1985 Usenix Computer Go Tournament

The second Usenix Computer Go Tournament was held June 12 at the Summer 1985 Usenix meeting in Portland, Oregon. This was the second full-board computer go competition ever to be held (the first being the previous year's Usenix tournament). The tournament organiser was Peter Langston. There were five entrants, written by three programmers.

Ogo just mirrored the opponents move, playing the center if it went first. Unfortunately, some last minute "enhancements" caused it to commit suicide with rhythmic regularity. Oog was an improved version of ogo, algorithm unknown. Both were by Peter Langston.

Gorilla and Goanna used a pattern-matching strategy that required an astonishingly small amount of cpu time; the game between these two programs was true speed go. They scanned the board looking at NxN (N = 3? 4?) squares, and choose the "biggest" play. Gorilla was designed to play a more cutthroat game in which killing opposition stones was a paramount goal. They had the bad habit of filling in their own eyes (even if there were only two to fill!). This bad habit, as well as ogo's persistence in not passing, caused Nemesis to trip over an internal error (out of buffers, or some such) and drop core after about 300 moves. How do you score a game where one of the players goes catatonic? [The catatonic player loses on time, as a catatonic human would, surely?] They were written by Bruce Ellis.

Nemesis was the continuation of the work of Bruce Wilcox. It was already for sale, for most Unix boxes and PCs. Bruce was willing to work on ports to other systems. It entered the tournament as the reigning champion and clear favorite, having demolished the opposition in the previous year and having played well in a recent tournament with humans. It was actually the strongest program, playing somewhere around 20 kyu. There was supposed to be a 15 kyu version in the works. Although judged to be the strongest entrant, it seemed to have trouble with the tense tournament atmosphere and crashed frequently (in 6 out of 8 games), thereby threatening its domination of the standings.

Peter Langston and Bruce Ellis of Gorilla and Goanna were present in person, Bruce Wilcox was not.

Each program played each of the others other twice, as black and as white. The results are presented in the foilowing table.

Black\White   oog   NemesisGorillaGoannaogoPoints
oog 011ogo suicided10
Nemesis1 Nemesis stopped playingNemesis stopped playingNemesis did a core dump-2
Gorilla0Nemesis stopped playing incompleteincomplete-6
Goanna0Nemesis stopped playingincomplete incomplete-6
ogo0Nemesis did a core dumpogo suicidedogo suicided -10

The "points" column is explained by the unusual scoring system:

[ This method of scoring may seem harsh. But I have organised computer Go events, and can sympathise with the wish to punish misbehavior. ]

It was far more difficult to judge the results in this tournament than in an ordinary tournament. Should a loss due to a blatantly illegal move be comparable to a loss from falling into a complete catatonic stupor and forgetting to move at all? How should either compare to uttering "out of buffers" before discretely resigning? The judges finally decided to pay most attention to genuine go competence. The results were then:

Here "Other" refers to all the ways a game can end without a proper result.

Thus oog was declared the champion, coming from complete obscurity to snatch the title from the favorite.

Oog was the second strongest player, being beaten only by Nemesis. The Oog-Nemesis games were the only ones that actually looked like Go games. There were rumors that Oog was going to be posted to the net. Both of oog's losses were to Nemesis; one of these games actually had a strong resemblance to a go game. The other genuinely interesting game was a practice game between Nemesis and itself; it was a close contest but finally black gave up the ghost and fell into a silent torpor.

The game of go is difficult to play well. A sense of broad strategical issues is very important; apparently none of the programs here used the familiar tree search techniques so popular for handling tactical situations. It seems that Nemesis plays at about the level of 20 kyu judging both from its performance here and its performance in an earlier tournament in which its opponents were human. This is substantially above the level of a beginning player but still very, very far away from being able to beat players drastically weaker than the program's author.

This information was provided by David Fotland; and by Bruce Wilcox,
who obtained it from net.games.go, where it was posted by "I'll be
mellow when I'm dead" (mwm@ucbopal.CC) and by "roland@inmet.UUCP".

Other computer Go Tournament results