Scale exploration with Gini impurity of policy values #44
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Bench: 1557738
Bench: 4361350
Bench: 4361350
Bench: 1490523
Bench: 1317423
Bench: 1317423
Viren6
pushed a commit
that referenced
this pull request
Aug 16, 2024
First, when computing policy values in a position, we also calculate the Gini impurity, defined as (1 - sum of squares of policy values). A high Gini impurity indicates that there are many strong candidate moves in a position, and vice versa. This Gini impurity is then used to adjust the exploration scaling using a logarithmic formula. For higher values of Gini impurity, we decrease the exploration value so that the search focuses more on exploring variations with high q values. Conversely, for positions with low Gini impurity where one move is much better than the others, we increase the exploration value to ensure that other potential lines are not prematurely discarded. The idea to use Gini impurity was first proposed and tested by @Viren6. This patch was shown to affect the quality of data produced, so it has been intentionally excluded for datagen. Passed STC: https://montychess.org/tests/view/66aef6280f6f1e65cfa2b1f8 LLR: 2.93 (-2.94,2.94) <0.00,4.00> Total: 6944 W: 1626 L: 1460 D: 3858 Ptnml(0-2): 54, 787, 1643, 915, 73 Passed LTC: https://montychess.org/tests/view/66af1fd90f6f1e65cfa2b235 LLR: 2.93 (-2.94,2.94) <1.00,5.00> Total: 8718 W: 1877 L: 1705 D: 5136 Ptnml(0-2): 40, 931, 2255, 1083, 50 Rebased STC: https://montychess.org/tests/view/66b053380f6f1e65cfa2b657 LLR: 2.91 (-2.94,2.94) <0.00,4.00> Total: 9216 W: 2083 L: 1913 D: 5220 Ptnml(0-2): 82, 998, 2291, 1142, 95 2nd Rebased STC: https://montychess.org/tests/view/66bd91bf68e8f7e2fe23cfde LLR: 3.03 (-2.94,2.94) <0.00,4.00> Total: 4448 W: 1158 L: 982 D: 2308 Ptnml(0-2): 55, 487, 997, 597, 88 Bench: 1317423
Viren6
added a commit
that referenced
this pull request
Aug 16, 2024
First, when computing policy values in a position, we also calculate the Gini impurity, defined as (1 - sum of squares of policy values). A high Gini impurity indicates that there are many strong candidate moves in a position, and vice versa. This Gini impurity is then used to adjust the exploration scaling using a logarithmic formula. For higher values of Gini impurity, we decrease the exploration value so that the search focuses more on exploring variations with high q values. Conversely, for positions with low Gini impurity where one move is much better than the others, we increase the exploration value to ensure that other potential lines are not prematurely discarded. The idea to use Gini impurity was first proposed and tested by @Viren6. This patch was shown to affect the quality of data produced, so it has been intentionally excluded for datagen. Passed STC: https://montychess.org/tests/view/66aef6280f6f1e65cfa2b1f8 LLR: 2.93 (-2.94,2.94) <0.00,4.00> Total: 6944 W: 1626 L: 1460 D: 3858 Ptnml(0-2): 54, 787, 1643, 915, 73 Passed LTC: https://montychess.org/tests/view/66af1fd90f6f1e65cfa2b235 LLR: 2.93 (-2.94,2.94) <1.00,5.00> Total: 8718 W: 1877 L: 1705 D: 5136 Ptnml(0-2): 40, 931, 2255, 1083, 50 Rebased STC: https://montychess.org/tests/view/66b053380f6f1e65cfa2b657 LLR: 2.91 (-2.94,2.94) <0.00,4.00> Total: 9216 W: 2083 L: 1913 D: 5220 Ptnml(0-2): 82, 998, 2291, 1142, 95 2nd Rebased STC: https://montychess.org/tests/view/66bd91bf68e8f7e2fe23cfde LLR: 3.03 (-2.94,2.94) <0.00,4.00> Total: 4448 W: 1158 L: 982 D: 2308 Ptnml(0-2): 55, 487, 997, 597, 88 Bench: 1317423 Co-Authored-By: Viren6 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
First, when computing policy values in a position, we also calculate the Gini impurity, defined as (1 - sum of squares of policy values). A high Gini impurity indicates that there are many strong candidate moves in a position, and vice versa.
This Gini impurity is then used to adjust the exploration scaling using a logarithmic formula. For higher values of Gini impurity, we decrease the exploration value so that the search focuses more on exploring variations with high q values. Conversely, for positions with low Gini impurity where one move is much better than the others, we increase the exploration value to ensure that other potential lines are not prematurely discarded.
The idea to use Gini impurity was first proposed and tested by @Viren6.
This patch was shown to affect the quality of data produced, so it has been intentionally excluded for datagen.
Passed STC: https://montychess.org/tests/view/66aef6280f6f1e65cfa2b1f8
LLR: 2.93 (-2.94,2.94) <0.00,4.00>
Total: 6944 W: 1626 L: 1460 D: 3858
Ptnml(0-2): 54, 787, 1643, 915, 73
Passed LTC: https://montychess.org/tests/view/66af1fd90f6f1e65cfa2b235
LLR: 2.93 (-2.94,2.94) <1.00,5.00>
Total: 8718 W: 1877 L: 1705 D: 5136
Ptnml(0-2): 40, 931, 2255, 1083, 50
Rebased STC: https://montychess.org/tests/view/66b053380f6f1e65cfa2b657
LLR: 2.91 (-2.94,2.94) <0.00,4.00>
Total: 9216 W: 2083 L: 1913 D: 5220
Ptnml(0-2): 82, 998, 2291, 1142, 95
2nd Rebased STC: https://montychess.org/tests/view/66bd91bf68e8f7e2fe23cfde
LLR: 3.03 (-2.94,2.94) <0.00,4.00>
Total: 4448 W: 1158 L: 982 D: 2308
Ptnml(0-2): 55, 487, 997, 597, 88
Bench: 1317423