Skip to content

Releases: TerjeKir/weiss

Weiss 2.0

04 Aug 07:49
0db49c7
Compare
Choose a tag to compare

This version adds support for Chess960, MultiPV, searchmoves, and querying online Syzygy.

It also adds about 110 self-play elo:

LTC
+7679 -1547 =10774
Elo difference: 110.06 +- 3.23

Major changes:

Chess960 / FRC
MultiPV and `go searchmoves`
Online Syzygy probing for perfect play with 7 or fewer pieces (off by default)
Improved history heuristics
Improved evaluation of passed pawns

Weiss 1.4

02 Jul 22:09
d1ea873
Compare
Choose a tag to compare

Since version 1.3 released 5 months ago Weiss has gained about 160 self-play elo:

STC
+5315 -910 =3775
Elo difference: 164.29 +- 5.56

LTC
+4949 -698 =4353
Elo difference: 157.70 +- 5.18

Major changes:

Additional evaluation terms
Additional pruning heuristics
Improved use of history heuristics

Weiss now expects to play with Syzygy tablebases. Without them, it will probably misevaluate some more drawn endings than before.

Weiss 1.3

10 Feb 00:14
7626aa5
Compare
Choose a tag to compare

Since version 1.2 released 3 months ago Weiss has gained almost 100 self-play elo:

10+0.1
+4516 -2170 =3776
Elo difference: 79.26 +- 5.39

60+0.6
+4236 -1519 =4261
Elo difference: 96.67 +- 5.21

Major changes:

King Safety (Thanks to No4b)
Retuned evaluation
Improved history heuristic

Weiss 1.2

17 Oct 22:08
f8d625d
Compare
Choose a tag to compare

Since version 1.1 released a few weeks ago Weiss has gained about 120 self-play elo:

10+0.1
+4705 -1169 =4174
Elo difference: 127.73 +- 5.28

60+0.6
+4463 -918 =4955
Elo difference: 124.19 +- 4.84

Major changes:

Tuning (eval and search)
Improved makefile
Static Exchange Evaluation for move ordering
Singular Extension
Improvements to history and LMR
Can limit NoobBook to x moves
Supports 'go mate x'

Weiss 1.1

02 Sep 16:04
5c3bd88
Compare
Choose a tag to compare

Since version 1.0 released 3 months ago Weiss has gained about 55 self-play elo:

10+0.1
Score of weiss 1.1 vs weiss 1.0: +6160 -3651 =5325
Elo difference: 58.13 +- 4.49

60+0.6
Score of weiss 1.1 vs weiss 1.0: +1888 -1044 =2076
Elo difference: 59.12 +- 7.41

For cyclic time controls like at CCRL the improvement is about twice as big:

40/10.0
Score of weiss 1.1 vs weiss 1.0: +2470 -920 =1602
Elo difference: 111.56 +- 8.19

Major changes:

Evaluation tuned using AdaGrad from @AndyGrant's paper
Fixed cyclic time management
Fathom replaced with Pyrrhic
Various minor search tweaks

And as usual a lot of code cleanup.

Weiss 1.0

02 Jun 10:38
f44a4c6
Compare
Choose a tag to compare

Since version 0.10 released a month ago Weiss has gained about 65 self-play elo:

10+0.1
Score of weiss 1.0 vs weiss 0.10: +10887 -5472 =15687
Elo difference: 59.28 +/- 2.72

60+0.6
Score of weiss 1.0 vs weiss 0.10: +3192 -1234 =5574
Elo difference: 68.92 +/- 4.51

Major changes:

Lazy SMP multi-threaded search
Probes DTZ tablebases for optimal moves in the endgame
Improved history heuristic
Fixed a bug causing rare crashes in 0.10
Can probe noobpwnftw's chess cloud database (like BookFish)

And as usual a lot of code cleanup.

Weiss 0.10

25 Apr 22:02
5a22d0b
Compare
Choose a tag to compare

Since the previous release a little over 2 months ago Weiss has gained about 250 self-play elo:

10+0.1
Score of weiss 0.10 vs weiss 0.9: +7473 -857 =2378
Elo difference: 250.69 +/- 6.77

Major changes:

  • Tuned evaluation parameters
  • Improved pruning, reductions, aspiration
  • Improved time management

Usability:

  • PV lines should no longer be cut short arbitrarily
  • More frequent updates during longer searches

And as usual a lot of code cleanup.

Weiss 0.9

14 Feb 20:26
75cdd12
Compare
Choose a tag to compare

Since v0.8 a little over a month ago Weiss has gained about 40 self-play elo:

10+0.1
Score of weiss 0.9 vs weiss 0.8: +10488 -6974 =10582
Elo difference: 43.76 +/- 3.22

Major changes in v0.9:

  • Improved TT overwriting scheme #132
  • Improved time management #142 #143
  • Added 'improving' heuristic #161
  • Fixed a bug #157
  • Added GPLv3 license

The vast majority of my time this last month was spent cleaning up the source code, reducing total lines of code by several hundred, improving readability, and getting some minor speedups.

weiss 0.8

07 Jan 22:49
Compare
Choose a tag to compare

Since v0.7 a month ago weiss has again gained over 100 self-play elo:

60+0.6
Score of weiss 0.8 vs weiss 0.7: +2184 -639 =1452
Elo difference: 131.50 +/- 8.79

Major changes in v0.8:

And as usual lots of refactoring and cleanup.

weiss 0.7

06 Dec 04:49
291176c
Compare
Choose a tag to compare

Over the last 3 weeks weiss has gained around 175 elo:

10+0.1
Score of weiss 0.7 vs weiss 0.6: +230 -43 =127 [0.734]
Elo difference: 176.10 +/- 29.76

Major changes in v0.7:

  • Dynamic delta margin #84
  • Smaller aspiration window #86
  • Movegen code is readable and 1/3 the length #93
  • Tuned evaluation values #95 #96 (still a lot more to be gained from this)
  • Various speedups #98 #99
  • Pruning:
  • Move ordering:
    • Underpromotion delayed #100
    • History ignores promotions #101
    • Avoid doubled killers #105
    • Include moves that cause beta cutoffs in history #106

Overall evaluation is still very basic, while the search has gotten better.