alreadydone / lz

Go engine with no human-provided knowledge, modeled after the AlphaGo Zero paper.
GNU General Public License v3.0
28 stars 8 forks source link

maybe useful idea #84

Open sewim opened 5 years ago

sewim commented 5 years ago

First of all, It might be rediculous idea. As you already knew AI seems to be stepped back when they are winning enough. I saw interesting manual method for that issue. Some players who is using lizzie make black and white opposite therefore at least their winrate successfully drop down and solve that issue. Maybe if you can make it as a cord which works real time, There is possiablity that AI will be increased performence. Thanks.

alreadydone commented 5 years ago

I think you can just use my komi+next branch (or the release of the same name) and append --nonslack to your command. It will automatically adjust winrate to the range 10-90% and you can also specify --max-wr and --min-wr if that's not enough. Latest 15b nets works well, but 40b nets may not.

sewim commented 5 years ago

Yes I had used your piece of art. It was good but It didn't work well for elf and 40b. I think Maybe...just maybe converting B&W can be applied for any weights comparing with color input plane method.

alreadydone commented 5 years ago

Setting --target-komi -7.5 has absolutely the same effect as switching black and white. I get other komi by interpolating or extrapolating from these two scenarios.