Awed By AlphaGo - Review of Games 1 & 2
Vložit
- čas přidán 11. 03. 2016
- In which I look at several themes within the first two AlphaGo-Lee Sedol games. Some of which are follow-ups from weaknesses Myungwan Kim identified from earlier matches, and others of which were raised in the first game.
Here is the Myungwan Kim review of early AG games I mentioned: • Myungwan Kim 9p review...
Here is a reddit guide to many other videos and analyses: / alphago_and_lee_sedol_... - Hry
Loved the video. Thanks, looking forward to hearing more! :) Probably going to give this one a shoutout in my next video.
Very good video even for me who never played Go himself, but this whole AlphaGo thing really got me. Thank you, looking forward to your next production.
i know nothing of Go and i am LOVING this video, you're so clear and concise. this really is amazing. keep up the awesome dude, i do tutorials myself and i love your style and respect the skill i see displayed, both in the players and in your teaching. very well done! *claps*
+CynicatPro I blush! Thanks!
I'm the same. I know basically nothing about Go except the absolute basics barely, and I watched all the Lee Sedol vs AlphaGo live matches all the way through because I was so interested in the games.
Your hypothesis on AlphaGo and aji is pretty interresting and innovating. Thanks for that! :-)
Very good video, looking forward to see another one for the remaining games.
Fascinating, just fascinating. Great video, thanks.
A year on now, I find the early speculation on AlphaGo's capabilities pretty interesting. Specifically I find interesting the thought that, because AlphaGo plays/played a peaceful style that maybe AlphaGo wouldn't be as good at fighting as a human player. I think it's pretty clear now that the truth is more along the lines of AlphaGo is so confident it can win most any fight that it can look beyond it to solidify a more consistent win strategy. Basically it doesn't need to seek out fights, because it's confident it can win any that come along. And if it looses one here or there, it's confident it can put itself in what is ultimately a better position for that loss.
Thanks for the video, Brady. You have quickly become my favourite CZcams Go contributor. Keep 'em coming! (14kyu).
Very well put together video. Even though I've looked at lots of commentary already, you still managed to add some new perspectives, and it was interesting to watch from start to end. Thank you!
Even though I really like your normal format, maybe you could try different things like this video more often? You have a very clear way of explaining things, and it would be interesting to see that used to explain more than just your blunders ;) For example, the video you did recently where you spent lots of time on exploring variations of the common two-space high pincer joseki was very helpful, and I think it would be cool with a longer video devoted entirely to joseki study. I know the basic idea behind your channel from the beginning was to make short videos, but now that I've seen such a good long video from you I'm hungry for more :)
Great and easy to follow video! I hope you're going to follow up with a Game 3 analysis exploring the same themes
19:00 probably well known at this point, but because I just looked over this game on the Deepmind website, and for anyone just finding these videos like me, I'll mention it, just in case.
Move 80 was an intentionally safe move by AlphaGo, looking for the most certain victory rather than the biggest one. At this point, AlphaGo put its odds of victory at 74%, an extremely conservative estimate, where 70% means the game is nearly decided. This was a display of confidence; AlphaGo knew virtually nothing Lee Sedol could do would turn the game around, and hung back to wait for an inevitable victory.
I am not sure about the long term aji argument regarding move 80. My understanding is that DeepMind website offered a SGF files including AlphaGo's expected play variations. AlphaGo I remember insisted that black move out around there instead of F3 approach (kakari) on the left bottom at move 79. Hence for AlphaGo, move 80 was the urgent point (kyuba) rather than a safe slack move. Today (2020)'s strongest AIs seem to think that double kakari is not necessarily huge, as humans thought back then, and that may be one factor for AlphaGo's not thinking white's keima response was a must.
Very interesting. Please do the other games aswell! :)
Thank you so much for the analysis
Such an amazing time to be a go player. I can't help but feel sad for Lee Sedol though. I hope he manages a win against Alpha Go.
Great ideas worth analyzing.
Fantastic video man, if you could do something similar for games 3-5 that'd be sweet.
This was great for someone who doesn't know Go.
As for improvement, you only need to mention twice that you'll be linking more detailed breakdowns in the description :P
+Colter Hawkstetler hehe. yeah, that and making sure the links actually work would be improvements. :) thanks!
Someone once told me that on your deathbed someone hands you a nine page pamphlet that explains the whole thing. Thanks for the show. Bob
Thank you so much and for your explanations for the non-players!
I know very little to nothing about Go, your videos ended up being recommended to me by the complicated algorithms of youtube. But your channel's videos intrigued me! I like various strategy games and principles behind them. I will probably be watching these videos every now and then just to be more acquainted.
Question is , have you played alphago? I don't know how much a skill gap can be demonstrated or measured. I figure a complete newbie like me would be easily dismantled in just a few turns. Someone more competent would surely do better and grandmasters have these games.
On another note,
I enjoy your explanations of why certain moves are played and how many different pathways you can get to those conclusions.
Hi Brady - you mentioned more detailed analyses - but I'm having trouble finding them on the reddit database you linked. The 'post-game analysis' links seem to just be newspaper articles without in depth discussion or board position images. Would be great if you could link your favourite analysis. Or anybody else. Thanks a lot for your videos. They're great.
It somehow feels like when the world's top players of a given game are seriously challenged for the first time by a computer, which have been notoriously weak during prior decades, these players somehow get kind of "intimidated" by the computer and try to outplay and outsmart it, rather than playing as they normally would against a strong human opponent.
Not that it has happened many times, of course, because this kind of thing is such a rare happenstance, but the previous clear example of this was Garry Kasparov's first serious tournament against Deep Blue where, in at least one game, Kasparov tried to outsmart the computer and throw it off by making an unusual, but sub-optimal, move, and the computer proceeded to destroy him. This was the first time that a computer was playing a serious tournament with long time controls against the top player in the world, so the situation is extremely similar.
Interestingly, computer chess has advanced enormously since the times of Deep Blue. Deep Blue was a dedicated supercomputer that could evaluate 200 million positions per second. Modern chess engines, running on a top-end PC, typically can calculate something like 10 or 20 million positions per second (depending on the speed of the CPU). Yet modern chess engines, running on such a regular PC, could easily beat Deep Blue without much problem. Heck, a modern engine running on a mid-range PC capable of calculating something like 5 million positions per second, could easily beat Deep Blue.
It's all because the computing algorithms have advanced so much. Modern engines can find better moves with only a fraction of searching needed, and are thus much stronger. (In fact, chess engine developers who know a bit of the internals of Deep Blue say that a few simple modern algorithms implemented into the original Deep Blue would immediately make it 100-200 ELO points stronger easily. And that's just with a few simple changes.)
I think that the attitude of chess grandmasters towards computer chess seems to be a bit different from what's happening in the go world. Chess grandmasters generally never play against computers, nor do they use computers for their own learning. Computers crush them even at their best, and they seem to have this thinking that they have nothing to learn from computers. This seems rather opposite to what the go masters are saying, as they seem to be excited about the new styles that Alpha Go is using.
What you said that the move was a slack, well, in Alpha Zero games of chess, the machine did some moves that are changing the way we see chess. For example, AlphaZero sacrificed a lot of pawns in the game against Stockfish because they were on his way, and in chess that works sometimes but not at that level, but it worked for Alphazero. So i believe we have a set of ideas that become paradigms and that stops us from seeing farther. I don´t know much about go so I don´t know if this is the case, but, as you say, we will see in some time.
AlphaGo do have a sort of database. It builds it learning and adapts but it still stores that somewhere.
Great video! about aji: you gave examples where AlphaGo seems to assign a high value to the aji within his territory, but it is less clear that it has an original valuation of the aji in his opponent's shapes. So at this point, it is difficult for me to see if what you say is in line with what other people are saying about AlphaGo tending to make simplifying moves if possible, or if you are saying something qualitatively different. This could be decided by studying how AlphaGo picks variations breaking his opponent's shape, or its openness to sacrifice. I'm a rather weak player, so I can't say myself, but from pro comments, I would tend to say that AlphaGo simplifies the game rather than over-values aji. What do you think?
+Thomas Carette I've been thinking on your point, and it is a good one. The two viewpoints are very similar, but my question arises from the widespread criticism of AG's tenuki in lower-left to reduce aji in the upper left. If we question whether it might not be a mistake at all, then it maybe that pros are undervaluing aji, or underestimating AG's lead. I think my concluding thoughts that maybe it views aji as a win percentage reducer, rather than an absolute point reducer sort of supports both views. Interesting stuff!
Im a chess player, so deep analyses into sequences are meaningless to me. Your commentary of these games through concepts however are fascinating and clear to me. I enjoyed this video and also applaud you for sticking to your theme and leaving deep commentary to others while you focus on themes.
it also has a database like predictor of most likely human moves, based off of thousands of amateur games
enjoyed your video, Just curious if you thought Lee Sedol's victory in Game 4 was a cultural concession by the AlphaGo team?
+James Hawkes Interesting idea, but no I don't think so. I discussed that game in some length in my newest video. Thanks!
8:30 This strikes at the question of how any mind -- silicon or meat -- learns Go. Alphago is layers of neural nets that have been trained to recognize strength and weakness in patterns of stones. That's what neural nets do: they recognize objects, ideas, patterns.
Lee's statement makes sense: he sought to play a move that Alphago would not recognize. But I think he underestimated the sophistication of his opponent by believing Alphago could not assess a novel stone arrangement.
the q14 variation has been around since 2010
Is the program (AlphaGo) Updated? Every game?
Nice video, thanks :)
+Alseki7 Your MK review link is dead.
+Alseki7 Fixed, I hope. Thanks!
+Brady Daniels No, it's still truncated.
+Klaus Weidner +Alseki7 Okay, I forgot that when you fix one link in the info, you need to redo them all before saving. I've done it again, and hopefully both links work. Let me know. (I am far from a YT expert, sigh) Thanks!
Alpha go is tireless emotionless and can focus on all areas of the board equally. In light of that I think it's play style isn't that surprising.
is watching alpha go more fun than hiruko no go?
if you could just use a better microphone, a sensitive and directional one
Brady I bet you have to get a new computer mouse often! lol
Playing an inno early in the game means nothing to a game engine which gets it's strength of play largely from neural net(s). An engine that uses a neural net actually has a 'feel' for the game in a similar way that a human does. Therefore your chances will be better if you stick with your normal style.
The komi (the points black gives back to white) is 6.5 in Japanese or Korean rules. Its only 7.5 in Chinese rules.
The Lee Sedol vs. AlphaGo 5-game match was played with a 7.5 komi to white (you can hear it announced if you watch the Deepmind documentary on CZcams). I'm not sure but I think the Japanese Go world moved generally to 7.5 komi some time ago as analysis suggested 6.5 komi was too much of an advantage to black.
Just like to point out a technical error you had. Alpha Go does in fact have a database, similar to how chess programs have an opening book. The AI only takes over once the precalculated sequences are broken.
I like you
You speak alot, but sometimes it's nonsense. You said that alphago loves to be aggressive especially when it has an advantage, which is actually the opposite of what alphago does. When alphago is ahead, it doesn't take risks.
This video was made in the middle of the AG - Lee series. AG has changed over time, as has my opinion of its play. I drew the same conclusions as you a couple of years ago in the video, AlphaGo - Whatever You Do I Wrong. Thanks for watching.
Alphago was backed by a database.
Database 😂😂
I liked your videos before, but now it seems to be all about computer software which doesn't interest me. I am sure many find it fascinating.