Fuego at the Computer Olympiad in Pamplona 2009: a Tournament Report

Size: px

Start display at page:

Download "Fuego at the Computer Olympiad in Pamplona 2009: a Tournament Report"

Cory Harris
5 years ago
Views:

1 uego at the omputer lympiad in amplona 009: a ournament eport artin üller ay 8, 009 bstract he 14th omputer lympiad took place in amplona, pain from ay 11-18, 009. he uego program won the 9 9 o competition, and took second place in o. his report provides some analysis of the games played by uego in the competition. 1 Introduction his was the second lympiad for uego. In its first participation, 008 in eijing, uego ended up in fourth place in both 9 9 and his year s competition took place in the beautiful historic alacio del ontestable in the center of amplona. omi was 7.5 and hinese rules were used in both tournaments. he uego team, software and hardware were as follows: arkus nzenberger is the creator of uego and its lead programmer. artin üller is team leader, programmer and opening book author. roderick rneson is a programmer. erry esauro and ichard egal of I Watson esearch contributed the distributed memory implementation. he software used was the current svn version of uego, with two experimental additions: irst, the distributed memory implementation allowed the program to run on a cluster of eight 8-core machines provided by I and operated by ich egal. econd, for the competition, an experimental version called uegox was used. his version of uego employs knowledge of the classical o program xplorer for move pruning in core parts of the U tree. 9 9 ournament he 9 9 competition was played on ay programs took part and played a double round robin. he time limit was 30 minutes per program. omplete tournament results and many game records are given on fr/icga/tournament.php?id=194. echnical eport 09-09, Version. ay 009. ept. of omputing cience. University of lberta. anada. ll rights reserved. 1

2 igure 1: ound 1, uego - Yogo. ound 1: uego - Yogo :0 igure 1 shows the games of round 1 against Yogo. Yogo is a strong opponent which took third place in this competition. ast year in eijing, uego and Yogo had split their mini-match with one win each. n the left in igure 1, Yogo (White) - uego (lack). Up to White 4 is a common opening. reviously lack 5 at 7 was in uego s book, but it led to a poor winning percentage. ove 5 at 6 was a new move prepared for this competition, and it worked perfectly in this game. White replied passively and lack got an overwhelmingly strong position. 18 is a desperation move. he white group 18-4 can be killed, but uego plays safe with 5 and wins on territory. 35 moves. lack wins by resignation. n the right in igure 1, uego (White) - Yogo (lack). Up to 6 is in uego s opening book. 7 is unexpected, usually lack would connect. White 8 leads to a long tactical sequence. White 1 is a solid move that leads to a very close game. White has a safe half point lead in the end. 8 moves. White wins by resignation. ound : uego - it :0 igure shows the games against the hinese program it. n the left, it (White) - uego (lack). 37 moves. lack wins by resignation. n the right, uego (White) - it (lack). 38 moves. White wins by resignation. ound 3: uego - udo o :0 igure 3 shows the games against udo o. udo o is ideki ato s program. It used to run on a set of ony playstations, but for this tournament it used more standard

3 igure : ound, uego - it igure 3: ound 3, uego - udo o.

4 hardware - a cluster of 6. n the left, uego (White) - udo o (lack). Up to move 7, the opening seems more promising for White. If White can establish two stable groups, it is often difficult for lack to get enough points to overcome the large 7.5 komi. or move 8, sliding into the black corner at would be more natural. he wedge at move 10 is very surprising but the result up to 16 is good for White. ove 16 at is a nice, restrained move. 17 is a desperation measure. With 4, White offers the top right in exchange for sure life in the top left. White 36 is an all-out move. onnecting at 3 would be easier to understand but 36 is strongest. 5 moves. White wins by resignation. n the right, udo o (White) - uego (lack). uego got a difficult game. or move 5, uego s recent book move was 3, but the results were not promising. o before the tournament I changed back to the older move at 3. his makes it a bit harder for White to invade. fter White 6, the extension 7 is strongest resistance for lack. apturing with 6 is safer, but may not be enough for lack. Up to 14 a complex fight starts. lack 15 is bad shape. ut it is also a kind of testing move. With 16, udo o decided to give way. lack 19 switches back to the corner. lack s play is a bit inconsistent here. If lack blocked at 3 instead, a semeai would result. White 4 is a good move, utilizing the double threat of 4. lack 7 seemed weak. If White lives on both sides, it is hard for lack to win. lack 9 surprised me but it is a good try. White cannot ignore it because of 1. he combination of and 3 makes it hard for White to reduce the bottom left. lack 31 closes off one side. fter 33, the top left corner is unsettled. White can live with 6 points of territory, while lack can create a seki in sente, or a mannen ko in gote. he main question is how big the black corner in the bottom left will become. I believe that White 34 is the losing move, and if White jumped in at it would win. looks thin but it seems to work. y blocking with 35 lack gets enough points to win. fter 54, lack wisely does not try to start the approach move ko. aking seki here is enough. 69 moves. lack wins by resignation. ound 5: uego - lexo :0 uego had a bye in ound 4. Its round 5 opponent was lexo. igure 4 shows the games. n the left of igure 4, uego (White) - lexo (lack). he opening is unclear. ove 1 seems doubtful. efending the other group with 8 or 3 seems preferable. or 13, jumping in at 8 seems more troublesome for White. 14 is an interesting defense. If White lives here the game is won. 3 is a blunder and must be on 6. 4 leaves an approach move ko. he question now becomes whether lack can create enough ko threats. 38, 40 and 4 are very solid. lack resigned without trying the ko. 4 moves. White wins by resignation. n the right of igure 4, lexo (White) - uego (lack). Up to 5, the same opening as against Yogo. White 6 is aggressive. White 14 at 4 leads to a difficult fight. uego s evaluation was pessimistic at 47%. White 16 and 18 are a strong combination. lack 7 is a surprising cut. lack s aim is to take sente and play 3 first. I believe that White 30 should be on 3. With 31, lack gets enough territory. 51 moves. lack wins by resignation.

5 igure 4: ound 5, uego - lexo. eft: 36 at 33. ight: 4 at 16, 40 at 3, 51 at 0. ound 6: uego - oo 1:1 igure 5 shows the games against oo. oth programs were undefeated after 5 rounds, so this promised to be an exciting match. owever, both games were mostly decided by problems with the opening books. n the left of igure 5, oo (White) - uego (lack). lack 5 is dubious, after White s wedge at move 6 lack is in trouble. I looked at various lines after the game, but could not find anything promising for lack. herefore I changed move 5 to 4, as we will see in the final playoff game. ove 7: the other atari on 7 does not work well, either. ove 9: this move looks crazy, since it allows White to capture the center stone on 5. owever, the solid move 4 seems to give White an easy win. Up to 7, lack plays for territory. ut there are too many weaknesses in lack s position. 30, 36 and 40 are strong moves. lack desperately tries to complicate the game with throw-ins such as 63. t this point oo is running quite short on time. White 68 seems to be a time-trouble move. owever, all lines still seem to win for White by at least 0.5 points. uego resigns after move 70. n the right of igure 5, uego (White) - oo (lack). White 8 is uego s last move in book. White 1 is a tesuji. It makes miai of building a position in the top and bottom left corners. pparently oo was in book up to move 13! he oo book is a huge automatically built one. I think it has over.3 million entries. owever, oo s evaluation on leaving the book was only 47-48%. rpad immel of the oo team was very unhappy with this. When lack kept sacrificing stones in the top right, everyone expected a vicious

6 igure 5: ound 6, uego - oo. eft: 16 at 1 33 at 4. assault on the right side White group. ure enough, lack attacked with 1 on. owever, the cut with is hard to deal with. fter 6 White was completely safe. 36 moves. White wins by resignation. ound 7: uego - o Intellect :0 igure 6 shows the games against o Intellect. n the left of igure 6, uego (White) - o Intellect (lack). Up to 10, the two white groups look. White 16 is strange. 5 would work and capture a big corner. White 18 is a blunder. lack 1 tries to attack, but simply capturing on 5 would give a reasonable game. When White lives easily with both groups, lack s territory is too small. 44 moves. White wins by resignation. n the right of igure 6, o Intellect (White) - uego (lack). uego is already out of book after move 4. lack 9 is a very interesting attachment. lack gets a strong position. With 16, white should probably live in the bottom right. 19 is a cool, solid answer to the hane. 63 moves. lack wins by resignation. ound 8: uego - immy :0 I am unclear about the opponent s name. he program was registered as immy, which has been a long time participant in computer o competitions. owever, on and in the interface this program appears under the name coldmilk. he engine message reads: coldmilk version 0.13, his program is designed by the student of immy. igure 7 shows the games against this program. n the left of igure 7, uego (White) - immy (lack). White 8 looks conservative, but it seems hard for lack to surround enough area. his game looks like a solid win

7 igure 6: ound 7, uego - o Intellect. ight: 44 at 10, 58 at 4, 61 at igure 7: ound 8, uego - immy. ight: 51 at 46.

8 igure 8: ound 9, uego - rica. ight: 43 at 33. for White throughout. aybe lack needs to try another opening at move 5 or moves. White wins by resignation. n the right of igure 7, immy (White) - uego (lack). gain, uego is out of book after White 4. lack 15 is surprising. White ignores it and lack captures the group with 17, but much aji remains. he game becomes very complex, and lack has several worrisome weaknesses. fter 5, it looks like the lack position will hold together. aybe White can try to wedge at 5 instead of is a blunder. White can get a ko with 9, 5, 9. ven so, if lack eventually connects at 8 the ko seems too indirect for White. 49 is silly but fortunately White answers it. 61 moves. lack wins by resignation. ound 9: uego - rica :0 igure 8 shows the games against rica. n the left of igure 8, lack 7 is the last book move. he result to 17 feels good for lack. 3 and 5 are good endgame. With 31 lack has a sure 1.5 point win. 47 moves. lack wins by resignation. n the right of igure 8, uego (White) - rica (lack). he opening is difficult. or 11, extending at 3 is an alternative. lack 15 seems soft, since with 18 - White makes a big enough reduction of this corner. lack 15 at, White, lack 7 would be unclear. 4 moves. White wins by resignation. layoff: uego - oo :0 igure 9 shows the playoff games against oo. fter 16 games, oo and uego were tied with 15 wins 1 loss. he playoff rules were as follows: games at first. If

9 igure 9: layoff, uego - oo. eft: 50 at 47, 51 at 8. the score is tied, more games. If the score is still tied, colors are drawn for a deciding fifth game. oth sides made changes to their opening book. oo switched to a different machine-built one. I changed the losing line with lack from the first match in uego. n the left of igure 9, layoff ame 1: uego (White) - oo (lack). he first game follows a familiar script. White has two groups, and lack must attack. he moves up to 8 follow the book. ctually, uego is pessimistic in this opening, but in my judgment White is hard to kill. t move 10 rpad already said that oo would lose. fter White 14, both white groups look safe. or 30, connecting at 3 would simplify the game. White can live. imilarly, 5 should be on 4 and 54 on 4. till, White lives and wins the game. t the end, lack could start a two step ko with 3, 1, 1 but cannot hope to win it. 58 moves. White wins by resignation. n the right of igure 9, layoff ame, oo (White) - uego (lack). ove 5 is a new try after the loss in round 6 with 5. fter 17, the corner looks big visually but it may not be enough to overcome the komi. I believe that with move 18, White is winning. owever, 4 is a simple blunder that throws away the group, the game and the tournament. 8 instead would be a sure win. White goes for complications after this but eventually resigns. 57 moves. lack wins by resignation.

10 ournament 6 programs played a double round robin. ive rounds were played on ay 14, and 5 more rounds with colors reversed on ay 16. he time limit was 1 hour per program for the first 50 moves, and unspecified with a possible referee s decision afterwards. hese rules will be clarified for next year. omplete tournament results and all game records are given on tournament.php?id=193. ound 1: uego (White) - udo o (lack) 1:0 he game is shown in igure 10. uego often plays the one point jump of is an important junction point. 40 would be safer at. 43 is good and takes away White s eye space. White fights very aggressively in lack s sphere of influence. 51 is a surprising tenuki and seems to give White the upper hand in the fighting. 69 is a tactical mistake, since 70 captures the key stones. White has a clear advantage from this point. ound : uego (White) - Zen (lack) 0:1 he game is shown in igure 11. In the early opening, White goes for territory, but the exchange after move 4 reverses the position, with White getting influence and lack taking profit. 3 lacks substance. Up to 45 again White builds influence but gives lack lots of profit. 60 and 68 are slow. 7 loses points is a bad combination. White can connect at is big. 10 is terrible. fter the sequence to 113 the game is hopeless for White. In this game, Zen was clearly the stronger program. ound 3: uego (lack) - o Intellect (White) 1:0 he game is shown in igure 1. White gets into trouble in the bottom left corner. Up to 77, many white stones die. White cuts with 78 to find a chance, but lack captures the key stones with 89 and has a big lead. ound 4: uego (lack) - onot (White) 1:0 onot is acques asaldua s program. It plays a beautiful opening, but is still incomplete, so it has trouble in the later parts of the game. he game is shown in igure 13. ound 4: Zen (White) - oo (lack) 0:1 his round 4 game against oo, shown in igure 14, is Zen s only loss in the tournament. he game looks favorable for White, but things start to go wrong in the top left corner and in the center fight. till, White s early resignation is surprising.

11 ound 5: uego (lack) - oo (White) 1:0 he game is shown in igure 15. he exchange in the top left seems for both. s against Zen, uego gives up territory in the next fight and builds influence. oo invades the bottom with 48 but goes wrong in the continuation. or 50, just extending at 3 would be fine. he resulting fight is good for lack. oth have low number of liberties but lack 67 works well. fter 81, another complex fight starts. ust extending at 6 would be simpler. 91 is the first of three failures by uego to simply capture at makes the cut of 9 possible. 97 is the second miss. With 103, lack should save the three center stones. White s capture with 106 is big. 117 ignores 1 yet again. lack gives up a large territory in exchange for an unclear attack. Up to 18, lack isolates some white stones at the bottom but White becomes extremely strong in the top center. t this stage the game seems favorable for White. 135 is bizarre. 137 tries to cut. Up to 150, the stage is set for the decisive fight. White s chances are better but oo is starting to run short on time, while uego still has over 30 minutes left. 156 is the first bad move by ogo in this fight. ttacking from the right or e.g. pushing on 16 seem better. 16 starts a huge ko fight. White wins it but the bottom changes hands with lack s ko threat on 1. urthermore, lack retains the option of living in the top right corner. he game is very complex but probably better for black from this point. more complex fighting breaks out in the bottom right center. lack manages to live in sente and switch to the huge attacking move 6. he remaining fights do not change the result of this exciting game. With this win by uego, the top three programs Zen, oo and uego were tied for the lead after the first day, with four wins one loss each. ound 6: uego (lack) - udo o (White) 1:0 he game is shown in igure 16. he sequence in the bottom left is worrying for lack. White gets a strong position. he cut at 19 seems to be an overplay. or 5, 9 seems to be the only move. ut White should not play out the loose ladder. Up to 47, the result is unbelievably good for lack. 61 and 63 should just maintain the ladder on is better on is trying too hard. White 100 is a blunder. White loses either 15 or 16 in a ladder. ound 7: uego (lack) - Zen (White) 0:1 he game is shown in igure 17. or 37, may be better. or 39, is safe. 47 is bizarre. 5 would be normal. 51 does not work. lack s loss in this corner is huge. 57 is a move uego loves, but its position is too cramped here. 65 should probably cut at is an overplay and should connect on 15. or 81, 18 is very large. White answer to the cut on 6 is odd. Up to 117, lack gets a large territory here. or 13, 1 is natural. 17 should either continue in the lower right, or come back to is very bad. 145 is risky, 15 is a safe move here. 149, 151 and 175 are bizarre. or 153, 11 is safer. he sequence up to 18 is very big and gives White the lead. 193 is a blunder, White could now kill with is also a blunder. till, getting a seki is

12 enough for White to win. 07 loses more points. he remaining lack moves have no meaning. tough game but in the end a clear win for Zen.

13 igure 10: ound 1, uego (White) - udo o (lack)

14 igure 11: ound, uego (White) - Zen (lack). 47 at 0, 01 at, 3 at 00, 33 at 04.

15 igure 1: ound 3, o Intellect (White) - uego (lack). 145 at 138, 165 at moves. lack wins by resignation.

16 igure 13: ound 4, onot (White) - uego (lack). 108 at 4, 109 at moves. lack wins by resignation.

17 igure 14: ound 4, Zen (White) - oo (lack). 119 moves. lack wins by resignation.

18 igure 15: ound 5, oo (White) - uego (lack). 91 at 50, 114 at 107, 115 at 7, 130 at 46, 13 at 38, 139 at 10, 149 at 140, 168 at 16, 170 at 165, 18 at 175, 191 at 188, 19 at moves. lack wins by resignation.

19 igure 16: ound 6, udo o (White) - uego (lack). 85 at moves. lack wins by resignation.

20 igure 17: ound 7, Zen (White) - uego (lack). 199 at 113, 00 at 11, 10 at 19, 14 at 145, 7 at 10, 9 at 113, 51 at 43, 5 at moves. White wins by resignation.

21 ound 8: uego (White) - o Intellect (lack) 1:0 he game is shown in igure should probably be at 15. Up to 41 White becomes extremely thick, but lack takes territory. or 5, atari on 19 is the obvious move. 61 is equivalent to a pass. With 64, defending the right side would be easier to understand. 67 is big, but White tries to go for a huge center. 7 is the wrong atari and must be on is painful for White. lack gets itself into trouble in the top left. Up to 98 White gets a huge profit here. ow White is better. ut 154, 5, is too deep. White dies in gote here and lets black make eyes in the top left. his makes the game close again. ut then lack ignores the threat of 170, so White wins. rom move 05, the huge block of black stones in the middle remains unsettled until 63 accidentally saves them. 67 is suicidal and makes them unsettled again. fter 78, lack ran out of time, but it had passed the 50 move mark. ince White has a safe lead of over 10 points even if lack s block lives, lack s operator resigned. ound 9: uego (White) - onot (lack) 1:0 he game is shown in igure 19. ound 9: oo (White) - Zen (lack) 0:1 he game that decided the championship, shown in igure 0. It was actually played last, after the end of round 10. he game was favorable for oo in the middle game, but again oo ran short on time, and got outplayed in the fight after move 170. ound 10: uego (White) - oo (lack) 1:0 he last game for uego is shown in igure 1. or this game, oo switched to the uygens supercomputer, oo used 0 nodes with 3 cores each, for 640 cores in total. White 10 is wrong, joseki is 4. Up to 7, lack s profit seems superior to White s influence. or 38, the only shape is ko at looks like a strong attack but 69 is very good. White gets no compensation for the loss of the bottom left. With 75 and 79 it is lack s turn to go wrong. 75 on 7 and 79 on 8 were better. Up to 86 the black stones are cut off. White should answer 105. he fight at the bottom ends with black capturing five stones and White solidifying its position. White 135 and 137 are extremely big. he game is good for White now. 16 moves. White wins by resignation. ppendix: uego echnical etails he program version was svn version 869 and 871. he only differences in 871 are some changes to the 9 9 opening book. he default settings were used, except for the following. lease see the uego documentation for explanations. go_param debug_to_comment 1 go_param auto_save (filename)

22 go_sentinel_file (filename) uct_param_search max_nodes uct_param_player reuse_subtree 1 uct_param_player ponder 1 go_rules kgs sg_param time_mode real uct_param_search number_threads 8 uct_param_search lock_free 1 uct_param_search virtual_loss 1 uct_param_search number_playouts he most interesting of those is probably uct_param_search number_playouts. It was used only on 9 9. he benefit of doing two playouts per leaf is to reduce memory contention, since fewer tree updates are performed. In informal tests before the competition, this was slightly better than the default of one playout. In addition to the standard uego settings above, there were more settings for controlling the experimental code in uegox and the I parallelization. evision istory Version 1 ublished ay 5, 009 Version ublished ay 8, 009. ixed typos and igure numbering. dded references to igures for games.

23 igure 18: ound 8, uego (White) - o Intellect (lack). 60 at 31, 03 at 10, 0 at moves. White wins by resignation.

24 igure 19: ound 9, uego (White) - onot (lack). 8 moves. White wins by resignation.

25 igure 0: ound 9, oo (White) - Zen (lack). 150 at 88, 15 at 7, 8 at 7, 5 at 0, 53 at 06, 67 at 60, 69 at moves. lack wins by resignation.

26 igure 1: ound 10, uego (White) - oo (lack). 160 at moves. White wins by resignation.

Game One: AlphaGo v. Lee Sedol

Game One: AlphaGo v. Lee Sedol Commentary by Antti Törmänen 1-dan Black: Lee Sedol 9-dan White: AlphaGo Lee Date: 9 March 2016 186 moves. White wins by resignation. White 22. With this cap the upper-side