These reviews are fantastic. Even though these games are ridiculously far above my pay grade, I enjoy them very much. Thank you very much Michael and Chris. Keep up the good work.
In the original series, the videos were too short to really enjoy the game. And there was too much Chris and too little Michael. Now, I feel the length of the review and the balance is really good. Great job on improving the content guys!
Have to say Michael is doing a great job. Maybe too self-critical, many much lower amateur players are giving more general commentary on the playstyle of AlphaGo, but Michael's comments have also been on point. Just don't be afraid to say your mind!
Isn't the tendency to play forcing moves and answering them with other forcing moves somewhat like playing sente for sente in the endgame? Like the basic Cho Chikun endgame problem with hanetsugis.
It would be great to know alpha's evaluations here. Did we really start finding mistakes by alphago that change the game result? This video is the first instance of such a claim I've seen since the Lee Sedol win. If this is true, go engines still have a long way to go and it's quite likely that "perfect play" looks radically different from the way humans play now as opposed to just some preferences in joseki choice strik8ng us as novel. Great thanks to Michael Redmond for sharing his enormous amounts of work on these games.
I'd love to see Michael get a chance to show us how he'd complete the kill with his hanging connection. Something tells me AG will have something to say about that.
I'm sure he would love to test it against the machine. We are lucky that he has the skill and fortitude to make any claims at all. "AG played it, so it must be right" would be useless.
Couldt he reason alphago likes to jump into the 3,3 point be to define the corners so it knows how to attack it from the sides. Jumping into the 3,3 looks like it will result in some kind of stick formation which is not good for territory, so maybe alphago likes to force its opponent to make a stick that wont give it much territory?
The wall that the opponent gets after 3-3 doesn't have any eyes. Maybe alphago thinks the wall is actually quite weak. Still it keeps playing 4-4 so can't be too bad.
I agree, from my limited perspective, that Alphago appears to go after sticks and that may be coupled with its ability to invade.. Perhaps influence without eyes has to be used very carefully.
My idea about why AlphaGo makes forcing moves for no reason sometimes is because it gives an equally big chance to win if you make a forcing move and then your real move as if you did in in another order. Therefore it is like random between two equally good moves. It might also be that it is pushing something bad beyond it's horizon if deepmind didn't fix this in some way.
Cho U (a Japanese top pro), after seeing many Monte Carlo based programs like AlphaGo, says that it is really difficult to define a "perfect play" or "god moves" in other words. The reason is that if the game is off the balance, there are a massive number of multiple paths leading to the same result. In go, thanks to the rule definition of "score", it is relatively easier to define god that maximizes score via backward induction play. However, in games like chess, there is no unified definition of score, hence the play of god, particularly for the losing side, is far from evident. A Japanese chess programmer responded me on Twitter on this point saying that the number of moves required till the checkmate is a candidate of what we can refer to as score, and it also fits human intuition. The similar point is actually discussed in game theory in general regarding the controversy of backward induction algorithm, when the opponent is boundedly rational.
Mr. Redmond, if you or Mr. Garlock are reading this, I have a comment on the "mistakes". I am no good at Go (2 kyu) but I have a background in AI and I have watched the development of chess AIs (which of course use a different type of algorithms than Go). You seem to be surprised with many points where you think AlphaGo is not playing good moves, and yet, as we know, AlphaGo is stronger than any human player. How can this be?? Here is what happens in chess. Chess programs are far stronger than the best humans - they easily beat the strongest grandmasters. Yet, when grandmasters analyze computer moves, they inevitably find errors. Well, yes these errors are real, they are artifacts of the computer "way of thinking". The point is, the AIs, despite having weaknesses, are so far superior than humans in other areas, that more than makes up for the weaknesses. It is quite possible that the same is happening here. AlphaGo makes certain kinds of mistakes. Yet in other areas, is so much stronger than humans, that it still wins handily against them. Hence comes a corollary, same as in chess. If you combine the forces of Go pro such as yourself, with an AI - if you could analyze together, with the aid of AI, the result will be a far stronger player still, than either alone. Right now you don't have your own copy of AI yet, that you can use to analyze, but it will probably happen soon. At that point, Go will benefit and advance. This last point is actually different than in chess - in chess, the name of the game is "avoid mistakes", but that is not easy to do for humans, even though they know that is what they have to do. In Go, the point is more like "what is the better intuitive way to play", and these are the kinds of questions that will start to be answered better, once the humans and AIs join forces.
I actually find it helps a lot to engage my critical thinking when there's some ambiguity, really drives your comprehension to work out for yourself what the point is that's being made, and it expands my appreciation of Michael's sometimes terse comments in general.
It's interesting, many years later when Katago is much stronger than Alphago was, it definitely doesn't play the moves Michael Redmond thinks are bad. It will 3-3 invade for example, but not nearly as eagerly or early, seeing other moves as better when the board is so open or when there are unsettled positions with high value results. And it's trained without any human knowledge. I wonder how Alphago Zero played on these points of criticism. Similarly, at 145, KataGo feels S7 clearly the best move on the board. And at 181, black could probably have lived at O8.
These reviews are fantastic. Even though these games are ridiculously far above my pay grade, I enjoy them very much. Thank you very much Michael and Chris. Keep up the good work.
You guys are on fire!
THANK YOU very much for the awesome reviews.
Amazing game.
The detailed explanation of 3-3 joseki variants is very useful to me.
Again, thank you both!
It's so amazing to see the game with all the variations! THANK YOU!
47:04 I think my head imploded here. Crazy to think of all the 47 games ahead... Thank you!
I love to take Go commentary out of context.
"Death is just everywhere."
"Yeah, it's very easy to die."
57:09 for anyone interested
In the original series, the videos were too short to really enjoy the game. And there was too much Chris and too little Michael. Now, I feel the length of the review and the balance is really good. Great job on improving the content guys!
Thanks so much for these, again. Amazing work you do and much appreciated!
Thank you both so much for these videos!
What would AlphaGo play if black had played G7 at the crucial moment Michael mentions? Can we get the Deep Mind folks to run the program for us?
16:19 is it too small for black to cut at b19?
Have to say Michael is doing a great job. Maybe too self-critical, many much lower amateur players are giving more general commentary on the playstyle of AlphaGo, but Michael's comments have also been on point. Just don't be afraid to say your mind!
Which version of alpha go is this? Is this zero?
Isn't the tendency to play forcing moves and answering them with other forcing moves somewhat like playing sente for sente in the endgame? Like the basic Cho Chikun endgame problem with hanetsugis.
It would be great to know alpha's evaluations here. Did we really start finding mistakes by alphago that change the game result? This video is the first instance of such a claim I've seen since the Lee Sedol win. If this is true, go engines still have a long way to go and it's quite likely that "perfect play" looks radically different from the way humans play now as opposed to just some preferences in joseki choice strik8ng us as novel. Great thanks to Michael Redmond for sharing his enormous amounts of work on these games.
I'd love to see Michael get a chance to show us how he'd complete the kill with his hanging connection. Something tells me AG will have something to say about that.
I'm sure he would love to test it against the machine. We are lucky that he has the skill and fortitude to make any claims at all. "AG played it, so it must be right" would be useless.
Of course perfect play is far from human and AG play
White is three for three in the series so far. Interesting.
There's a pattern
Couldt he reason alphago likes to jump into the 3,3 point be to define the corners so it knows how to attack it from the sides. Jumping into the 3,3 looks like it will result in some kind of stick formation which is not good for territory, so maybe alphago likes to force its opponent to make a stick that wont give it much territory?
The wall that the opponent gets after 3-3 doesn't have any eyes. Maybe alphago thinks the wall is actually quite weak.
Still it keeps playing 4-4 so can't be too bad.
I agree, from my limited perspective, that Alphago appears to go after sticks and that may be coupled with its ability to invade.. Perhaps influence without eyes has to be used very carefully.
My idea about why AlphaGo makes forcing moves for no reason sometimes is because it gives an equally big chance to win if you make a forcing move and then your real move as if you did in in another order. Therefore it is like random between two equally good moves.
It might also be that it is pushing something bad beyond it's horizon if deepmind didn't fix this in some way.
28:11 open the door! Ah - 29:00
Cho U (a Japanese top pro), after seeing many Monte Carlo based programs like AlphaGo, says that it is really difficult to define a "perfect play" or "god moves" in other words.
The reason is that if the game is off the balance, there are a massive number of multiple paths leading to the same result. In go, thanks to the rule definition of "score", it is relatively easier to define god that maximizes score via backward induction play. However, in games like chess, there is no unified definition of score, hence the play of god, particularly for the losing side, is far from evident. A Japanese chess programmer responded me on Twitter on this point saying that the number of moves required till the checkmate is a candidate of what we can refer to as score, and it also fits human intuition.
The similar point is actually discussed in game theory in general regarding the controversy of backward induction algorithm, when the opponent is boundedly rational.
Mr. Redmond, if you or Mr. Garlock are reading this, I have a comment on the "mistakes".
I am no good at Go (2 kyu) but I have a background in AI and I have watched the development of chess AIs (which of course use a different type of algorithms than Go). You seem to be surprised with many points where you think AlphaGo is not playing good moves, and yet, as we know, AlphaGo is stronger than any human player. How can this be??
Here is what happens in chess. Chess programs are far stronger than the best humans - they easily beat the strongest grandmasters. Yet, when grandmasters analyze computer moves, they inevitably find errors.
Well, yes these errors are real, they are artifacts of the computer "way of thinking". The point is, the AIs, despite having weaknesses, are so far superior than humans in other areas, that more than makes up for the weaknesses. It is quite possible that the same is happening here. AlphaGo makes certain kinds of mistakes. Yet in other areas, is so much stronger than humans, that it still wins handily against them.
Hence comes a corollary, same as in chess. If you combine the forces of Go pro such as yourself, with an AI - if you could analyze together, with the aid of AI, the result will be a far stronger player still, than either alone. Right now you don't have your own copy of AI yet, that you can use to analyze, but it will probably happen soon. At that point, Go will benefit and advance.
This last point is actually different than in chess - in chess, the name of the game is "avoid mistakes", but that is not easy to do for humans, even though they know that is what they have to do. In Go, the point is more like "what is the better intuitive way to play", and these are the kinds of questions that will start to be answered better, once the humans and AIs join forces.
Michael, use a pointer when you're talking! You're amazing, but can be difficult to tell which part of the board you're referencing sometimes
I feel like that would be distracting to be honest... But a couple of times I do find myself pausing for a bit and checking where something happened.
I actually find it helps a lot to engage my critical thinking when there's some ambiguity, really drives your comprehension to work out for yourself what the point is that's being made, and it expands my appreciation of Michael's sometimes terse comments in general.
It's interesting, many years later when Katago is much stronger than Alphago was, it definitely doesn't play the moves Michael Redmond thinks are bad.
It will 3-3 invade for example, but not nearly as eagerly or early, seeing other moves as better when the board is so open or when there are unsettled positions with high value results. And it's trained without any human knowledge. I wonder how Alphago Zero played on these points of criticism.
Similarly, at 145, KataGo feels S7 clearly the best move on the board. And at 181, black could probably have lived at O8.
Black got baited into taking white's upper left corner.