AlphaGo Zero vs. Master with Michael Redmond 9p: Game 8
Vložit
- čas přidán 13. 04. 2018
- Michael Redmond 9p, hosted by the AGA E-Journal's Chris Garlock, review the eighth game of the new AlphaGo Zero vs. Master series. Click here for the sgf file: www.usgo.org/news/2018/04/alph...
Video produced by Michael Wanek and Andrew Jackson. The sgf files were created by Redmond, with editing and transcription by Garlock and Myron Souris. - Hry
I believe I watched every review of Michaels games and the alphago matches at least twice. Thank you so much for the work you put into these videos. Thanks Michael, your way of explaining these difficult situations is just great, even a ddk player like me gets something out of it. And of course thank you Chris. Keep it up!
I've been taking small naps and watching this on pieces. Looks like Chris would have preferred to watch it the same way 😂 thanks guys
Thank you for taking the time to share your insight into these games. I sincerely appreciate your efforts, thoroughly enjoy your analysis, and rather like that you're able to fasten a comprehensible narrative to the often deeply strange and brilliant play we see these AIs churning out. A testament, I think, to both your communicative ability and the profound nature and beauty of Go.
Lovely commentary, thanks Michael and Chris!
Fantastic video as always. Thank you :)
Awesome game and assessment - thank you to Michael and Chris.
That was very interesting commentary during the unintended "starting from zero" pun by Michael (glad you caught that Chris...I had the same reaction you did). I can see why pros wouldn't want to be reading anew from every move their opponent played, so these types of matches would be very rare to see at any level of play. It's a game like this that makes me want to see more AI play, since they aren't going to back away from variations that involve deep reading and large branching. Clearly we aren't just getting high level matches from AIs, we're getting games that humans would likely never even attempt to play. And that's not a knock on people, it's just due to understandable considerations for human beings like time constraints and the need to have some mid- to long-term plan in order to have a chance to win the game.
Which is to say that AI are not human. Sometimes the most surprising thing is when they do play like people.
I have a theory on why AlphaGo loses points at the end. From its experience, it probably concluded that a "won" game is a game where your score is greater and as close as possible to the opponent's. It was never actually taught the rules of go and what a "win" means, so it developed its own notion of winning, which is different from that of humans.
It's very interesting looking at real time move percentages from strong AIs, such as Leela Zero. There are some visualizations that show each move it's considering and how many moves it's read out from that position. Often, it will read variations from one move for 1000+ positions, then switch to a different move, meaning that it does in fact make mistakes with its initial "instincts", but is able to brute force read them out and try a different variation if it doesn't like the results. Humans can only read out 10s or 100s of moves, AIs can read tens of thousands or even more in the same time.
I meant 10's or 100's counting every move in every variation they read, not (usually) in a forced sequence. All the permutations count.
But it's interesting the discussion about these very fluid plans makes me wonder if the "normal variations" that are shown in SGF only represent what happens if both players stick to a single plan for a few moves...but maybe some of the moves that begin "losing" variations actually are completely viable, they just wouldn't play the rest of that variation.
I just caught myself pressing the like button before even starting the video
I want to see Redmond beat AlphaZero someday. That would be my dream
I want to see Michael PLAY Alpha Zero, or even Leela Zero, let alone beat! :)
Yes thank you! I suppose you chose to publish on a Saturday this time, because Friday the 13th would have been to dangerous? ;)
Poor Chris needed a shot of espresso.
or anti-histamines
or amphetamines
chris seems a little tipsy towards the end xd
I think the next step is to try to figure out what Alphago is actually optimizing for when it makes those "mistakes". Obviously it is "chance of winning", but how exactly is that measured?
Michael, you might soon get your wish to have a AI tool comparable to AlphaGo Zero, to analyze games with cooperatively. There is an "open source" AlphaGo Zero project underway. The neural network source code is written, based on the DeepMind research paper. Of course, Google uses proprietary and very fast hardware to train. But this open source project, uses distributed computing to do the training, several hundred people with fast GPU's currently. It may still take several years to reach AlphaGo Zero level, but probably not longer than that. Maybe you can try and play against the current best trained version to see how good it already is. The project is at github.com/gcp/leela-zero
The people who host LZ on OGS would be absolutely delighted to have Michael play against it. Michael, if you see this comment, please get in touch :)
Amen!
I actually asked David Silver, the lead programmer of DeepMind, if they were going to publish their most recent network weights. He honored me with a reply, that they did intend to save the trained state in some way, but as far as I know, it is not in the public domain.
Better than pro-level intuition, coupled with reading out a few thousand moves before playing every turn is a pretty fearsome combination. To put that in context, a human joseki might have the Go masters analyze a similar number of moves across hundreds of possible variations from a given starting point over the course of however many years that particular joseki is in use. The AIs are effectively inventing and thoroughly analyzing a new joseki starting at the current whole board position every single turn of every game they play!
As for the level of intuition the AIs have, I dare say that if a human pro were to somehow play 10's or 100's of millions of games against equally strong opponents before dieing of old age, they would be be able to reach and probably surpass the AI. Alas, a human lifetime isn't long enough for this.
Good point! Why would they program an AI so that it wins by far fewer points then it could?
Is it so the next gen has the ability to improve easily?
Starts at 0:59
Word of the day: obstreperous
After two failed marriages I never understood why husbands and wives call their spouse "honey," but now that I've watched Michaels GO videos, I realized they are saying "Hunai," because it's the Japanese term for someone who is always preventing you from going where you want to GO!
And Chris, we love your enthusiasm for the best game ever invented, and also for the 2nd best (tennis!) but you need to use the mute button. Coughing into the mic is extremely rude to the audience. I feel I'm going to catch CO-vid watching you and Michael's CO-VID.
Teoría no computación al de la mente
The reason why it seems like Zero & Master play silly moves when they are winning/losing is because the algorithm makes it impossible for them to differentiate between good and bad moves. Zero reads that no matter what he plays, the evaluation is always the same: 100%. It basically plays what it happens to read most.
Same thing happens when you're losing. No matter what you play you lose, so you start playing self ataris etc. It's the engines' way of saying "This game is over, whatever I do doesn't matter".
This could be fixed by setting a higher resignation percent. It seems AG resigns only when it's 99% sure (or more) that it's losing.
It actually resigns when it's 90 % sure that it's losing. Should probably be 80 - 85 % though.
The humans no is diferent the machine y think in a way and we act diffently Roger penrose