Skip to content

Project History

Jump to bottom Edit New page
MaurizioDeLeo edited this page Oct 19, 2018 · 29 revisions

This page is no longer maintained. For new status:

  • See our blog

  • Discord chat #announcements and #dev-log have the most detailed updates.

  • The summary below mostly covers the main training pipeline. We are also preparing a transition to a completely rewritten engine (lc0.exe). The main reason is for speed, up to 10X faster for Nvidia GPUs.

  • For details of future plans, see our projects page.

  • For more detailed status updates, see our Discord chat #annoucements and #dev-log.

Summary of major changes and milestones.

  • 2018-06-10, ID 396: Switch main training pipeline to use scs-ben's bootstrapped net to try to shortcut fixing rule50 issues. Elo dropped for awhile because of this transition, but tests show the rule50 issues are mostly gone.

  • 2018-06-05, ID 379: rule50 plane normalized. Batch Norm Betas enabled. The main training pipeline is using these changes, but so far recovery from the rule50 norm fix is very slow. In parallel some 3rd parties (scs-ben, logosgg) are bootstrapping a 15x192 and a 20x256 net from existing data. And also the lc0 client is starting from scratch. After we observe progress on these we will decide how to proceed.

  • 2018-05-23, ID 330: Generate new networks once every ~40K games, but only do 10K steps (was 50K steps). Revert value_loss_weight back from 0.25 to 1.0.

  • 2018-05-13, ID 287: Change value_loss_weight from 1.0 to 0.25 as a temporary workaround for oversampling.

  • 2018-05-10, ID 271: v0.10

    • Fix #577 - Two inputs to the Neural Network (rule50 and the all 1s plane) were not connected properly.
    • Set --fpu_reduction=0 for training games.

  • 2018-05-08, ID 267: v0.9 disable tree reuse in self-play (wider search at root).

  • 2018-05-06, ID 253: Large Elo fluctuations began.

  • 2018-05-02, ID 237: v0.8 fix a bug in FPU, tune FPU and PUCT.

  • 2018-04-30, ID 227: Increased network size from 10x128 to 15x192. Self play games will be about twice as slow now.

  • 2018-04-18, ID 145: v0.7 fixed a bug in the code for black pawn promotions.

  • 2018-04-17, ID 125: Leela 125 vs Fruit 2.1 on TCEC, 6 15+10 games. +4 =1 -1 https://lichess.org/study/CmnpSyrx

  • 2018-04-17, ID 125: Leela 125 vs Stockfish 1.0 on TCEC, 20 15+10 games. +7 =3 -10 https://lichess.org/study/6cDZXrjY

  • 2018-04-17, ID 125: An informal bullet match against IM lovlas +17 =0 -0 https://lichess.org/study/YlP9Chea

  • 2018-4-15,16, ID 125: First match on TCEC, a rapid match vs Scorpio 2.79: +4 =3 -13 https://lichess.org/study/MFUV2n3I (games 5 & 6 are a highlight, as Leela won both sides of the opening)

  • 2018-4-14, ID 125: An informal 20+5 match on Lichess vs IM lovlas: +9 =2 -0 https://lichess.org/study/kLXrkT4M

  • 2018-4-12, ID 123: Increase net size from 6x64 to 10x128, and architecture changes. Revert from -v1600 back to -v800.

  • 2018-4-7: Promotion gating changed from -50 to -150.

  • 2018-4-6, ID 94: Self play games changed from -v800 to -v1600. This will produce higher quality games but run twice as slow. The main motivation for this change is to balance the rate of new games coming in to new networks being trained.

  • 2018-3-29: During the transition from client V3 to V4, the move_count input to the NN was disabled. This caused some strange effects while self play and match games were being produced from a mix of V3 and V4 clients. Also networks trained from before this period may not work well in V4 clients.

Add a custom footer
Add a custom sidebar
Clone this wiki locally