On-line engine blitz tourney October

[hgm]

The monthly on-line engine blitz tourney for October will take place on:

Saturday October 13, 3pm Boston time (21:00 Amsterdam time)

(This assumed daylight-saving time too be still in effect everywhere.) To connect:

winboard -zp -ics -icshost winboard.nl -icshelper timeseal -fcp ENGINE.exe -fd ENGINEFOLDER -autoKibitz

(for UCI engines, add -fUCI)

[op12no2]

Hi hgm,

Could I have a login for Roller please. colin@sailwave.com

Roller is a lot of fun and I created it for my niece who is just learning the game. I tried weakening Lozza in various ways but didn't like the results.

Roller is Javascript and what wikipedia calls a "pure mcts algorithm": from the current position it rolls out complete random move games with no other search structure - and plays the root move with the best net score when time is up. It has absolutely no knowledge other than the rules (but does not know draw by 3 rep); I even refrained from adding that KK is a draw (*). Obviously it's very very weak, but it will beat a true random engine 99% of the time in my tests and from what I can see that 1% is falling for draw by rep. On my machine I get about 2.8k rollouts per second.

Play it in browser (**EATS LOTS OF CPU TIME**) here: http://op12no2.me/toys/roller/play.htm

Source: http://op12no2.me/toys/roller/roller.js


(*) But I do assume a draw after 450 ply because otherwise you can get some very very strange and long games.

[Sven]

Nice work, Colin!

I found one very minor issue: when checkmating by promoting, the display does not change the pawn into the promotion piece.

[Sven]

Regarding strength: would it make any sense to increase the weights (i.e. the frequency of being selected randomly) of all captures, promotions and possibly also checks during a rollout? I have never tried MCTS myself but I could imagine that this might increase the probability of actually losing a rollout game after a bad move.

[hgm]

I am as inexperienced as you in this, but what you propose seems dangerous. The largest part of all rollouts will be unique, the nodes never being visited again. Preferring captures will almost always lead to the Queen capturing a protected piece (mostly Pawns!) before the end of the game, with a high chane of being recaptured. In other words, this bias builds in a compulsive need to sacrifice your most powerful pieces for the opponent's most abundant pieces.

But limiting it to LxH and equal captures could be a good idea.

[op12no2]

Nice work, Colin!

I found one very minor issue: when checkmating by promoting, the display does not change the pawn into the promotion piece.

Thanks Sven and good catch! I suspect that bug is in the Lozza UI as well. I'll fix it.

[op12no2]

But limiting it to LxH and equal captures could be a good idea.

Can't do that as that it's semantic knowledge

[op12no2]

Regarding strength: would it make any sense to increase the weights (i.e. the frequency of being selected randomly) of all captures, promotions and possibly also checks during a rollout? I have never tried MCTS myself but I could imagine that this might increase the probability of actually losing a rollout game after a bad move.

I have some little experiments rippling away in the background so I can try it

[hgm]

Code: Select all

:Tourney Players: Round 9 of 9 
:
:     Name              Rating Score Perfrm Upset  Results 
:     ----------------- ------ ----- ------ ------ ------- 
:  1 -Berta             [2006]  12.5  [2183] [ 107] +09w +16w +06b =02w =03b +04w +05b +11b =07w 
:  2 +ArasanX           [2088]  11.5  [2096] [   0] =03w +12b +18w =01b =06w =05w +16b =04w +09b 
:  3 +Schooner          [1803]  11.0  [2079] [ 837] =02b +07w =04b +12w =01w =06b +10w =05w =11b 
:  4 -nemorino          [2072]  10.5  [1972] [   8] +08w =13b =03w +05w =16b -01b =06w =02b +15w 
:  5 -WaspX             [2002]  10.5  [1991] [  43] =10b +17w +13w -04b +18w =02b -01w =03b +06w 
:  6 +rpiStockfish      [1940]  10.5  [1979] [ 140] +14w +11b -01w +08w =02b =03w =04b +16w -05b 
:  7 -NightmareX        [1997]  10.0  [1761] [   4] -11w -03b =19b +15w =08b =09w +23b +14w =01b 
:  8 -Marvin            [1798]  10.0  [1845] [ 166] -04b +10w +15w -06b =07w -11b +24b =18w +17w
:  9 -RofChade          [1766]  10.0  [1883] [ 235] -01b +19w =12b =11w =13w =07b +17w +10b -02w 
: 10 +Pirarucu          [1764]  10.0  [1815] [ 337] =05w -08b +20w +14w =17b +12w -03b -09w +16b 
: 11 -Phalanx           [1743]  10.0  [1883] [ 540] +07b -06w -16b =09b +32w +08w +18b -01w =03w 
: 12 +Goldbar           [1902]  9.5  [1675] [   0] +15b -02w =09w -03b +20w -10b =19b =23w +22b 
: 13 +Singularity       [1871]  9.5  [1702] [ 100] +23b =04w -05b -17w =09b +24w =14w -15b +20w 
: 14 -Viper             [1663]  9.5  [1732] [ 374] -06b =24w +22w -10b =15b +20w =13b -07b +18w 
: 15 +rpiArminius       [1598]  9.5  [1759] [ 305] -12w +23w -08b -07b =14w +32b +27w +13w -04b 
: 16 +Arminius          [1844]  9.0  [1813] [ 158] +19w -01b +11w =18b =04w +17b -02w -06b -10w
: 17 -tomitankChess     [1646]  9.0  [1735] [ 427] =18b -05b +24w +13b =10w -16w -09b +19w -08b 
: 18 -WaDuuttie         [1933]  8.5  [1670] [   0] =17w +20b -02b =16w -05b +19w -11w =08b -14b 
: 19 +Jumbo             [1533]  8.5  [1646] [ 428] -16b -09b =07w =22b +27w -18b =12w -17b +23w 
: 20 +rpiMinkoChess     [1517]  8.5  [1635] [   5] +27b -18w -10b +31w -12b -14b +32w +24w -13b 
: 21 -Barbarossa        [1494]  8.5  [1470] [  25] -24b -32w -27b +29b +33w =23w +28b -22w +31w 
: 22 -RookieMonster     [1462]  8.5  [1578] [ 151] +31b +27w -14b =19w -23b =26w +29b +21b -12w 
: 23 +Lozza             [1545]  8.0  [1514] [ 178] -13w -15b =32w =24b +22w =21b -07w =12b -19b 
: 24 +Bliep             [1512]  8.0  [1571] [ 101] +21w =14b -17b =23w +31b -13b -08w -20b +27w
: 25 -CTdccc            [1331]  8.0  [1417] [  45] +35w -28b -29w =30b +37w +38b +36w =33b +34w 
: 26 -Weini             [1263]  8.0  [1506] [ 465] =30w =36b +35w +34w +29b =22b +33w -27b +38b 
: 27 +Floyd             [1522]  7.5  [1347] [   0] -20w -22b +21w -32b -19b +31w -15b +26w -24b 
: 28 +Joker             [1431]  7.5  [1325] [   0] -32b +25w -31b -33w +30w +34b -21w =36b +39b 
: 29 +KingSlayer        [1395]  7.5  [1409] [ 105] +33b =31w +25b -21w -26w +35b -22w +40w =32b 
: 30 +Embla             [1294]  7.5  [1361] [ 146] =26b =34w +33b =25w -28b +40w +39b =38w +37w 
: 31 +Spartacus         [1486]  7.0  [1350] [   0] -22w =29b +28w -20b -24w -27b +37b +32w -21b 
: 32 +rpiFruit          [1462]  7.0  [1481] [ 133] +28w +21b =23b +27w -11b -15w -20b -31b =29w
: 33 -Isa               [1422]  7.0  [1291] [   9] -29w +35b -30w +28b -21b +39w -26b =25w +40b 
: 34 +microMax          [1215]  6.5  [1298] [  39] +36w =30b +37w -26b +38w -28w +40b +39w -25b 
: 35 -Eichhoernchen     [1323]  6.0  [1175] [   0] -25b -33w -26b +39w +40b -29w +38w -37b +36w 
: 36 +rpiPhalanx        [1175]  5.0  [1173] [ 172] -34b =26w +40b -38w +39b +37w -25b =28w -35b 
: 37 +Skiull            [1157]  4.0  [1140] [ 166] =38w +39b -34b +40w -25b -36b -31w +35w -30b 
: 38 +rpiEmbla          [1146]  3.5  [1072] [ 108] =37b +40w -39w +36b -34b -25w -35b =30b -26w 
: 39 +rpiBikjump        [1047]  2.0  [1023] [  99] +40b -37w +38b -35b -36w -33b -30w -34b -28w 
: 40 -Skipper           [1044]  0.0  [ 841] [   0] -39w -38b -36w -37b -35w -30b -34w -29b -33w
:
:     Average Rating    1580.3 
:

[hgm]

But limiting it to LxH and equal captures could be a good idea.

Can't do that as that it's semantic knowledge

I guess you can disguise it as a game move, though: suppose you alter the game such that instead of just starting to moves, the players decide on a strategy first, and only then start moving, sticking for that strategy for the rest of the playout. The strategy is a 5x5 matrix, where each element corresponds to an (attacker type, victim type) pair, and can have two values (high weight or normal weight). Each turn they start with all elements set to normal. The first 10 moves of each player must be used to set the value of a strategy element of their choice to high weight, or a turn pass. Then the rest of the rollout is done by selecting moves according to that strategy.

I would be surprised if the effect on the score of setting capture of a Queen by a lower piece to high weight would not have a far larger effect on the average score than picking the best move in the current position. So MCTS should discover very rapidly which captures are good, and then only exploit that strategy in the 'tactical part' of the rollout, wasting very little time on rollouts with poor capture strategies. (Note that the strategy can adapt later, e.g. if it gets in KNNPKB it will discover that the all-important thing is to capture the Pawn.)

I kind of like that approach to game playing: think before you act!