The 1992 US Computer Go Championship - by David Fotland
This year the US Computer Go Championship was held August 2, 3, and 4
at Williamette University in Salem Oregon. This was the 5th Computer Go
Championship held in the USA at the Go congress. Previous winners were:
1988 - Cosmos (Many Faces of Go)
1989 - Go Intellect
1990 - Go Intellect (actually won by Goliath, a program from Europe)
1991 - Many Faces of Go
This year's results were:
1) Many Faces of Go, by David Fotland
2) Go Intellect, by Ken Chen
3) Nemesis, by Bruce Wilcox,
4) Contender, By Lynn Beus
5) Poka, by Howard Landman
6) Prototype, by Arthur McGrath
7) Fumiko, by Peter Dudey
With 7 programs entered there was time to play a full round robin with one
game between each pair of programs. Time limits were 1 hour to play 125
moves for each program, with the game adjudicated after move 250 if the
programs were too slow or the games were too long. Ing's SST Laws of Wei Chi
were used since these are the rules used by the World Computer Go Championship.
None of the games had to be adjudicated.
Five programs have participated in this tournament
several times before: Many Faces of Go, Go Intellect, Nemesis,
Contender, and Poka. Go Intellect has several second place finishes
in world competition and was the favorite to win this year. Many Faces of
Go was based on the new second release that had been completed the week before,
and is about 3 stones stronger than last year's Many Faces of Go. Nemesis
was part way through a major overhaul to allow him to add full board
lookahead, so Bruce was not expecting to do very well.
Poka was pretty much the same program as last year.
There were two new programs this year, Fumiko and Prototype. Although
neither of them beat any of the established programs, both of their
authors were inspired by the competition, and plan to do much better
next year.
Ken chose white for Go Intellect whenever he could because the tournament was
played by the Ing SST Laws of Wei Chi with an 8 point komi. His
opponents all chose black when they had the choice of colors, so Go
Intellect played white in every round.
I knew that Go intellect would be my toughest competition this year. Many
Faces beat Go Intellect last year only because Go Intellect got into time
trouble and stopped reading and dropped two big groups in the endgame.
Before that it was 100 points ahead. Last year's Many Faces was
definately weaker than Go Intellect. I didn't expect to have any problem
beating Nemesis, Contender, Poka, or the new programs.
I had hoped to have some time to tune Many Faces against Go Intellect
before the competition, but the second release for Ishi Press was a little
late. I delivered the final Many Faces to Ishi on July 27, after 3 months
of intense bug fixing, and spent the rest of the week relaxing in front of
the TV watching the Olympics. On Friday I played a test game between
Many Faces and last year's Go Intellect, and Many Faces lost by more than 30
points. I fixed problems in this game from Friday night right up until
the game against Go Intellect on Monday afternoon.
Unfortunately, one of the changes I made before the round against Contender
was incorrect and caused the program to crash, so Many Faces lost to
Contender. Programs can be restarted after a crash, with the clock
still running, but cannot be modified during a round. Many Faces survived
two restarts, but the third time it crashed for good. I fixed the bug after
the round and played Go Intellect the next round.
Go Intellect and Many Faces were the only two programs that do full board or
life and death reading, and they crushed all other opponents. Many Faces
beat Nemesis by over 140 points. The game between Go Intellect and
Many Faces was very close. Go Intellect was ahead by about 3 points in the
late endgame when it tried to live inside a large secure
territory. It gave up a lot of points and lost by 11. Go Intellect and
Many Faces ended up tied with 5 wins each, and tied on sum of opponent's
scores. Since Go intellect lost to Many Faces and Many Faces lost to
Contender, Many Faces won on sum of defeated opponent scores.
Many Faces and Nemesis were the only programs to implement the Standard
Go Modem protocol, so only one game was played over a cable. All other
games required the authors to act as operators. (A copy of the Standard
Go Protocol specification, and sample code from Many Faces of Go and Nemesis,
is available from either Bruce Wilcox or myself).
Howard was also tuning up Poka during the tournament between rounds.
In the first round Fumiko had an incorrect position on the board and
couldn't correct it, and so forfeited. They continued playing, and Fumiko
suicided. Poka wouldn't accept it, but had already won. Howard fixed
Poka to accept suicide, but unfortunately did it wrong, and later lost to
Nemesis when it took both the capturing and captured stones off the
board in a ko fight.
Contender looked quite a bit better this year. The game between Contender
and Many Faces looked close before Many Faces crashed. After I fixed the
bug we completed the game and Many Faces killed a big group and won. Contender
got way ahead of Poka, but gave back a big dead group in the endgame and
lost. Poka would have finished above Contender if Many Faces hadn't
crashed.
Prototype plays a large moyo game, and looked very good against Many Faces
in the opening, but let most of its stones get captured in the middle game.
It also made a huge moyo against Nemesis, but defended the boundary incorrectly
and let Nemesis break through the wall. The game between Poka and Prototype
was interesting since both programs make big moyos and both are weak at
tactics and life and death. It came down to a big group of Poka's which had
a single 3 space eye. Prototype didn't kill it, and Poka didn't make it
live since Poka thinks groups only need one eye to live. Finally after all
dame were filled, Poka played to make two eyes as a safety play (since in
Chinese rules it doesn't cost to play inside your own territory) and won.
Fumiko was written during the summer and was a very young program. It crashed
or ran out of time in its first four games, and forfeited the last two
without playing. It was running on my portable 386 SX, which was not
fast enough to allow it to finish its games.
I borrowed an HP9000/720 "Snake" workstation to run Many Faces, but ended
up not using it. I couldn't duplicate the programming environment from my
PC onto it, and since I was making changes up to the last minute I traded
machines with Howard. He used my Snake, and I used his 40 Mhz 486. I played
all tournament games at level 15. At one point Howard played two games
simultaneously on the Snake.
1 2 3 4 5 6 7 Wins
1 Go Intellect X w-loss w-win w-win w-win w-win w-win 5
2 Many Faces b-win X b-win w-crash b-win w-win win 5
3 Nemesis b-loss w-loss X w-win b-win w-win w-win 4
4 Contender b-loss b-win b-loss X w-loss w-win win 3
5 Poka b-loss w-loss w-crash b-win X w-win b-win 3
6 Prototype b-loss b-loss b-loss b-loss b-loss X b-win 1
7 Fumiko b-loss forfeit b-loss forfeit w-loss w-loss X 0