@@qaz120120 It's commentary. Think of it like watching any professional sports on TV. There are usually two people on commentary. One of them is extremely skilled at breaking down the game, giving you a step by step analysis of what each play does. That's mainly because they themselves were an athlete of the sport they're commentating on. The other is more talkative and engaging, giving you the play by play as it happens and giving the analyzer someone to bounce ideas off of. The same applies here. Michael is the breakdown guy and the other is the play by play guy.
I think a lot of his questions and comments are really insightful actually. Strong pros need to be reminded of things that seem obvious to them sometimes
Chris Garlock and Michael Redmond, the enthusiast and the master that's happy to share his knowledge about the game. They make it very easy to become interested in Go and build such a natural duo, something classic I've seen in sports /competitions in the past. I hope there's gonna be more footage of the two together, as I've watched the Lee Sedol vs AlphaGo series and was astonished by their commentary. Greetings from Germany!
I have never played a game of GO in my life but can not get enough commentary from Michael Redmond and Chris Garlock. Michael's talents are on full display commentating live games which I wish we could get some more of. I've watched Lee Sedol games like 5 times each... Thank you guys!
Just like the games of previous extraordinary Go players, which challenged the concepts of traditional play, AlphaGoZero games deserve to be comprehensibly analyzed in detail to advance our perception of Go and it's evolving possibilities. Michael Redmond's commentaries are excellent in providing detail and overview in context. Thank you AGA.
This is incredible and it's sort of thematic even with Limited AI Chess Engines matches where the better engine usually only has one weakness instead of two and has a subtle positional advantage. The use of simple, straightforward approach by Zero seems so obvious when played - except when you have to think of the moves on your own. Amazing and thank you for the analysis Michael. Great commentary!
b e9 at 25:41 : This move here is actually dealing with the weakness at 'A', the *cut* at 'A'; *it's making it less important* To be honest, I don't think i've heard that before. "Deal with a cut by making it less important." Seems obvious now that Redmond has said it ... this is why i watch these things.
Much appreciated Michael (and Chris). Keep these awesome commentaries coming. Surely Deep Mind will turn this into a money maker. Pros would be lining up to test moves out on Alpha Go Zero and get the winning %.
They are actually not pursuing it! They released the details of the algorithm in their Nature paper and are moving on to applying the general idea to other fields.
The details of the AI system won't translate to most human players being able to develop their own Go AI. However if Deep Mind released the software and the trained neural system itself free of charge so that players can install it on their computers then of course the serious Go competitive players could go ahead and test ideas with the AI guiding them. In spite of the fact that the hardware used by Deep Mind is faster it doesn't matter. It just means people with ordinary PC's would have to let the AI run for a longer time to test ideas. People do the same thing with chess sometimes letting a computer run for days on a position to see what the PC comes up with. Plus people aren't restricted to that approach. Another thing they can do is present the AI with various variations and see which one(s) it likes most. Those are only two ways to use the AI to test the soundness of ideas.
Great commentary again, Michael, thank you! Re: the variation (at c.8 minutes) where black tries the hane at D3. Then W E3 (as you say this seems inevitable) up to W fixing up the shape at G3, you have black sliding at B6. But I found the continuation after D6 (nose) far from clearly good for W. I would be very glad to see the continuation there, doubtless I am missing something...
For a lowly kyu player like me, it's really informative to have a high-level player like Redmond explain what's probably going through the computers head when it plays a certain move.
I feet AlphaGo shows us two very interesting insights from two points of view. One is that the AI proved and confirmed what human has discovered for 2000 years by discover very similar game play as how human would play at a very grand sense of view. The other is how a 2000+ year old game helped computer scientist and engineers to prove AI theories in a very measurable way. And because there are very advanced human players, AlphaGo shows how strong and "smart" it is. Without such a controlled environment with advanced human players and theories, it would be very hard to comprehend what AI has done. This is why I don't view this as a threat to human intelligence. rather a confirmation of human ability. Just last a calculator can compute much faster than a human, but it also shows how advanced human mathematics is.
What you said is logical. However this is just an inkling of the potential of AI so I think that there is much more to come. For example it seems apparent, (to the experts at least), that even Zero may not be playing every move perfectly. AI will certainly continue to improve because humans will definitely continue to improve the AI systems plus computers will continue to get faster.
You hit the nail on the head. This idea of "man versus machine" is such an egotistical way to look at it. This, just like the calculator, is man PLUS machine.
It would be nice if they gave the information about these engines a little bit. Is the Master one who played against Lee Sedol? Same as who played in 60 games series against top 15 players on an online server? And the Alpha Zero one is a version created without using data from human games ?
One of the key takeway is that they have very good reading and do not face like human does on time. That give a better game to watch and analysis. Just not for winning. As those days in 1940s-1980s, Just good to watch.
To answer the question at 15:00 i believe that Master was the same program as the Lee Sedol version but instead of being trained on human games it was trained on AlphaGo vs AlphaGo games from the Lee Sedol version
Wow, Zero really loves those star points! Even later on into the game, if the opponent hasn't placed a stone there, it looks like Zero will grab it. There must be a lot of value in controlling that point according to Zero's calculations.
Plus it's not just calculations, ie. not just a tree search, but the trained neural net(s) it uses to guide the tree search by sorting the moves at each ply in the tree so that it probes the variations that seem the most promising. Also at the end of each variation it uses at least one neural net to evaluate the position since of course even the computer hardware used by Deep Mind can't search every variation to the end of the game, (until the endgame), therefore the AI has to stop somewhere in each variation and evaluate, (make a judgement call), of the value of the last position in the variation.
Seeing how AlphaGo seems to be inclining towards nearly even results rather than winning by a huge margin, my guess is it may be artifact of its learning method. During learning periods, longer games have bigger impact than short games (more positions to adjust the neural network with) and since the program will give up (and make the game shorter) when it sees clearly it's lost the game, it might get accidentally taught to keep the game going as long as possible. On another topic, it might be interesting to see Chess Zero too.
copied from reddit: I think the notion of winning percentage that AlphaGO calculated is misunderstood here. AlphaGO isn't about playing "strong" when he's losing and "safe" when he's winning, it is about always, consistently about playing the "best move toward winning". So that 1 point difference is "actually" or "statistically" harder to overcome than 100 points difference in your case. For example, if AlphaGO is present with 1 points with 99% chance vs 100 point with 80% chance, I think this is what AlphaGO is having in its mind before it makes it move: If I make move at A, There are 20 out of 100 possible variations that opponent could do to reverse my 100 point difference If I make move at B-- My point difference is down to 1 but there are only 1 out of 100 possible variations that opponent could do to reverse my 1 point difference. So basically 1 point route is "actually" harder for opponent make comeback tesuji than the 100 point difference. www.reddit.com/r/baduk/comments/4a3mlf/possible_alphago_weakness/
The Deep Mind team has been promising to leave us with a "teaching tool" which may either be a playable version of AG or perhaps something that can simply analyze a board and generate a heat map showing the most attractive moves. Either way, this should be enough to eventually be able to ask it "what if" questions about all of the mysterious moves we've been seeing and use that to figure out stories that will be the equivalent of getting AG's thoughts. That's my guess anyway.
I think Michael is right to doubt that the 50 self-play games that he has been working with, involve the same player that faces Zero in this game. In the Deepmind video where Ke Jie and Fan Hui review (I think game 2) of the future of go summit alongside AG, Fan Hui admits that the version Ke Jie fought against is stronger than Master. But in the recent "Mastering Go w/o Human Knowledge" publication, the authors state that Master was the same version that fought Ke Jie. In the paper, Master is still winning some games against Zero. I wonder what kind of result would emerge if AlphaGo Ke Jie version played a 100 game match against Zero.
Has anyone considered the possibility that an early 3x3 invasion is *not* actually a good move? I seem to recall that AlphaGo has the ability to read ~50 moves ahead. Given that the putative disadvantage of an early 3x3 invasion is giving influence for territory too early, it seems likely to me that disadvantages would require very deep searches in order to detect, possibly more than the search depth allows at the very beginning. Given that AlphaGo does not have any heuristics hard-coded in and only relies on computation perhaps it is falsely thinking it is a good move, whereas humans don't need to read a hundred moves in to appreciate that jumping in early on gives a power disadvantage much later in the game.
The thing is AlphaGo does not work the way the traditional gaming AI's do. The traditional AI's rely heavily on how much of the game tree they can read in a given time. So what you are speculating does apply to the traditional AI's but not so much to neural net systems. Neural net systems achieve a type of 'feel' for Go which is similar to the feel that humans achieve. Add to that the ability to read at blinding speed you get an entity that leaves humans in it's wake.
Is there any way we could add a rule to things like Alphago? Like it knows the rules of the game, obviously, but what if we essentially lied to it and told it that one of the rules of go was that you had to win by the most you possibly can or by 20+ points or something like that. I wonder how it would play then. I just want to see the AI not give points because it thinks it'll win.
It's not so much that it gives up points because it knows it will win anyway, as if it's just careless and confident, it's that it thinks that small sacrifices are worth making here and there if it increases the overall chance of winning by any margin. It's all about avoiding risk for it, making sure that it wins in the most amount of possible futures. If you trained AlphaGo to try to win by as much as possible, it would stop doing these sacrifices, but it would also lose a lot more games, because going for bigger plays means taking bigger risks. And if it's a Zero-type situation where it only trains against itself, one can assume that it would have to adopt a safer style to win at all against itself - if both of them were playing for big points they'd probably only end up winning by small margins anyway due to being evenly matched, and adjust strategy accordingly, which would end up looking the same as the current training, so far as I understand.
I am from the Dog clan. Spokesperson of the Dog his Greatness. Let me explain to your human how we dogs think. We play away quite a lot, just to dis each other, it is an attitude thingy, trying to cause a reaction. It is like “you think you are a tough guy? I don’t even care, what can you do man?”
Chris I know this video is 2 years old so my comment may no longer be relevant but if you are concerned with the lighting, rather than directing light on your face reflect light off something(s) onto your face. For example if you have a wall in front of you reflect the light off the wall onto your face will do the job.
I wonder if zero is playing conservatively and territory based because it is playing white. Is there a difference in its play style between white and black?
In the paper they use it for both training and playing, you can see the naked neural net is like 3k elo with search it's like 5k (pulled these values out of memory, might be bad). Is there some source of information more up to date than the paper I'd love to hear about it.
Michael, Chris, I have a comment on what Michael is saying on "jumping into 3-3 early", that you teach humans not to do this, yet AlphaGo Zero is doing it; how come? Well, I am a professional programmer, and we have many such situations in programming practice. In my area of expertise, I teach beginners to not do x, y, z, yet... I sometimes do these things myself... But, I do it from the position of "strength", "I know what I am doing" :) It's always like that - "don't do X", _unless_ of course, you know what you are doing! In this case, it seems, all humans, compared to AlphaGo, "don't know what they are doing". So, at our level, we can't do 3-3 invasion early. To do it properly, you have to be at the higher level, at the level of AlphaGo. For AlphaGo, even human pros are "beginners" in a way. Sorry about that Michael :)
You speak of, (in this game anyway), Zero playing the endgame in such a way that it seems to chose variations which barely win the game instead of playing for more points. I will speculate that there are two reasons, (perhaps linked?), why it plays that way. 1. Since it doesn't matter what the margin is, in Zero's, (and Master's play iirc according to videos I have watched), view it is irrelevant. 2. I think what is happening in the endgame, (not the late endgame of course), is the AI's look at possible approaches and when they look at an approach that barely wins they can read it out perfectly but for some reason, (perhaps due to the nature of the game of Go and perhaps partly due to the way AG analyses the positions), playing for more points is too murky and the AI's can't see clearly to the end of every one of those variations. If I am understanding you correctly I get the impression that humans will play for more points in the endgame because intuitively it feels like the winning chances are higher. That human tendency makes sense to me because I get the impression that even top players can't read most endgames out perfectly, (not the late endgame of course), (ie. every variation), to the end? Therefore they want to chose the variations that seem to win more points. I barely know the game btw I'm speculating based on my interpretation of expert's opinions.
Ai as far as my play is not great, beginner most days. If AI on OGS 20b-Turbo-Elf-v1 6Dan, wants to ask in as early as possible get cash then get enough control of the center (the 3-3 early allows the flow of the game to be contested more seeing as later stones alive are always in the way and dead ones lose any miracle reincarnation with each move) yose in Chinese count Dame now with 7.5 komi we must take Komi away from AI and bring Edo styles back 🤔 dosaku and shuei are the goats. wu liked the MIAi that shuei capitalized on. The 4-4 was never revolutionary
These commentaries and analyses are going to be Michaels true legacy in Go. Tireless work, and brilliant.
The other guy is dumb as fck though, why the heck is he in this movie?
@@qaz120120 It's commentary. Think of it like watching any professional sports on TV. There are usually two people on commentary. One of them is extremely skilled at breaking down the game, giving you a step by step analysis of what each play does. That's mainly because they themselves were an athlete of the sport they're commentating on. The other is more talkative and engaging, giving you the play by play as it happens and giving the analyzer someone to bounce ideas off of. The same applies here. Michael is the breakdown guy and the other is the play by play guy.
@@Crypton16 Thank for the explanation but the other guy is trying very hard to sound smart and actually very annoying for how "engaging" he is.
I think a lot of his questions and comments are really insightful actually. Strong pros need to be reminded of things that seem obvious to them sometimes
@@qaz120120 what's wrong with you?
My mind is always blown when Michael Redmond explains.. thanks... great game
Lars Rye Jeppesen not for nothing Redmond is one of the most popular pros in Japan, when it comes to explaining with great clarity.
Chris Garlock and Michael Redmond, the enthusiast and the master that's happy to share his knowledge about the game. They make it very easy to become interested in Go and build such a natural duo, something classic I've seen in sports /competitions in the past. I hope there's gonna be more footage of the two together, as I've watched the Lee Sedol vs AlphaGo series and was astonished by their commentary. Greetings from Germany!
I have never played a game of GO in my life but can not get enough commentary from Michael Redmond and Chris Garlock. Michael's talents are on full display commentating live games which I wish we could get some more of. I've watched Lee Sedol games like 5 times each... Thank you guys!
"If it was me, I would just die like a... like a sad dog" best quote ever :D
I enjoy watching these master and ground breaking pieces.
Just like the games of previous extraordinary Go players, which challenged the concepts of traditional play, AlphaGoZero games deserve to be comprehensibly analyzed in detail to advance our perception of Go and it's evolving possibilities. Michael Redmond's commentaries are excellent in providing detail and overview in context. Thank you AGA.
This is incredible and it's sort of thematic even with Limited AI Chess Engines matches where the better engine usually only has one weakness instead of two and has a subtle positional advantage. The use of simple, straightforward approach by Zero seems so obvious when played - except when you have to think of the moves on your own. Amazing and thank you for the analysis Michael. Great commentary!
b e9 at 25:41 : This move here is actually dealing with the weakness at 'A', the *cut* at 'A'; *it's making it less important*
To be honest, I don't think i've heard that before. "Deal with a cut by making it less important." Seems obvious now that Redmond has said it ... this is why i watch these things.
Great review!
YESSSS!!!!!!
Why do you have the exact same earphones?
Love this series btw!
Much appreciated Michael (and Chris). Keep these awesome commentaries coming. Surely Deep Mind will turn this into a money maker. Pros would be lining up to test moves out on Alpha Go Zero and get the winning %.
They are actually not pursuing it! They released the details of the algorithm in their Nature paper and are moving on to applying the general idea to other fields.
The details of the AI system won't translate to most human players being able to develop their own Go AI. However if Deep Mind released the software and the trained neural system itself free of charge so that players can install it on their computers then of course the serious Go competitive players could go ahead and test ideas with the AI guiding them. In spite of the fact that the hardware used by Deep Mind is faster it doesn't matter. It just means people with ordinary PC's would have to let the AI run for a longer time to test ideas. People do the same thing with chess sometimes letting a computer run for days on a position to see what the PC comes up with. Plus people aren't restricted to that approach. Another thing they can do is present the AI with various variations and see which one(s) it likes most. Those are only two ways to use the AI to test the soundness of ideas.
Great commentary again, Michael, thank you!
Re: the variation (at c.8 minutes) where black tries the hane at D3. Then W E3 (as you say this seems inevitable) up to W fixing up the shape at G3, you have black sliding at B6. But I found the continuation after D6 (nose) far from clearly good for W. I would be very glad to see the continuation there, doubtless I am missing something...
For a lowly kyu player like me, it's really informative to have a high-level player like Redmond explain what's probably going through the computers head when it plays a certain move.
Mindboggling stuff
great video
I feet AlphaGo shows us two very interesting insights from two points of view. One is that the AI proved and confirmed what human has discovered for 2000 years by discover very similar game play as how human would play at a very grand sense of view. The other is how a 2000+ year old game helped computer scientist and engineers to prove AI theories in a very measurable way. And because there are very advanced human players, AlphaGo shows how strong and "smart" it is. Without such a controlled environment with advanced human players and theories, it would be very hard to comprehend what AI has done. This is why I don't view this as a threat to human intelligence. rather a confirmation of human ability. Just last a calculator can compute much faster than a human, but it also shows how advanced human mathematics is.
Beautifully said. :)
What you said is logical. However this is just an inkling of the potential of AI so I think that there is much more to come. For example it seems apparent, (to the experts at least), that even Zero may not be playing every move perfectly. AI will certainly continue to improve because humans will definitely continue to improve the AI systems plus computers will continue to get faster.
You hit the nail on the head. This idea of "man versus machine" is such an egotistical way to look at it. This, just like the calculator, is man PLUS machine.
Its happening
Fan of Michael Redmond
This game is not just a rock, baby, it's a boulder!
It would be nice if they gave the information about these engines a little bit. Is the Master one who played against Lee Sedol? Same as who played in 60 games series against top 15 players on an online server? And the Alpha Zero one is a version created without using data from human games ?
Burning question: Were the sneezes around 21:00 from the cameraman, a bystander, or one of the pets?
Pretty sure it's a wife. My guess would be Michael's.
thanks for the Video, i dont understand alot but im im really new so i think its ok. Michael looks alot thinner, maybe his screen wasnt fitting?
One of the key takeway is that they have very good reading and do not face like human does on time. That give a better game to watch and analysis. Just not for winning. As those days in 1940s-1980s, Just good to watch.
Is there any footage of Michael Redmond playing?
quite a lot online, and he goes over some of his professional games with garlock on this channel in the Redmond Reviews vids
Hello
White is going to win easier with "copy cat" or symmetrical play strategy against Alpha go??
Because black komi 6.5??
To answer the question at 15:00 i believe that Master was the same program as the Lee Sedol version but instead of being trained on human games it was trained on AlphaGo vs AlphaGo games from the Lee Sedol version
in hindsight what about black forcing at E14 at 32:38 and solidifying the position. At 47:15 what about black forcing first and then extending?
+
Wow, Zero really loves those star points! Even later on into the game, if the opponent hasn't placed a stone there, it looks like Zero will grab it. There must be a lot of value in controlling that point according to Zero's calculations.
Plus it's not just calculations, ie. not just a tree search, but the trained neural net(s) it uses to guide the tree search by sorting the moves at each ply in the tree so that it probes the variations that seem the most promising. Also at the end of each variation it uses at least one neural net to evaluate the position since of course even the computer hardware used by Deep Mind can't search every variation to the end of the game, (until the endgame), therefore the AI has to stop somewhere in each variation and evaluate, (make a judgement call), of the value of the last position in the variation.
I think someone whispers tengen at 14:16 haha
that was creepy, it happened right when I was starting to drift off to sleep
Wtf..:O
It's sounds just a litttttle bit like xhu.
Seeing how AlphaGo seems to be inclining towards nearly even results rather than winning by a huge margin, my guess is it may be artifact of its learning method. During learning periods, longer games have bigger impact than short games (more positions to adjust the neural network with) and since the program will give up (and make the game shorter) when it sees clearly it's lost the game, it might get accidentally taught to keep the game going as long as possible.
On another topic, it might be interesting to see Chess Zero too.
That's the reason why we have evolved to have the longest sexual relationships possible for beings like us, haven't we?
copied from reddit:
I think the notion of winning percentage that AlphaGO calculated is misunderstood here. AlphaGO isn't about playing "strong" when he's losing and "safe" when he's winning, it is about always, consistently about playing the "best move toward winning".
So that 1 point difference is "actually" or "statistically" harder to overcome than 100 points difference in your case.
For example, if AlphaGO is present with 1 points with 99% chance vs 100 point with 80% chance, I think this is what AlphaGO is having in its mind before it makes it move:
If I make move at A, There are 20 out of 100 possible variations that opponent could do to reverse my 100 point difference
If I make move at B-- My point difference is down to 1 but there are only 1 out of 100 possible variations that opponent could do to reverse my 1 point difference.
So basically 1 point route is "actually" harder for opponent make comeback tesuji than the 100 point difference.
www.reddit.com/r/baduk/comments/4a3mlf/possible_alphago_weakness/
tldr its easier for AlphaGo to win by 1 point, so that's its only goal. It only cares about winning, not the point differential
any chance in getting master's or zero's thoughts on this game?
Nope.
010010010010000001110111011010010110111000101110
I have no mouth and I must scream.
This opponent is too easy. Who taught this master how to play Go?
The Deep Mind team has been promising to leave us with a "teaching tool" which may either be a playable version of AG or perhaps something that can simply analyze a board and generate a heat map showing the most attractive moves. Either way, this should be enough to eventually be able to ask it "what if" questions about all of the mysterious moves we've been seeing and use that to figure out stories that will be the equivalent of getting AG's thoughts. That's my guess anyway.
I think Michael is right to doubt that the 50 self-play games that he has been working with, involve the same player that faces Zero in this game. In the Deepmind video where Ke Jie and Fan Hui review (I think game 2) of the future of go summit alongside AG, Fan Hui admits that the version Ke Jie fought against is stronger than Master. But in the recent "Mastering Go w/o Human Knowledge" publication, the authors state that Master was the same version that fought Ke Jie. In the paper, Master is still winning some games against Zero. I wonder what kind of result would emerge if AlphaGo Ke Jie version played a 100 game match against Zero.
It's the same version (non zero) just with more training. Zero dominated Master so no human would survive vs Zero.
These are so fun. I kinda prefer Master's playing style. It's more interesting, imo.
Has anyone considered the possibility that an early 3x3 invasion is *not* actually a good move? I seem to recall that AlphaGo has the ability to read ~50 moves ahead. Given that the putative disadvantage of an early 3x3 invasion is giving influence for territory too early, it seems likely to me that disadvantages would require very deep searches in order to detect, possibly more than the search depth allows at the very beginning. Given that AlphaGo does not have any heuristics hard-coded in and only relies on computation perhaps it is falsely thinking it is a good move, whereas humans don't need to read a hundred moves in to appreciate that jumping in early on gives a power disadvantage much later in the game.
The thing is AlphaGo does not work the way the traditional gaming AI's do. The traditional AI's rely heavily on how much of the game tree they can read in a given time. So what you are speculating does apply to the traditional AI's but not so much to neural net systems. Neural net systems achieve a type of 'feel' for Go which is similar to the feel that humans achieve. Add to that the ability to read at blinding speed you get an entity that leaves humans in it's wake.
Is there any way we could add a rule to things like Alphago? Like it knows the rules of the game, obviously, but what if we essentially lied to it and told it that one of the rules of go was that you had to win by the most you possibly can or by 20+ points or something like that. I wonder how it would play then. I just want to see the AI not give points because it thinks it'll win.
It's not so much that it gives up points because it knows it will win anyway, as if it's just careless and confident, it's that it thinks that small sacrifices are worth making here and there if it increases the overall chance of winning by any margin. It's all about avoiding risk for it, making sure that it wins in the most amount of possible futures. If you trained AlphaGo to try to win by as much as possible, it would stop doing these sacrifices, but it would also lose a lot more games, because going for bigger plays means taking bigger risks. And if it's a Zero-type situation where it only trains against itself, one can assume that it would have to adopt a safer style to win at all against itself - if both of them were playing for big points they'd probably only end up winning by small margins anyway due to being evenly matched, and adjust strategy accordingly, which would end up looking the same as the current training, so far as I understand.
I am from the Dog clan. Spokesperson of the Dog his Greatness.
Let me explain to your human how we dogs think. We play away quite a lot, just to dis each other, it is an attitude thingy, trying to cause a reaction. It is like “you think you are a tough guy? I don’t even care, what can you do man?”
Will alphago zero or AI ,beat Enblox players?
Chris I know this video is 2 years old so my comment may no longer be relevant but if you are concerned with the lighting, rather than directing light on your face reflect light off something(s) onto your face. For example if you have a wall in front of you reflect the light off the wall onto your face will do the job.
I wonder would would happen if 2 alphago zeros play each other could there be a stalemate
Go doesn’t really have stalemates
Noice
I wonder if zero is playing conservatively and territory based because it is playing white. Is there a difference in its play style between white and black?
They're phasing out MCTS?
My understanding was that they already had (in Zero), but I might be mistaken.
In the paper they use it for both training and playing, you can see the naked neural net is like 3k elo with search it's like 5k (pulled these values out of memory, might be bad). Is there some source of information more up to date than the paper I'd love to hear about it.
Has anyone tried to mirror it's moves obviously we have a sophicasted enough brain to learn off this ai just like its doing to us
Michael, Chris, I have a comment on what Michael is saying on "jumping into 3-3 early", that you teach humans not to do this, yet AlphaGo Zero is doing it; how come?
Well, I am a professional programmer, and we have many such situations in programming practice. In my area of expertise, I teach beginners to not do x, y, z, yet... I sometimes do these things myself... But, I do it from the position of "strength", "I know what I am doing" :)
It's always like that - "don't do X", _unless_ of course, you know what you are doing! In this case, it seems, all humans, compared to AlphaGo, "don't know what they are doing". So, at our level, we can't do 3-3 invasion early. To do it properly, you have to be at the higher level, at the level of AlphaGo. For AlphaGo, even human pros are "beginners" in a way. Sorry about that Michael :)
white = AlphaGo Zero
black = Master
You speak of, (in this game anyway), Zero playing the endgame in such a way that it seems to chose variations which barely win the game instead of playing for more points. I will speculate that there are two reasons, (perhaps linked?), why it plays that way.
1. Since it doesn't matter what the margin is, in Zero's, (and Master's play iirc according to videos I have watched), view it is irrelevant.
2. I think what is happening in the endgame, (not the late endgame of course), is the AI's look at possible approaches and when they look at an approach that barely wins they can read it out perfectly but for some reason, (perhaps due to the nature of the game of Go and perhaps partly due to the way AG analyses the positions), playing for more points is too murky and the AI's can't see clearly to the end of every one of those variations.
If I am understanding you correctly I get the impression that humans will play for more points in the endgame because intuitively it feels like the winning chances are higher. That human tendency makes sense to me because I get the impression that even top players can't read most endgames out perfectly, (not the late endgame of course), (ie. every variation), to the end? Therefore they want to chose the variations that seem to win more points.
I barely know the game btw I'm speculating based on my interpretation of expert's opinions.
monobrow
36:51 Chris does not approve
Ai as far as my play is not great, beginner most days. If AI on OGS 20b-Turbo-Elf-v1 6Dan, wants to ask in as early as possible get cash then get enough control of the center (the 3-3 early allows the flow of the game to be contested more seeing as later stones alive are always in the way and dead ones lose any miracle reincarnation with each move) yose in Chinese count Dame now with 7.5 komi we must take Komi away from AI and bring Edo styles back 🤔 dosaku and shuei are the goats. wu liked the MIAi that shuei capitalized on. The 4-4 was never revolutionary