Ke Jie to play AlphaGo in May

[wolfking]

According to China's Xinhua News Agency(http://news.xinhuanet.com/sports/2017-0 ... 766628.htm) and China Mind Sports Net(http://www.imsa.cn/archives/45947) (both links in Chinese), Ke Jie will play AlphaGo during the Wuzhen Weiqi Summit this May.

The format is best of three, but all three games will be played regardless of results of the first two games. The Mind Sports website has a little more details, suggested that there will be games between AG and a team of top Chinese pros, and the team games and Ke - AG games will alternate (does that mean there will be two team games?). The prize money is expected to exceed that of Lee - AG matches.

A press conference is scheduled to take place on April 10, when the news will be officially announced.

[pookpooi]

News in English
http://english.cctv.com/2017/04/07/ARTI ... 0407.shtml
http://www.ecns.cn/cns-wire/2017/04-07/252446.shtml

[Kirby]

Is it true that this version of AlphaGo was trained without the policy network (I.e. No initial training based on human games)?

[Uberdude]

Kirby, to clarify terminology I understand the suggestion is that it may be trained with a policy network solely from self-play games rather than human games and then self-play games as in the Fan Hui or Lee Sedol or Master versions, not that there is no policy network. As for what is true, that'll have to wait until the 10th, if there is indeed an announcement then.

[Kirby]

Kirby, to clarify terminology I understand the suggestion is that it may be trained with a policy network solely from self-play games rather than human games and then self-play games as in the Fan Hui or Lee Sedol or Master versions, not that there is no policy network. As for what is true, that'll have to wait until the 10th, if there is indeed an announcement then.

I see. So there still be some sort of human origin behind the AI, since the self-play games were only possible after starting from a policy network trained by human games.

[pookpooi]

since the self-play games were only possible after starting from a policy network trained by human games.

Can policy network trained by MCTS AI (that could produce good 5-6 dan games)? So that human origin factor is eliminated.

[Uberdude]

Kirby, to clarify terminology I understand the suggestion is that it may be trained with a policy network solely from self-play games rather than human games and then self-play games as in the Fan Hui or Lee Sedol or Master versions, not that there is no policy network. As for what is true, that'll have to wait until the 10th, if there is indeed an announcement then.

I see. So there still be some sort of human origin behind the AI, since the self-play games were only possible after starting from a policy network trained by human games.

No. The approach would be something like:
- Start with a bot that can play legal moves (let's call it IdiotBot)
- IdiotBot plays vs IdiotBot loads of times with some randomness, record winner
- Train a policy network to favour moves chosen by the winner
- Feed this back into IdiotBot so it become DumbBot
- Play DumbBot vs DumbBot gazillions of times, keep improving policy network.
- Probably start doing a value network too
- Repeat
- ???
- New AlphaGo
- Profit!

When they made their Atari game playing AI it didn't start off training on humans playing the games, it just started doing random things and then eventually got good at the games.

[Kirby]

I see. I misinterpreted your meaning behind this part:

trained with a policy network solely from self-play games

I thought you were suggesting that a policy network would be trained from self-play games played by an existing version of AlphaGo (e.g. take games like the AlphaGo vs. AlphaGo ones that DeepMind released, and train a policy network based on this). It seems that you meant that the self-play games would be completely from the ground up, only knowing the rules of the game.

That's a bit more interesting to me, since it could lead to some creative play.

On the other hand, given that AlphaGo had issues with Lee Sedol's move 78, it seems like it could be challenging to reach a high level this way - the branching factor from the early self-play games would be huge.

[Kirby]

I guess the real improvement to the policy network, if they start from "scratch" like this (without human input), would be the reinforcement learning on the policy network, like they did for AlphaGo.

So the first versions of the policy network might suck, since it's basically just trying to predict random moves. But maybe it will get a bit better after the policy network is improved.

[pookpooi]

http://www.cnbeta.com/articles/tech/600981.htm

Report that Google China send invitation to media for this afternoon press conference (Chinese Timezones) Which is 6 hours left. Waiting, waiting....

[djhbrown]

Alfadog will trounce Ke Jie

[pookpooi]

Watch live press conference of AlphaGo vs. Ke Jie here at 3pm Chinese Timezone

http://bn.sina.cn/video/live/index/sports/alphago2

[macelee]

Key information:
- Ke Jie vs AlphaGo, best-of-3 match played on 23rd, 25th and 27th May
- Chinese rule, 3 hours main time each side followed by 1-minute byoyomi
- AlphaGo will use distributed resources
- Winner prize 1.5 million USD, Ke Jie has a game fee of 300,000 USD.
- On 26th of May, a game between AlphaGo and a Chinese team consisting of Shi Yue, Mi Yuting, Tang Weixing, Chen Yaoye and Zhou Ruiyang. (2.5 hours each side)
- Some more entertaining games: Gu Li + AlphaGo vs Liao Xian + (another instance?) of AlphaGo (not sure how human players and AI split their roles)

[Charlie]

I wonder: will this event be broadcast with English commentary and a bit of pomp and ceremony, like the one in March 2016? I'd book time off work... again.

[wolfking]

I wonder: will this event be broadcast with English commentary and a bit of pomp and ceremony, like the one in March 2016? I'd book time off work... again.

It seems AGA will do something in English https://www.youtube.com/watch?v=aJBGUvNatnQ&t=357s.