search

pytorch / ELF

ELF: a platform for game research with AlphaGoZero/AlphaZero reimplementation
Other
3.37k stars 566 forks source link

[quick question] How much stronger is the weight v2 than the weight v1? #136

Closed y-ich closed 5 years ago

godmoves commented 5 years ago

According to the test from lz website, they are roughly the same strength.

l1t1 commented 5 years ago

here is some test
http://www.yss-aya.com/cgos/19x19/cross/LZ_05db_ELFv2_p800.html

y-ich commented 5 years ago

@godmoves san, @l1t1 san, Thank you for your information.

So does it mean that extending filters from 224 to 256 was not effective?

Even so, it seems that opening play style of v2 is different from the one of v1. Very interesting. I really appreciate the hard work of ELF developers.

alreadydone commented 5 years ago

v2 is trained from scratch with 20 million new games.

y-ich commented 5 years ago

@alreadydone san. Yes, I knew it. "extending" was misleading.

l1t1 commented 5 years ago

some new test result on github of leelazero now

l1t1 commented 5 years ago

2440 games vs elf v1 59.96%

y-ich commented 5 years ago

@l1t1 san, Thank you for your information.

59.96% means +70 Elo rating. I think that your information is the result in the condition of the same number of playouts. Is my understanding right?

224 channels is faster than 256 channels roughly by 1.14 times. Assuming that x2 playouts is 200 Elo stronger, v1 get +39 Elo rating by x1.14 playouts. So in the condition of time parity, v2 is +31 Elo stronger than v1.