I've been experimenting with GNU for a few weeks and I have some questions about the "right" settings. By default, gnubg plays and analyzes at the expert level. I've played with those settings for a while, thinking it was enough.

Then, I set the analysis level to its highest level, grandmaster. Analyzing a game now takes about between 30 sec and 1 min, depending on the number of moves, instead of being a question of a few seconds. I imagine that if I set the play level to grandmaster as well, it would take gnubg more time per move.

The question is, is it worth it? What's the ideal settings according with your experience with it?

P.S. Little theorical puzzle: How quick (number of moves) can you close your board, assuming you don't get hit and there's no anchor in your home board?

"expert" (0-ply) is somewhat similar to a very strong player. Its checker play is fairly strong, but not fantastic; its cube action tends to be somewhat off. Though it doesn't play exactly like a 'human'.

"world class" (2-ply) is the standard setting. Cube action here is much better, and checker play is also significantly stronger. "supremo" (2-ply with expanded move filter) will find a move that WC filter didn't see once in a while but takes significantly more time, especially if you do rollouts.

"grandmaster" (3-ply) is a further setting beyond that. Cube action is marginally weaker, but checker play is marginally stronger. But the time increase over 2-ply is substantial.

Standard setting is 2/2, that is, the "world class" preset.

Thanks a lot.

For the puzzle, I meant what is the minimal number of moves to close the board with the opponent's checkers removed (so no hit and no anchor in our board).

Since I didn't gather a lot of buzz from my problem, I'll simply post the answer here : 4. I'll let you figure how.

I've also gotten into the habbit of doing a rollout in GNU any time the 2 best plays are reasonably close-- say less than .01 apart in equity, and sometimes more if the position is complex or if I'm confused by the suggested play. It is quite often the case that GNU changes its mind.

For speed, I favour a rollout truncated at 10 moves. The simulation is esentially converging to what would in theory be the 10-ply evaluation. I usually stop the simulation manually when the standard deviation of the best move is less than half the equity difference between the best and second best moves; using the empirical rule from statistics, this very roughly corresponds to 95% certainty that the ordering of the top two moves is correct. This works pretty well in most situations, and can let you get a resonably good rollout in a minute or two.

If the plays are less than 0.01 apart in equity, does it really matter which one is better, aside curiosity? I mean, if the second best play (really close to 1st) will lead to a type of game that you're more familiar with (e.g. blot-hitting contest vs mutual holding game, etc.) or is a play that you think will lead your opponent to more errors later in the game, then why not?

Thank you in advance for any feedback.

If the *true* equity is less than 0.01 then I think you're right it doesn't matter, but I find that plays that the 2ply evaluation thinks are close, but that are doing completely different things, are often not-so-close when you do the rollout.