Life In 19x19

WindCaliber wrote:Interestingly, the AlphaGo tool has the first 24 moves of the 1st game between Park and Ke in its database. Aside from the first couple of moves up to move 6, the moves by Park that Master seems to disagree with most are , , and (?). Meanwhile, every one of Ke Jie's moves was Master's optimal move. Maybe Ke Jie has finally become AlphaGo!

AlphaGo Master to be more precise! If Ke Jie 9p played an assistive role in the development of this tool, it may not be surprising.

I couldn't help but be intrigued by the winning percentage difference from move 6; Master seems very sure, it reminds me of when someone said that Cho Chikun 9p evaluated that a game as within half a point in the early middle game.

In addition, AlphaGo Master is, I believe, two iterations behind the latest version-- so I wonder, would be difficult for professionals to take its evaluations as definitively as it could be!

For example, regarding maybe someone could tell as to whether it is a case of Master preferring territory compared to AlphaGo Zero, which I think Michael Redmond 9p may have mentioned if I remember correctly.

PS: (WindCaliber) Having not posted in a while was a little intrigued by some similarities (Meant in the best way )

Uberdude wrote:That game to move 20 is the same as the team go from the Wuzhen event.

Funny you say that because I was thinking to myself, "Hmm, this game seems familiar.." Then, I tried browsing through all the AlphaGo games to find which one it was, but finally got too lazy to finish searching.

Elom wrote: I couldn't help but be intrigued by the winning percentage difference from move 6; Master seems very sure, it reminds me of when someone said that Cho Chikun 9p evaluated that a game as within half a point in the early middle game.

In addition, AlphaGo Master is, I believe, two iterations behind the latest version-- so I wonder, would be difficult for professionals to take its evaluations as definitively as it could be!

I think 4858 ELO is a good reason to believe its evaluations
It's still a bit of an open ended question as to whether AG Ke and AG Master are different. DeepMind said that they could be considered the same, but previously I think they said that they were going to further train Master before the Wuzhen event. My guess was that they had the same code architecture, but the only difference was extra training, which maybe didn't improve it too much due to plateauing.