출처 : http://www.bbsdigest.com/thread/index?bid=87&tid=31270999

I'm a PhD student in Computational Neuroscience/Machine Learning who attended a talk today at UCL by David Silver, lead author on the AlphaGo Nature paper: http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html Although unfortunately I didn't take any pictures, I can confirm he did show us some slides on AlphaGo's evaluation functions and how their assessment of the probability of winning changed during the course of each match. 

Since I heard there was some interest about whether the game is balanced in its initial state, I looked carefully at the initial intersection with the y-axis, corresponding to the point at which no moves had been played. I can confirm that according to the Value Network's assessment, white has an initial probability of winning bigger that 50%, my guess being 53% (the scale used a 25% interval and so was very hard to read). By contrast, the Monte Carlo rollout had what looked like an almost exact 50-50 assessment. Although AlphaGo uses both assessments, David Silver said that the Value Network is in general more accurate and is thus more highly weighted in the latest versions of AlphaGo. So it's fair to say AlphaGo thinks white (with 7.5 komi) has a small advantage. 

Also, one other interesting tidbit is that according to Deepmind's internal assessment, AlphaGo's ELO was much, much bigger than Lee Sedol's (above 4000 iirc). However, David Silver acknowledged that he didn't really trust their internal assessment as AlphaGo was only playing versions of itself by reinforcement learning, and so may have been completely unaware of surprising strategies which can beat it. Which is why they had to play Lee Sedol in those exciting matches (and indeed, according to their internal ELO they probably would have expected 5-0 not 4-1) 

Hope you find this interesting anyway, I sure did!

요약

• 알파고 가치망에서는 덤7.5집일 때 백이 약간이나마 유리하다고 평가
• 알파고 ELO 4500 은 간접측정이라 신뢰성이 부족.