[바둑 AI] RN 에 대한 기본 정보 확인

1. 공식 개발 사이트?

https://github.com/zakki/Ray/releases

 

2. 개발자 닉네임 

위 사이트 주소에서 알 수 있듯 zakki

 

3. 성능

http://www.yss-aya.com/cgos/19x19/standings.html

 

With GPU 

i7-2600K (or 6700K) + GTX980ti (출처 : https://lifein19x19.com/viewtopic.php?p=216659&sid=1c0591d97ba778b9818bde362672812d#p216659)

Rn.4.3-4c

Rating 3349

109 Games

2017-06-01 16:28:23

 

Without GPU 

Rn.3.3.1-4c0g Rated: 2591 as of 2017-01-31 15:09:24

 

4. GPU 없을 때 왜 이리 약한가?

https://github.com/zakki/Ray/issues/6

zakki commented on 23 Feb

A known issue about CNTK before 2_0_Beta_8 is CPU mode doesn't support batch normalization, so value network returned garbage value.
That makes CPU mode rn weaker than gnugo.

Another problem is that DCNN evaluation on CPU is too slow to reflect tree-search. It was fixed on rn-3.10.2 by waiting DCNN. (That fix strength, but it slow down simulation near 100po/sec as side effect)

 

zakki commented on 27 Feb

I tried 20 games of Rn.3.10.2.exe --no-gpu --thread 1 --playout 5000 vs White command: Rn.3.10.2.exe --thread 6 --playout 5000.
Win rate of CPU only was 40% (±11). I cannot say it is strong or weak.

결론 : 같은 playout 이면 시간의 차이가 있을 뿐이지 기력차이는 없다.