Document Elo impact of various parts of search
authorJoost VandeVondele <Joost.VandeVondele@gmail.com>
Sun, 1 Apr 2018 01:13:29 +0000 (03:13 +0200)
committerStéphane Nicolet <cassio@free.fr>
Sun, 1 Apr 2018 01:21:45 +0000 (03:21 +0200)
commite408fd7b10503a9114962cb5972abd9957bc67c2
tree73d4d7af870f1a0ef89ea87ad73e8de177f138d5
parent76f9cd4df1292ffa919e039edcfb8069f576e698
Document Elo impact of various parts of search

In order to understand better the impact of various techniques used in search,
Elo estimates have been run at STC for 60000 games (statistical error ~1.8 Elo),
disabling each feature in turn. This should help future improvements and
simplifications to pick suitable targets.

The list of tests is:
step  7 : http://tests.stockfishchess.org/tests/view/5abcbb4b0ebc5902926cf1ca
step  8 : http://tests.stockfishchess.org/tests/view/5abcbb680ebc5902926cf1cc
step  9 : http://tests.stockfishchess.org/tests/view/5abcbb850ebc5902926cf1ce
step 10 : http://tests.stockfishchess.org/tests/view/5abcbbeb0ebc5902926cf1d2
step 11 : http://tests.stockfishchess.org/tests/view/5abcbbbf0ebc5902926cf1d0
step 13 : http://tests.stockfishchess.org/tests/view/5abd03680ebc5902926cf20b
step 13a: http://tests.stockfishchess.org/tests/view/5abd29660ebc5902926cf22a
step 13b: http://tests.stockfishchess.org/tests/view/5abd29820ebc5902926cf22c
step 14 : http://tests.stockfishchess.org/tests/view/5abd03860ebc5902926cf20f
step 14a: http://tests.stockfishchess.org/tests/view/5abd2b6c0ebc5902926cf230
step 14b: http://tests.stockfishchess.org/tests/view/5abd2b8d0ebc5902926cf232
step 14c: http://tests.stockfishchess.org/tests/view/5abd2bad0ebc5902926cf234
step 14d: http://tests.stockfishchess.org/tests/view/5abd2bcf0ebc5902926cf236
step 14e: http://tests.stockfishchess.org/tests/view/5abd2bf10ebc5902926cf238

This patch documents this in the code.

Note that it will be a waste to recompute these estimates often, even a couple
of [0,5] patches are unlikely to change them by more than the error bars. The
interest of the Elo annotations in the code is not in the details, but in high-
lighting trends such as razoring (2 Elo) and singular extensions (60 Elo). These
estimates should be recomputed at most once a year.

Closes https://github.com/official-stockfish/Stockfish/pull/1522

No functional change.
src/evaluate.cpp
src/search.cpp
src/timeman.h