GIPFboard study

Benchmarks

Stronger than every other GIPF program we could find and run.

We searched for every other GIPF-playing program, got the ones that work running, and played full matches against them. The machine on this site won all of them, by a wide margin.

Head to head

Match results

Olympiad champion

GF1 “Gipf for One” 1.03

Against its search depths 3 to 6191 games, early version of our AI
190–1
Against its strongest tested setting, depth 6final version at 0.3 s per move; GF1 took about 8 s per move
12–0
our winsour losses

Online game server

BoardSpace.net robots

Against all four robots246 games overall
246–0
Against Bestbotfinal version at 0.3 s per move
24–0
Against Bestbot, with a handicapour AI limited to 0.1 s per move
23–1

Handicap test

How little thinking does it need?

Against ourselves

Our previous site model

+1,700Elo

The rest of the field

Other GIPF programs

These exist, but could not be played against. Here is why.

  • LudiiIts GIPF implementation is unfinished and has no capture rules (source).
  • GipftedRunner-up at the 2001 Computer Olympiad. Not available for download.
  • Hobby projectssooryan/gipf-ai, frsfnrrg/gipf and a few others on GitHub. Unfinished, or written for simplified rules.

Know a GIPF engine we missed? Open an issue.

Method

How the matches were run

Each match pitted the real browser engine against the other program, with colours alternated and varied openings. After every move the complete board was compared between both programs, so differences in how they read the rules could not skew a result. The details are in the full technical report.

Fine print: no games against human experts yet · GF1 was not tested beyond depth 6 · standard GIPF rules.