AlphaGo vs. AlphaGo with Michael Redmond 9p: Game 4
ฝัง
- เผยแพร่เมื่อ 23 ก.ย. 2024
- Michael Redmond 9p, hosted by the AGA E-Journal's Chris Garlock, review the 4th game of the amazing AlphaGo vs. AlphaGo selfplay games.
The 50 game series was published by Deepmind after AlphaGo's victory over world champion Ke Jie 9p in May 2017.
Produced by Michael Wanek & Andrew Jackson
Thumbnail image by Lorie Shaull - Own work, CC BY-SA 4.0, commons.wikime...
![[Full Episode] The Restaurant War Thailand ศึกพ่อค้าซ่าแม่ค้าแซ่บ Episode 1 | 22 ก.ย. 67](http://i.ytimg.com/vi/BuzQ70v-UZQ/mqdefault.jpg)
Michael's love and appreciation for Go Seigen is so beautiful to feel through his words.
I'm super impressed with the rate at which these videos are being published!
To be fair they probably prerecorded several if not most of the game analyses. But nonetheless Michael just does an amazing job.
Yves M. Yeah you could tell that the first couple are during the same session, but it's still a lot of Michael Redmond's time, not including the actual part where he studies the game, which is presumably longer than the time he takes to go over the games with us.
ConsciousBreaks Yes sure, you're right of course. To be doing all 50 games. Hats off! We can be grateful.
to be fair he would probably study the games anyways. But he does create a lot of nice variations for us
It's super impressive to me to see rational explanations for this chaos.
Likewise. So many moves are otherwise completely opaque to me.
A new Redmond review almost every day! This is awesome! I felt a huge drop in my Go interest after the excitement of the Master/KeJie match ended, but this helps fill the void.
This is better than the Ke Jie match. The fact that the games look so different from when playing against humans just shows how much stronger AG is above the top pros.
At some point in every review the ending is given away which makes me lose a bit of interest. Strange since it's just an AI playing itself and the result doesn't matter. Still I like to see everything play out the first time without knowing how it will end.
AlphaGo won all 50 games. Sorry I gave away all the endings xD
I thought maybe there'd be a triple ko...
Those commentaries on alphago's games (as well as Mr Redmond previous ones) are such a treat to us amateurs, both as an understanding of alphago's abilities and as an understanding of the game. It nice to see that even if Master's moves were considered god-like because of his winning record, it still makes some mistakes... I really liked Mr Garlock commentary about the history of go, he has nice ideas, would be cool if he could participate a little bit more :)
I have a couple questions though :
_ how long does it take to review a game of Master, how long compared to a usual pro game and what makes the difference ?
_ did Michael had access to some other Japanese/Chinese/Korean commentaries from other pros (I would be interested to know what top pros like Ke Jie or Park Juhngwhan might think about those games)
_ what would it take for the human players to catch up to alphago ? Is it just a matter of creativity or pushing the reading further or something else ?
Anyway thank you so so much & keep it up, you're working for history :)
Maybe the moves aren't simple complete mistakes, I think we just don't see the reason.
Thanks for all the hard work!
I saw this video was published when I went to sleep yesterday and I've waited the entire night to watch it!
When he is explaining go, it makes so much sense. When I play though i just lose lol
What books are you referencing by Go Seigen in the early part of the discussion of this game (within the first 15 minutes or so). Thanks
Wow! This was a great match. It really proves that a player can truly reach their potential only when they have an opponent that challenges them to do so.
I really hope some day Michael gets access to AlphaGo or a similar AI to see what might happen in his variations. Especially the two "losing moves" by black in this video, which would be fascinating if AlphaGo were forced to play Michael's suggested moves (since Michael justified their selection quite well). Similarly, I've also been interested in seeing what AlphaGo or a similar AI would do if handed some of the more famous Go games of the past, and see what it would think of the positions and whether the outcome of the match would change in its hands. The more well-studied the famous game is, the more I think we might learn from having an AI offer a "fresh set of eyes" to the positions.
We are waiting for the rest! especially the next one, game 5, constains a lot of decisions and local tsumego particularly hard to anderstand.
With integer komi Alphago self play ending would have been breathtaking, no more suboptimal moves that doesn't change the end result if both player can take shelter in the draw area...
Thanks to Michael for doing such a great job for us
The timing of resignation reminds me of Toya Koyo resigning in the match against Sai.
Will we see any more videos in this series from Mr. Redmond?? (Very enjoyable; thank you).
The link to the page to obtain the SGF file is broken. Anyone know of an alternative?
Thank you as usual for a great video.
11:12 the reason why it doesn't throw away ko-threats in close conjunction with kos is because it sees the ko in the game tree and can see the result of throwing away a ko-threat when it explores the ko. If there is no potential ko close by, the effect of throwing away a ko-threat will not be apparent since it doesn't have a ko to simulate.
In its evaluation throwing away a ko threat might be advantageous as it simplifies the game.
I also thought this first, but the value network should take who has the most ko threats into account. If throwing away ko threats simplifies the game it will still only do it when it is certain of winning.
What these first four games suggest to me is that high speed reading is really the key to Alphago's success, not deep strategic vision. Whatever that means. Anyway it seems like the policy inputs are giving enough hints to allow the right subset of positions to be crunched. That doesn't mean the policies themselves are all that great.
Alphago doesn’t read deeply, its power comes from pattern recognition that is achieved using neural networks. The creators go over this in their paper. In fact without reading at all (no simulation, no search) it was still better than every other go program at that time.
It's been 5 days. I am having Michael Redmond withdrawal symptoms. Help!
wow that was intense! I like it! Keep it up!
28:26 AlphaGo fan attacked Michael Redmond!
Chapeau @ Mr Redmond.
That dog is cute
I get the sense that Alphago processes the idea of yose very differently from humans because of the power of its reading ability. Then again it's a bit disconcerting to think that it is relatively weak in taisha & nadare and that weakness is its current level in yose, fighting, tesuji, etc. That would imply that there are several revolutionary advances in AI strength required before getting to 'perfect' or 'solved' 19x19 go.
Getting anywhere near 'perfect' game of Go, would require a computer of size bigger than the observable universe. So it's not going to happen anytime soon for this reason alone.
That statement is true for brute forcing Go, but is that really necessary? Suppose a quantum NN would be able to eat a board position and spit out the winning percentages of each legal move. I wonder how it would handle triple ko.
Found another link at the following page: www.usgo.org/news/2017/08/alphago-alphago-game-4-reminders-of-go-seigen-escalating-trades-and-semeais-and-a-final-ko/
also this video is been released before.....
16:33 had to play that one back
Does Alphago lose with white ever?
Black will get lucky next game.
Go Seigen? Dosaku? Just say AlphaGo is Sai, we all know it...
i heard that alphago was trained via amateur games ... why not pro games? and If it were trained on pro games, then wouldn't it make sense for the Go Seigen fuseki to show up in the policy network, if it's seen it before?
That was only in the very beginning. The number of human games it learned from is now a tiny fraction of the number of self play games it learns from by now, so it's virtually impossible for a particular move that one person, or a few people, played to show up. I think the takeaway is that Go Seigen(and others) suggested a good move, and Alphago agrees that it's a good move.
It was not trained on pro games since there aren't enough, which would lead to overfitting. After being trained on 6d+ KGS games the policy network was playing at around 3 dan amateur. I don't think it would gain much from only train on professional games.
I do think professional KGS games were also in the dataset.
Guys, slow down! 1.5 hour review each day is a bit much. I can barely keep up!
Spread out the games a bit.
No it's awesome. Really love the rate of video releases. Nobody is forcing you to watch them on release day you know :)