Leela Bots Leaderboard

New binary by Lucario6607 which allows the two hybrid search methods to be blended in a smooth manner rather than with an abrupt transition. A logarithmic formula was used. Testing shows a 21 elo gain after nearly 800 games with each version against four different "human" bots.

New t3V2 net by @Ira_Jed. Testing showed a 73 elo gain over the current bot net based on 600 total games against three "human" bots. So far it won the first twelve games it played vs humans.

Increase to temp settings that happened with the switch of LQO to onnx-trt was reverted as results were worse after that change

Switched from CUDA to ONNX-trt once problems with the latter were resolved. This should reduce average time per move by about 30%, which should help in bullet chess by reducing human ponder time.

New binary by Lucario6607, which modifies search contempt by blending contempt search with normal search above the SC limit, which was lowered from 60 to 55. Node limit was set back to 60000 as that is what was tested. Testing showed that when giving queen odds results improved about 50 elo against four different human-derived nets, The new net remains in place. The bot is still running CUDA until problems running the new binary with onnx-trt are resolved, which should give a substantial further speedup. It is likely that the new binary will end up in the other bots as well once testing confirms that it is good at those odds.

Net updated from t3v2to t1v1, by @Ira_Jed. Testing showed that t1v1 performed 55 elo better against 4 opposing "human-like" bots. This net is so far scoring better against all levels of human opposition than earlier nets.

Net updated from t3v1 to t3v2. Both were trained by @Ira_Jed and used games generated by @Lucario6607. Testing showed that v2 performed significantly better against 3 out of 4 opposing "human-like" bots. As compared with Noah's V2, the new version should score better against opponents strong enough to break even or better at the odds, but may not score as overwhelmingly against weak opposition.

Using new net created by @Ira_Jed and @Lucario6607.

Revert the special black net after Hardware switch and move back to 4090 binary, since there were some issues with the settings.

Revert Net back to lqo_v2 by @Noah13 since real-world performance was worse then tests.

Move the bot to a different hardware (RTX 5090), including a special binary for speedup on that hardware.

Switching Net to experimental t3d version, which looked good in various tests.

Using specially trained net by @Noah13 when LQO plays Black.

Using new search-contempt binary by @amjsh, which boosts performance against humans by ~200 elo. This leaderboard shows all the games from this point on.

Raising the node limit of LQO to 3000.

Updating the Bot to use specially trained queen-odds net (lqo_v2) by @Noah13 instead of contempt implementation.