Lozza 4.0

A bigger 768 → 128 → 1 SReLU network (still white-relative) trained on self-generated data using a DIY Javascript pipeline: datagen.js → filter.js → trainer.js. Includes the 3.1–3.16 development builds listed below.

Originally released with the weights in a separate .bin file; re-released on 6 Jan 2025 with the weights inlined into lozza.js again after users reported problems. Same strength.

Changes

• Minor tweaks for datagen + net command. (3.1)
• Defer UE to after legal check (doh!) and don't check pre-determined legal moves. (3.2)
• Make sure all UE updates (inc. castling etc) are a single accumulator loop. (3.3)
• Bigger net 768x128x1 srelu. (3.4)
• Allow successive NMP and beta pruning. (3.5)
• Simplify search recursion. (3.6)
• Optimise castling a bit. (3.7)
• Put eval in TT before search. (3.8)
• Prune QS with quickSee(). Only count nodes that iterate moves. (3.9)
• Fix some web stuff. Add improving indicator. (3.10)
• Optmise deferral of h1 accumulator update a bit more. (3.11)
• Use hash move in q search. (3.12)
• Move futility alpha test to move loop. (3.13)
• Use the Mersene Twister from Cwtch for randoms. (3.14)
• Load net from the .bin file created by trainer.js. (3.15)
• Add perspective (currently unused). (3.16)

Rating

CCRL 40/40: 2968 · CCRL Blitz: 2900

Usage

Use lozza.js with Node.js in any UCI chess GUI (Arena, Cute Chess, etc.), or in a web page/web worker.

Binaries

While Lozza is intended for web use, I can supply binaries on request. Please email me op12no2@gmail.com noting the target platform; thanks.