Redmond's Reviews, Episode 4: Honinbo Dosaku 9P v. Yasui Chitetsu 7P
ฝัง
- เผยแพร่เมื่อ 6 ก.พ. 2025
- In this commentary, Michael Redmond 9P and Chris Garlock, Managing Editor of the American Go E-Journal take a look at another game from the great master Honinbo Dosaku 9p. This is a game where Dosaku 9p played against Yasui Chitetsu 7p, and some of Dosaku's play in this game reminded Michael Redmond 9p of of AlphaGo's moves.
www.usgo.org/ne...
Of all these redmond videos this analysis by Michael is just so deep. The history is beyond fascinating and this is just amazing analysis. It really makes me want to learn GO. Clearly its the game that may never stop giving no matter how many years you play. Thete is no way you can imagine how much you add to our lives by putting these out there Michael. I think Chris agrees.
I love the two of you together, great dynamic!
Awesome commentary, what a great game. Thank you.
Thank you so much for this video: it was fantastic, really, even more than usual, though I could feel that way just because it was a Dosaku game :p and because of the confirmation that there will be more like this to come, yay! :)
Amazing! Thank you so much for this.
Love these Dosaku games and love all the insight into Chitetsu as well!
love Redmond talking about Go History :)
I like using the Master series as a light inspection into the individual games, and then while going over the game pointing out the "AlphaGo moves" while the Redmond Reviews are a full length analysis of games Michael likes. Hour format is great!
Yes! Thanks again for this
I am really sad Yasui lost after such an amazing performance, well thats life for you.
Suggestion -- How about creating a way for viewers to send questions to Michael Redmond? Here's my question. I know how to estimate the territory balance and the the power balance in a game. This give me an idea who's ahead, and what I need to win. Master/AlphaGo starts with estimating the probability of winning. I have no idea how to do that. How can I use Master's idea of a probability of victory to improve my plans?
I don't think you can. As far as I understand, Alpha Go makes its moves by considering three factors: The policy network; the value network; and doing monte carlo playouts.
The monte carlo playouts yield a probability of winning by playing out many thousands of games per move using random simulations all the way to the end of the game. That's obviously not something a human can do.
The policy network uses probabilities of a given move being played by a strong player based on a function that is generated by training a neural network on strong players' games (well, at least this was the case for the version that played against Fan Hui and was described in the Nature paper).
The value network also uses a neural network, but this time it is a function that yields a probability of winning by training the neural net with millions of self-played games (each of these games also going all the way to the end).
Since digesting games in this manner is not practical for human beings, the best you can really do is to develop your intuition for evaluating a position - which is roughly analogous to the probability of winning assigned by the value network.
Furthmore, Alpha Go doesn't really care about the margin of victory, only its degree of likelihood. However, the only way to apply this strategy as a human player is to develop a very strong end game - and not just the late end game, but in the early end game.
I guess if you wanted to try to take something away from the way Alpha Go sees the game, it would be to try to learn to simplify the game when you're ahead and to develop as strong an end game as possible so that you can win by the smallest of margins if necessary... That's something that professionals have always done anyway though - it's just that Alpha Go is even better at it than top human players are.
Thanks, Vladlevin100. I agree with you. I'd like to come up with something that does a human simulation of the computer reality. :-) Right now, I've got a good checklist of to think about before making a move. I wonder what solutions others have arrived at.
Phil what is that checklist you use b4 you move?
It's not AlphaGo, but I would think that playing around with Leela or Zen with a good machine and giving the AI a lot of time + exploring different possible moves, should be helpful for Michael to perhaps get a sense of positions that he mentions wanting his own AlphaGo to look at...
I second this. On a MacBook pro, Lizzie beats 2p players by enough to make said player think he could maybe beat her once out of 100 games. Lizzie on a upper tier gaming desktop is way stronger than that. Ive had her look at 9p pro games and say their moves are bad
Much better bumper.
You might add the [W] and [B] next the players names in the title like usual. :)
The Alphago move E2 ? which Alphago move E2? Show the move E2 or I will start blackmail !
43:17 Alphago sensei
Less Chris Garlock, more Redmond
Always trolling. :P
Well, his preamble was a bit long, but I liked what he did during the game
If a 2 minute introduction in a 1 hour 3 minute video is too much for you guys, you should stop watching TH-cam videos completely. Read books or get a teacher.