes in the 1000 games I was runing tonight, and they all occurred when CrazyWa had black in the position after 1.d4 ! Unfortunately that does not reproduce easily, because CrazyWa randomizes in the first 5 moves. I guess I should make it print the value of its initial randomizer key (based on starting time), and a way to enter that key inorder to reproduce things. As a new instance of CrazyWa is started for every game it is really strange it already crashes on its first move!
BTW, will CrazyWa have to play games with ponder on, in the upcoming championship? I have not implemented pondering yet, and if it is not needed I can better spend the time on other stuff, as there still is much to improve.
-
- Posts: 4833
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Re: CCVA CrazyWa 0.0.4 Gauntlet Tour Nr 21
Thanks for the update.hgm wrote:I again uploaded a new version (0.06). I finally found a King-Safety term that worked. I had to include a measure for the material in hand for that, something that earlier versions did not have available. But in this version it is kept track of incrementally. This new King Safety makes it much more aggressive.
I tried a breadth-first QS to prevent search explosion, with totally disappointing results. Even in cases were the search explosion was extreme, it did not do anything to avoid it. On the contrary, the search just took far more nodes. I had expected that making sure short refutation would always be seen before long ones would narrow the window on the lines that drag on forever so much that it cold be searched much faster (perhaps even closing the window completely, so that it would not have to be searched at all). But either my implementation was totally wrong, or this is not the case, and the long lines are really score-determining.
So I resorted to a hack, and limited QS to 6 levels. After which I just give a hefty bonus to the side to move. This did improve the results against Sjaak. In difficult positions the depth doesn't drop below 4 anymore (before it could even drop to 2, and then usually a blunder would be played that cost it the game).
This change also reduced the number of crashes. I suppose most of these were asserts triggering because the maximum depth of 100 ply was exceeded, which now cannot happen anymore. Strange enough I still had a number of crashes in the 1000 games I was runing tonight, and they all occurred when CrazyWa had black in the position after 1.d4 ! Unfortunately that does not reproduce easily, because CrazyWa randomizes in the first 5 moves. I guess I should make it print the value of its initial randomizer key (based on starting time), and a way to enter that key inorder to reproduce things. As a new instance of CrazyWa is started for every game it is really strange it already crashes on its first move!
BTW, will CrazyWa have to play games with ponder on, in the upcoming championship? I have not implemented pondering yet, and if it is not needed I can better spend the time on other stuff, as there still is much to improve.
The champ tour is ponder off.
-
- Posts: 2929
- Joined: Sat Jan 22, 2011 12:42 am
- Location: NL
Re: CCVA CrazyWa 0.0.4 Gauntlet Tour Nr 21
I treat in-hand pieces as always threatening squares near the King. Maybe with a condition that they can be dropped safely, but I don't think so.hgm wrote:I again uploaded a new version (0.06). I finally found a King-Safety term that worked. I had to include a measure for the material in hand for that, something that earlier versions did not have available. But in this version it is kept track of incrementally. This new King Safety makes it much more aggressive.
Interesting. The last time I tried something like that it was a complete disaster.So I resorted to a hack, and limited QS to 6 levels. After which I just give a hefty bonus to the side to move. This did improve the results against Sjaak. In difficult positions the depth doesn't drop below 4 anymore (before it could even drop to 2, and then usually a blunder would be played that cost it the game).