What's new
This version mostly focuses on search tweaks. Most every tweak I could get to
gain has been included, so I imagine we're getting into diminishing returns and
~5 Elo patches.
Simbelmyne 1.8 is around 180 Elo stronger than 1.7 in STC self-play.
A small gauntlet places Simbelmyne at around 3050 Blitz 🎉
# PLAYER : RATING POINTS PLAYED (%)
1 Avalanche 1.3 : 3085.0 322.0 727 44.3%
2 Princhess 0.15 : 3073.0 330.5 726 45.5%
3 Akimbo 0.5 : 3058.0 405.5 728 55.7%
4 Simbelmyne 1.8.0 : 3052.5 3221.5 5811 55.4%
5 Koivisto 3.0 : 3027.0 328.0 726 45.2%
6 Nalwald 16 : 3001.0 389.5 726 53.7%
7 Viridithas 3.0 : 2985.0 296.0 726 40.8%
8 Black Marlin 2.0 : 2938.0 265.0 726 36.5%
9 Polaris 1.7.0 : 2937.0 253.0 726 34.8%
Added features
🔍 Search
- Tweak aspiration window implementation (20.5+/-14.2) (#207)
- Reduce PV nodes less instead of non-pv nodes more (10.1+/-8.1) (#208)
- History table refactor (#209) (#210)
- Tweak quiet history penalties and ageing (35.9+/-20.1) (#211)
- 1 ply continuation history (21.6+/-14.6) (#215)
- Countermove table (13.4+/-10.4) (#216)
- Don't overwrite TT move in all-nodes (21.6+/-14.7) (#217)
- Include promotions in tacticals in (somewhat) staged movegen (#220)
- Use "improving" heuristic in LMP (18.8+/-13.4) (#221)
- Use "improving" in NMP (11.1+/-7.5) (#222)
- Reduce more/less according to history score (16.7+/-12.3) (#223)
⚖️: Evaluation
- Make incremental eval term branchless (#204)
🐛 Bugfixes
- Fix SEE to work with non-zero margins (#202)
- Only report "stop_early" when there's one legal move in root (#205)
- Fix nodecounts and nps reporting (#206)
- Fix LMP move threshold (23.8+/-15.7) (#219)
♟️ Lichess Bot
- Tweak Dockerfile to fetch and build source from
main - Add deploy script for easy redeploys
- Tweak deploy script to GC old containers/images
🛠️ Misc
- Use fixed-point multiplication trick instead of modulo based indexing of TT (12+/-8) (#203)
See the respective PRs for self-play results where relevant
Choosing a binary
This release comes with precompiled binaries for all major platforms. Because
the engine benefits tremendously from more modern CPU instruction sets, there
are binaries compiled for major instruction sets, following the x86-64
microarchitecture levels
as a naming scheme.
Realistically, on modern hardware (< 10 years old), you should be okay to use the
binaries labeled V3. If the engine crashes within the first seconds, try V2,
and so on.
What's Changed
- Fix/refactor SEE by @sroelants in #202
- Swap TT indexing from modulo-based to fixed-point multiplication by @sroelants in #203
- Use match statement in incremental eval by @sroelants in #204
- Only flag "one legal move" in root by @sroelants in #205
- Share the time controller between search iterations by @sroelants in #206
- Tweak aspiration window search (20.5+/-14.2) by @sroelants in #207
- Reduce pv nodes less, rather than non-pv nodes more by @sroelants in #208
- History table refactor by @sroelants in #209
- History refactor, pt2 by @sroelants in #210
- Simplify history penalties and ageing (35.9 +/- 20.1) by @sroelants in #211
- Lichess deploy script by @sroelants in #212
- 1ply continuation history (21.6 +/- 14.6) by @sroelants in #215
- Add countermoves to history (13.4 +/- 10.4) by @sroelants in #216
- Don't overwrite TT move in fail-low nodes by @sroelants in #217
- Fix LMP moves threshold by @sroelants in #219
- Include promotions in tacticals by @sroelants in #220
- Search/improving lmp by @sroelants in #221
- Use improving in NMP by @sroelants in #222
- Reduce less/more according to history score by @sroelants in #223
Full Changelog: v1.7.0...v1.8.0