Predict AlphaGo on Future of Go Summit in Wuzhen, China

[pookpooi]

In this news http://sports.sina.com.cn/go/2017-05-12 ... 4772.shtml

Meng Tailing 6p suggest Ke Jie to stop looking for AI weakness and play in his usual style.
This confirm that Ke Jie has been train under AI-specific program for a while (his losing streak against FineArt maybe one of his experiment) and it might affect his results against human recently.

[djhbrown]

Meng is wrong; Ke Jie is right. You can't beat her at her own game, you have to probe beneath the surface.

The dog does have weaknesses, although finding them is harder for her than for most opps. based on how she works, i'm guessing one of them is a lack of positional judgment, compensated for by extraordinary reading ability.

We will see

[Waylon]

The dog does have weaknesses, although finding them is harder for her than for most opps. based on how she works, i'm guessing one of them is a lack of positional judgment, compensated for by extraordinary reading ability.

We will see

From my experience with Zen, Crazy Stone and Leela, I'm guessing just the opposite: positional judgement is superb, the reading ability is not yet perfect.

[djhbrown]

that's the difference between us, a difference between an objective examination of the technology, and a subjective impression gained from personal experience

[dfan]

From my experience with Zen, Crazy Stone and Leela, I'm guessing just the opposite: positional judgement is superb, the reading ability is not yet perfect.

The consensus of the 9ps that lost to Master seemed to be that its positional judgment was qualitatively better than theirs.

[Bonobo]

The consensus of the 9ps that lost to Master seemed to be that its positional judgment was qualitatively better than theirs.

Methinks this would fall under …

a subjective impression gained from personal experience

[pookpooi]

One more week to go, DeepMind team on the way to China (I don't know if they arrived at the airport like celebrity like what happened in Seoul), while many Chinese pro are estimating Ke Jie 9p chance
Yu Zhiying 5p, the highest ranked Chinese female player gave Ke Jie 30% chance against AlphaGo
Gu Li 9p said if Ke Jie play with his 100%, he will get 5% chance against AlphaGo
Nie Wieping 9p simply said AlphaGo has 100% win chance
Ke Jie gets a lot of advice from his pro friends. He said this is the fight to death, there's no winrate, lost is 100% lost, win is 100% win
Also, don't forget to vote! AlphaGo version was frozen two weeks before the real match (may be to test if it's working correctly on big machine)

[alphaville]

One more week to go, DeepMind team on the way to China (I don't know if they arrived at the airport like celebrity like what happened in Seoul), while many Chinese pro are estimating Ke Jie 9p chance
Yu Zhiying 5p, the highest ranked Chinese female player gave Ke Jie 30% chance against AlphaGo

I agree with her. Ke Jie has 30% chance to win, if he receives 3 stones handicap.

[yoyoma]

pookpooi what is the source of the AlphaGo rating graph you posted?

[djhbrown]

Ke Jie has 30% chance to win

He has a 100% chance to win the only thing that counts: the appearance money, because he's already got it. Which is rather unfair to Poor Porky Pie PRers who have to feed parents and brother on a few clickbait crumbs from the rich man's table, sourced-up by a dash of Auntie Dee's makeitupasyougoalong Homemade graphic Catsup.

[pookpooi]

Official website launched
http://events.google.com/alphago2017/

[pookpooi]

Rumor around weibo that not only the new AlphaGo will learning without human game record, but it also doesn't use Monte Carlo Tree Search

http://www.weibo.com/p/23041873040b820102wwci

My opinion, is this even possible? Please take this as a grain of salt. But there's information that DeepMind dedicated the whole day (24th) to conference. (Is someone roaming through DeepMind presentation slides without permission? That's very bad)

[alphaville]

Rumor around weibo that not only the new AlphaGo will learning without human game record, but it also doesn't use Monte Carlo Tree Search

http://www.weibo.com/p/23041873040b820102wwci

My opinion, is this even possible? Please take this as a grain of salt. But there's information that DeepMind dedicated the whole day (24th) to conference. (Is someone roaming through DeepMind presentation slides without permission? That's very bad)

Any chance you can post the weibo info without the need for us to login there?

The Nature paper stated that neural-network only (no MCTS) version of AlphaGo was already stronger than all other commercial computer-go games at the time of Fan Hui match, but way weaker than the MCTS version.
It sounds impossible at first sight to think that a pure static analysis version, without MCTS, would be stronger than a pro; on the other hand, with a very deep neural-network architecture, I guess it may be possible, since some of the layers of the network would effectively do some "reading ahead". Just speculating here.

[pookpooi]

For me it doesn't require sign up to access weibo, it just that they're in Chinese language, but here in spoiler.

柯洁与AlphaGo的比赛又要开始了，这次有什么看点呢？想了几天还没有想出来。并不是没有看点，而是谷歌一直没有透漏，或者说我没有看到，今天的AlphaGo与去年的AlphaGo有什么不同，不知道有什么不同，就不知道看什么。

以前在文章中我曾经猜想过，也许新的AlphaGo会尝试从零学习，也即抛弃从人类棋谱中学习，完全靠自我互博，逐步提升自己的能力，但是一直得不到证实或者正否。我一直觉得这是有可能的，不过可能需要的计算资源要大的多，虽然一直有朋友反对我这种看法。

前几天和另外的朋友聊天时，得到了更加激进的消息，说新的AlphaGo不仅是从零学习，而且还舍弃了蒙特卡洛树搜索，让我大吃一惊，如果真是这样，那么将是又一次大的进步。不过我总有点不太相信，从零开始是有可能的，但是舍弃蒙特卡洛树搜索，还是有点不可想象。

前几天看到聂卫平的文章，提到有些棋手已经在跟AlphaGo学习，在比赛中走出了之前认为不好但是AlphaGo走过的类似的棋，我想这就是AlphaGo对围棋发展所起的推进作用，如果AlphaGo真的舍弃人类棋谱，从零开始学起的话，这种贡献可能会越来越多。

期待在这次比赛之后谷歌团队能透漏一些技术上的问题，这样对于我们如何看待这次比赛，会有很大的帮助。

总之这次比赛关注的已经不是谁输谁赢问题了，而是技术上是如何发展的，这对AI才是有意义的，否则如果只是重复过去，则意义要大打折扣。

[Uberdude]

It's important to note that saying this new version of AlphaGo doesn't do Monte Carlo Tree Search is not the same as saying it doesn't do tree search, though I don't know if that technical precision was lost in translation. You could have a tree search version which constructs a game tree using a policy network to suggest moves, and a value network to evaluate nodes in that tree without doing monte-carlo rollouts to the end of the game. In their published work so far combining the rollouts with the value network gives the best results, but as the networks keep getting better it wouldn't surprise me if tree+policy+value+rollouts was now 15p strength and tree+policy+value "only" 13p and therefore good enough for the upcoming match.