search

LeelaChessZero / lc0

The rewritten engine, originally for tensorflow. Now all other backends have been ported here.
GNU General Public License v3.0
2.37k stars 523 forks source link

Policy tournament #1962

Closed Tilps closed 5 months ago

Tilps commented 5 months ago

Note that in preparing this for submission I renamed a number of things from PolicyMultiGames to MultiGames.

The intention is that after this is merged, MultiGames will get a replaceable multigame evaluator which is used to switch between policy and value modes - allowing for a cleaner merge of value_tournament, without massively redundant code.

This update preparing for merge has also changed the behaviour of policy tournament a bit. Since the default value of the flag kTotalGames is -1, which doesn't make sense for policy tournament, you are now required to always provide a total games parameter if using multi game mode where before the code treated -1 like -2 and didn't handle -2 correctly at all.

Some additional checks have been added, no trying to play an odd number of games in multi-game mode. Use of an opening book is now enforced, not just encouraged.

Also note the cuda/cudnn changes are obsoleted by PR1961 which I'll aim to submit first.

Tilps commented 5 months ago

PR1961 merged and thus this is now ready to be reviewed and maybe submitted...