Randomly use classical eval #3107
Conversation
|
We already use classical eval for large imbalances so this patch is a natural extension of that idea to smaller imbalances albeit we do it randomly. And of course, we only do it rarely. Also this patch essentially passed twice since the contempt tweak is just about neutral STC: (with contempt=12) LTC: (with contempt=12) I initially thought that the contempt change is necessary for this to work but the tests suggests that it is not. |
|
Who is going to update the evaluation guide to specify that evaluations are nondeterministic? |
|
Does the "analyze" feature benefit from this kind of patches? or is this just for games? |
|
it does @FauziAkram |
|
@FauziAkram I would say, of course.... makes the engine find better moves more quickly, that ought to be good for analysis as well. @ddugovic this is not so different from e.g. dynamic contempt where the evaluation result depends on the search state. |
If there is a moderate imbalance, use classical eval with small probability (1/16),
as derived from the node counter.
STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 32320 W: 3562 L: 3377 D: 25381
Ptnml(0-2): 144, 2609, 10478, 2776, 153
https://tests.stockfishchess.org/tests/view/5f520615ba100690c5cc5f80
LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 21032 W: 1116 L: 974 D: 18942
Ptnml(0-2): 20, 837, 8664, 971, 24
https://tests.stockfishchess.org/tests/view/5f522eaaba100690c5cc5f8c
closes official-stockfish/Stockfish#3107
Bench: 4109324
|
Is this safe for search stability? |
|
@gonzalezjo |
8 participants
If there is a slight to moderate imbalance, use classical eval randomly for 1/16 or around 6.3% of the time.
STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 32320 W: 3562 L: 3377 D: 25381
Ptnml(0-2): 144, 2609, 10478, 2776, 153
https://tests.stockfishchess.org/tests/view/5f520615ba100690c5cc5f80
LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 21032 W: 1116 L: 974 D: 18942
Ptnml(0-2): 20, 837, 8664, 971, 24
https://tests.stockfishchess.org/tests/view/5f522eaaba100690c5cc5f8c
Bench: 3696738