Pikafish test data
What happens when the same engine with the same configuration plays itself?
Self-play data for the 20240314 GitHub version of Pikafish.
Test 1
- Threads: 1
- Hash: 256MB
- Time control: equivalent to 60 seconds + 0.6 seconds on a normal core, no pondering.
- Openings: random 3-move cloud-database openings from 2022 at ±10 points, the vast majority balanced, over a hundred thousand positions in total, paired games.
Results:
- 8438 games in total, 152 of them decisive, draw rate 98.2%.
- 81 wins for the side moving first, 71 wins for the side moving second.
- Average 248 plies per game (124 moves).
Among the wins:
- 116 were checkmates
- 35 were stalemates
- 1 was a perpetual check (which can be regarded as a checkmate; when the number of moves by which perpetual check loses equals the number of moves to being checkmated, Pikafish may choose perpetual check.)
Among the draws:
- 4592 were drawn by threefold repetition
- 3299 were drawn by the 60-move rule
- 395 were automatic draws
Test 2
The same test, changing only the openings: 3-move Red high-advantage openings, where each position's draw rate should be between 40% and 80%, over a million positions in total, paired games.
Results:
- 5914 games in total, 2512 of them decisive, draw rate 58.54%.
- 2427 wins for the side moving first, 25 wins for the side moving second.
- Average 320 plies per game (160 moves).
Among the wins:
- 1698 were checkmates
- 749 were stalemates
- 5 were perpetual checks
Among the draws:
- 2245 were drawn by the 60-move rule
- 995 were drawn by threefold repetition
- 222 were automatic draws
Across these openings:
- One win plus one draw occurred 1066 times
- Two wins occurred 6 times
- Ties occurred 1885 times
- The Ptnml tie rate was 63.75%