Arche 0.4.6 - new version chess engine
Arche is a UCI chess engine in Rust. Most of the effort goes into being able to tell whether a change to it helped: every change states in its commit message how much of the tree the search looks at afterwards, one that changes how the engine plays is measured in games before it merges, and one that claims to be faster states how much faster it measured.
Author: Andrew
What's new?
Corrections to this section, written by hand after it was generated. The
generated lines below are left as they were published.
Two entries here print [bench 48354803] and their trees pin 47836191:
"Count the iteration a budget gave up in the nodes a search reports"
(3298289) and "Add an effort instrument that differences two
configurations by the node" (b9325ae). Both were written on quiet
futility pruning, which counts 48354803, and rebased onto the late move
count before they landed. Neither moves a pin, so their bench is the
late move count's 47836191.
Features
(search) Add an adaptive null move reduction that grows with depth and the eval's margin over beta [bench 50127050] [elo +27 ±16 (sprt [0, 10] passed, 1000 games, 10+0.1, vs 54d85b9)]
(search) Decide the deep reduction by depth and move index [bench 50263773] [elo +2 ±9 (sprt [-10, 0] passed, 3500 games, 10+0.1, vs ac53360)]
(search) Add quiet futility pruning at depths one to three [bench 48354803] [elo +21 ±13 (sprt [0, 10] passed, 1500 games, 10+0.1, vs 92922de)]
(search) Add a late move count at depths one to three [bench 47836191] [elo +37 ±16 (sprt [0, 10] passed, 1000 games, 10+0.1, vs bdf0bef)]
(search) Add mate distance pruning [bench 2153533]
Bug Fixes
(search) Count the iteration a budget gave up in the nodes a search reports [bench 48354803]
(search) Fall back to a smaller transposition table at startup [bench 47836191]
(search) Measure an aborted search's time over the nodes it reports [bench 47836191]
(uci) Write a final info line with every node the search spent
Documentation
(search) Record the attention model's objective as measured and rejected
(search) Correct what the late move count's line is pinned to [bench 47836191]
Development
(deps) Bump the actions group across 1 directory with 4 updates
(bench) Hold a term by name rather than by a flag of its own
(deps) Bump the mache actions from v0.3.0 to v0.4.0
(bench) Add an effort instrument that differences two configurations by the node [bench 48354803]
(bench) Read the effort test's outcomes above the depths the rule decides
(deps) Bump the mache actions from v0.4.0 to v0.5.0
(docs) Correct the quiet futility figure and four counts the new rules staled
(release) Add a second calibration gauntlet against the ccrl 40/15 list
(bench) Let the effort instrument ablate the late move count [bench 47836191]
(ci) Chain up to four sprt batches in one strength run
(ci) Read the ladder's numbers off resolve rather than off the inputs
(bench) Raise the bench depth from nine to eleven [bench 6900228]
Docker image: ghcr.io/aywrite/arche-lichess-bot:v0.4.6 (sha256:a7cafa044be61347555d684869af5d2a24d0fdc62234c6132e8f4f54ab60e545)
Performance compared to v0.4.5 | +117 ±22 Elo (500 games)
Estimated rating | 2760 ±30 (95%) on the ccrl blitz scale (400 games)
Estimated rating | 2686 ±52 (95%) on the ccrl 40/15 scale (96 games)
Arche 0.4.2 took part in our tests and ended up near the bottom of the standings in the lowest qualifying group, Group E, with a score of 3.5 out of 9 games (44 place).
Comments
Post a Comment