Tweak beta update formula + Tuning #5039
Conversation
|
Congrats, but always as some additional code change is combined with a big tuning a verification should be done that this is not slipped in only of the tuning. So here is the simplification test based on this PR: https://tests.stockfishchess.org/tests/view/65c3f43cc865510db0283ec4 |
|
@locutus2 since this came from @cj5716 he already had a passer stc, If there isn't, please remove it and resubmit the patch minus the beta update vs patch at LTC non regression bounds. Aka locutus test but just as LTC... and please dont mix in random patches into one in the future, unless there proof that these only pass together. |
|
Here is the requested test: https://tests.stockfishchess.org/tests/view/65c66bb8c865510db0286bbe So If I understand correctly, if this test fails, we will merge the patch as it is. |
|
@Disservin the LTC test is taking forever (fluctuating between -0.4 to -0.6 Elo), I think that now we have a clear indication (+500k STC games and +250k LTC games) that simplifying the beta update part would not be as easy as one might assume, so can we please proceed in merging the PR? |
Search parameters were tuned using 16k games at VVLTC. They were tuned starting with the new parameters (in search only) of PR #5039. Passed VVLTC: https://tests.stockfishchess.org/tests/view/65c8a8fc1d8e83c78bfcd163 LLR: 2.94 (-2.94,2.94) <0.00,2.00> Total: 20826 W: 5355 L: 5100 D: 10371 Ptnml(0-2): 1, 1941, 6275, 2194, 2 Passed 2nd VVLTC: https://tests.stockfishchess.org/tests/view/65cadc2d1d8e83c78bfcfdaf LLR: 2.94 (-2.94,2.94) <0.50,2.50> Total: 17710 W: 4611 L: 4352 D: 8747 Ptnml(0-2): 1, 1586, 5422, 1845, 1 STC Elo estimate: https://tests.stockfishchess.org/tests/view/65cb6aed1d8e83c78bfd0802 Elo: -1.46 ± 1.8 (95%) LOS: 5.5% Total: 40000 W: 10267 L: 10435 D: 19298 Ptnml(0-2): 200, 4860, 10023, 4742, 175 nElo: -2.77 ± 3.4 (95%) PairsRatio: 0.97 Bench: 1198939
Search parameters were tuned using 16k games at VVLTC. They were tuned starting with the new parameters (in search only) of PR official-stockfish#5039. Passed VVLTC: https://tests.stockfishchess.org/tests/view/65c8a8fc1d8e83c78bfcd163 LLR: 2.94 (-2.94,2.94) <0.00,2.00> Total: 20826 W: 5355 L: 5100 D: 10371 Ptnml(0-2): 1, 1941, 6275, 2194, 2 Passed 2nd VVLTC: https://tests.stockfishchess.org/tests/view/65cadc2d1d8e83c78bfcfdaf LLR: 2.94 (-2.94,2.94) <0.50,2.50> Total: 17710 W: 4611 L: 4352 D: 8747 Ptnml(0-2): 1, 1586, 5422, 1845, 1 STC Elo estimate: https://tests.stockfishchess.org/tests/view/65cb6aed1d8e83c78bfd0802 Elo: -1.46 ± 1.8 (95%) LOS: 5.5% Total: 40000 W: 10267 L: 10435 D: 19298 Ptnml(0-2): 200, 4860, 10023, 4742, 175 nElo: -2.77 ± 3.4 (95%) PairsRatio: 0.97 Bench: 1198939
|
This is outdated now, I will close it, to save time for the maintainer. |
Tweak beta update formula : Thanks to @cj5716 for this idea, and if possible, add him as co-author of this PR
Tuning various parameter : manually chosen by @FauziAkram based on the observations derived from the tuning patches:
https://tests.stockfishchess.org/tests/view/65ba41c2c865510db02782d8
https://tests.stockfishchess.org/tests/view/65ba418dc865510db02782ce
Passed STC:
LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 123296 W: 32261 L: 31807 D: 59228
Ptnml(0-2): 515, 14684, 30841, 15048, 560
https://tests.stockfishchess.org/tests/view/65c2a84dc865510db0282484
Passed LTC:
LLR: 2.95 (-2.94,2.94) <0.50,2.50>
Total: 34308 W: 8801 L: 8482 D: 17025
Ptnml(0-2): 21, 3764, 9263, 4087, 19
https://tests.stockfishchess.org/tests/view/65c368e9c865510db02833da