BGB v GNU Methodology

I had BGBlitz play a series of 7-point matches against itself using the following settings:

Engine: TachAI II / Torino (see note)
Level: Expert
Search depth: 3 (equivalent to GNU 2-ply)
Search filter: normal
Cube policy: normal

Note: The early matches (through position 40) were played in BGB version 2.1. Later matches were played in version 2.2.1, where the TachAI II engine has been renamed Torino. According to Frank Berger, author of BGBlitz, Torino includes some bug fixes and minor tweaking, but is essentially the same engine.

I then analyzed the matches using GNU 0.16, set to "world class" (2-ply) for cube and checker play.

For each disputed position, I did a "forced" 2-ply eval of all candidate plays, followed by a 1296-trial, 0-ply rollout of all candidates withing 0.100 of the top play. Finally, the bot's choices, along with any plays that performed better in the initial rollout, were rolled out using GNU's "world class" settings. These rollouts were a minimum of 1296-trials, and were extended if necessary to reach 2 JSDs. Positions where the equity between the bots' choices was less than 0.020 were discarded.

BGB v GNU Methodology

Copyright 2007 GammonQuest