Harmon Chess App Update - Android

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

User avatar
MikeB
Posts: 4889
Joined: Thu Mar 09, 2006 6:34 am
Location: Pen Argyl, Pennsylvania

Harmon Chess App Update - Android

Post by MikeB »

Source, app and notes can be found here:

https://github.com/MichaelB7/droidfish/releases/tag/v13

Enjoy and stay Safe!

MikeB
Image
purechess
Posts: 82
Joined: Wed Nov 28, 2018 1:28 pm
Full name: Heinrich Pulliter

Re: Harmon Chess App Update - Android

Post by purechess »

Thanks!!

Just saw mb.bin in the Assets.
Can you provide more info about this NNUE network?
User avatar
Scally
Posts: 232
Joined: Thu Sep 28, 2017 9:34 pm
Location: Bermondsey, London
Full name: Alan Cooper

Re: Harmon Chess App Update - Android

Post by Scally »

Hi,

That is Mike’s own trained network mentioned in a few threads here, here’s his main thread:

http://talkchess.com/forum3/viewtopic.php?f=2&t=76790

mb for Michael Byrne.


Al.
User avatar
MikeB
Posts: 4889
Joined: Thu Mar 09, 2006 6:34 am
Location: Pen Argyl, Pennsylvania

Re: Harmon Chess App Update - Android

Post by MikeB »

purechess wrote: Wed Mar 17, 2021 11:14 am Thanks!!

Just saw mb.bin in the Assets.
Can you provide more info about this NNUE network?
it is lightly trained to depth 12 on AI generated positions (25M positions for each depth ,4, 6, 8 up to depth 12 ( no FENS from real chess games were trained on it) and although quite strong at 1 min and 1 sec increment roughly equal to Zappa Mexico on a PC ,if you set it to 6 second game with 1 second increment on the Harmon app plays a solid human level type game and is beatable. The training is not mature, so position score evaluations are often not the best since for the most part it was trained in "pure" mode with a lambda setting of 0 ("zero"), meaning it was trained on wins losses and draws and not evaluation scores ( for most part, a few training sessions were done on evaluations , but not many anyway) I enjoy playing against it, ymmv. It also finds many of the hard positions to solve , best move anyway , but the score is not always there. Just copy and paste the mb.nn file into you nnue folder under the droidfish folder and change the EvalFile setting under engine settings to mb.bin.

Currently generating depth 14 positions, still keeping it to 25 M , for the next training session.
Image