Overview

Overview

p3achygo is a AlphaZero-based Go engine, able to reach 5-6D after playing 500k games, for about a week of training time on an A100. The most recent run is currently ongoing. I have not currently made models/SGFs public, nor have I integrated with GTP protocols. These should all be coming soon :).

Status

Current Run: v3 (8-15-2023 - Ongoing)

Games Played: 910,000

Current Strength: Coming Soon

Training Efficiency

To measure p3achygo’s training efficiency relative to the original AlphaGo Zero, we can compare p3achygo to Leela Zero. Leela Zero is an almost 100% faithful reproduction of the original AlphaGo Zero, with minor differences in visit count per move and MCTS algorithm. At 256 visits, p3achygo is slightly stronger than LZ-066 with 1000 visits. LZ-066 played ~3.2 million games, for an approximate lower bound of ~1.4T queries (Leela used 3200 visits per move). By this measurement, p3achygo is ~140 times more efficient than LZ in early training. The original AlphaGo Zero ended its 20-block run after ~2T queries, and 4.9 million games.

For more docs on runs and experiments, see runs.md.

Sayuri

Sayuri is another Go Engine that uses similar techniques as p3achygo, plus some. Sayuri seems to be capable of even better training efficiency than p3achygo. Their results are remarkable and I hope to see where their current run lands :)