Skip to content

Search: retry quiet main-search SEE pruning - #28

Open
sebhofmann wants to merge 3 commits into
mainfrom
issue-3-quiet-main-search-see-pruning
Open

Search: retry quiet main-search SEE pruning#28
sebhofmann wants to merge 3 commits into
mainfrom
issue-3-quiet-main-search-see-pruning

Conversation

@sebhofmann

Copy link
Copy Markdown
Owner

Closes #3

Summary

  • rebase the quiet main-search SEE experiment onto current main
  • preserve TT moves and early ordered candidates
  • derive the quadratic threshold from history-adjusted LMR search depth
  • only prune quiets with negative history, preserving likely killers/countermoves
  • retain dedicated UCI controls, statistics, and quiet SEE regression coverage

Why the first attempt failed

PR #22 reached -2.7 +/- 9.4 Elo after 1,030 LTC games. The move-count gate alone still allowed SEE to remove useful quiets. The additional negative-history gate targets only moves that ordering already considers weak.

Validation

  • make clean && make both
  • ./see_test.sh: 112 passed, 0 failed
  • ./perft_test.sh --fast: 29 passed, 0 failed, 4 slow cases skipped
  • make test-nnue: passed
  • paired STC screen (2+0.02, 600 games): 133 / 103 / 364, +17.4 +/- 17.4 Elo, LOS 97.5%

LTC SPRT

Running: TC=60+0.6 ./tournament.sh sprt 3_review_new 3_review_base

  • H0/H1: 0 / +5 Elo, alpha=beta=0.05
  • base commit: 513ee20d7f9e9e82995cb0078fc9ef10f27fcbef
  • test commit: 57554b2e7e8da7309be5bdf1ae4c20dac83ad372
  • paired colors, -repeat, /home/paschty/Downloads/2moves_v2.pgn

The PR remains contingent on the LTC boundary; the STC screen is not final acceptance.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Search: add main-search SEE pruning for quiet moves

1 participant