Chess engine: Arcanum 2.3

Arcanum is a free UCI chess engine under the MIT license.

Author:Lars Aurud, Rating Chess Engines Diary CEDR=2922

v.2.3:

Convert to search to PVS.

Only return TT scores when not in PV

Use adaptive reduction in LMR based on depth and move number.

Use adaptive reduction in NMP based on depth.

Use a linear function of depth and PV to set the RFP threshold.

Fix a bug in move ordering causing wrong ordering due to overflow.

Avoid RFP and razoring when alpha and beta respectively is close to mate to find more checkmates.

Add LMP (Late Move Pruning) reducing the number of quiet moves searched at low depths.

Add a layer to the NNUE. New architecture: 786->256->32->1.

Performance

Check for timeout every 256th node to reduce slow calls to getMs()

Use T0 as prefetch hint for TT and NNUE

Improve caching of board features such as opponent attack squares

Other

Add option NormalizeScore (default true).

Normalizes the cp score, such that 100cp represents ~50% chance of winning in self play.

Note that this does not affect search, only the scores sent to UCI.

Allow setting the hash option to 0 to disable the TT.

Only update seldepth in PV nodes

Add a help command listing all the supported UCI commands,

including those for NNUE training and Fengen.

A lot of code quality improvements

Elo Estimate

Arcanum v2.3 is estimated to be ~160 Elo stronger than Arcanum v2.2.

Self Play

Time Control 120s + 1s

Score of v2.3 vs v2.2: 106 - 20 - 74 (Win - Loss - Draw)

v2.3 playing White: 57 - 8 - 35

v2.3 playing Black: 49 - 12 - 39

Elo difference: 159.8 +/- 39.6, LOS: 100.0 %, DrawRatio: 37.0 %

Other Engines

This is the result of a gauntlet with time control: 60s + 0.5s.

Their CCRL blitz ratings are displayed in the last column.

Arcanum 2.2 - results

Arcanum 2.3 download

Chess Engines Diary