AlphaGo Zero vs. AlphaGo Lee with Michael Redmond 9p: Game 2

Michael Redmond 9p, hosted by the AGA E-Journal's Chris Garlock, reviews the second game of the new AlphaGo Zero vs. AlphaGo Lee series (note that it's Game 11 in the AG-Lee series, but the second game in this video series). Check out the sgf file here: www.usgo.org/ne...
Video produced by Michael Wanek and Andrew Jackson. The sgf files were created by Redmond, with editing and transcription by Garlock and Myron Souris.

[@Kaiasky]

This is one of my favorite games just because of how clearly it tells the story of the 3-3 invasion's development

[@kfm1242]

Michael's ability to enlight stuff for me that would be soooo much over my head in my games is amazing.

[@tomewall8982]

"The shoulder hit heard round the world." Nice.

[@kwilatek]

20 vs 40 block refers to the size of the neural network that evaluates a single position on the board. 40 block version has a bigger, deeper, more sophisticated neural network for evaluating a single board position (and recommending the next move).

[@DiapaYY]

In human terms you can say the difference is how complex it's intuition is, but both can read equally many variations. (20 block probably even reads faster, but 40 blocks doesn't need to read as much since it's intuition evaluates if it is good or bad better than the 20 block.)

[@ClarkPotter]

Thank you for this. I was going to ask this.

[@RyanSmith-ow6cm]

The number of blocks refers to the depth of the neural network, which you can think of as it's ability to think abstractly. In the context of Go, more abstract representations give the 40 block version the ability to form a more refined "intuition" for the correct moves to analyze and it's judgement of variations as favorable or unfavorable. It's worth pointing out that AlphaGo Lee used different network structure(s) that function a little differently than AGZ's "blocks". AG Lee used 12 and 14 layer networks for it's value and policy networks if I remember. A lot of information specific to the board position was computed as features to the network so it's not an Apples to Apples comparison against Zero since it got nothing other than the stones, but 20 blocks is over 40 layers.

[@dougwan8150]

Thank you all for making the great videos that are so professional and thorough.

[@samuelzero3307]

4 years late lol… Nice commentary Mr.Redmond, it’s very helpful!

[@RoryMitchell00]

It's nice to hearken back to the Lee match that started this all, and I loved the commentary from Michael and Chris that referenced that important point in history. Overall, this game had fairly straightforward play for us lowly humans to try and understand, but still produced some interesting shapes. Seems like most of the game was AG Lee wishing/expecting the moves to go one way, then being surprised by Zero and having to scramble and change its plans, all leading to its eventual demise.

[@Maharani1991]

Awesome game and beautiful analysis, as always. Thank you, guys. :) When is Michael gonna play Leela Zero? ;)

[@kfm1242]

GerSHAK Chris, please make it happen :) .... With the help of some wine or something :)

[@SamuelGrguric]

yeah the beginning was a little confusing. i thought i was watching a rerun or something because the title and stuff. Also, i should point out that not all the viewers are AGA members (or even living in the US), so we never get welcomed at the start or thanked for watching at the end :( Edit: should also say thanks for the great video and analysis as always! I look forward to them every week :)

[@alekerickson4301]

would it be possible at some point to have a full lecture on the 3-3 invasion joseki based on michael's analysis of these AG v AG games?

[@sandybarnes887]

kzbin.info/www/bejne/aJzLf6F7arZmfa8

[@guestgreenstein4726]

It's funny how Chris had a 50/50 and was like: so AlphaLee thinks it's... winning?.. No, Chris.

[@sashasaffron1630]

Why is there not any footage on KZbin of AlphagoZero versus AlphagoZero in go? If I join the AGA and donate can I see that? Smh

[@sandybarnes887]

kzbin.info/www/bejne/rJvWf2x4h66hr9U

@@sandybarnes887 it says AG vs AG - not Zero vs Zero?

[@sandybarnes887]

@ AlphaGo Fan, AlphaGo Lee (Sedol), AlphaGo Master (60 online games vs humans), AlphaGo Zero (40 block), AlphaZero (20 block)

[@benjaminschooley3108]

Alphago Lee is the one that had the human games as a reference right?

[@dwightchase4013]

AlphaGo Lee is the one that used the 300K games as a reference you are right about that

[@chuang-yg1vc]

there are variations in lower right corner, and I wonder if Michael would comment on how each variation would play out locally, in the game in the middle, and the overall gain/loss.

[@thebigv5573]

After black plays P5 to let live white in the lower right corner *why* cant white stick to his initial plan to sacrifice the corner invading stones to squeeze black in the corner and attach on top of the P5 stone (at P6?) Black will still have to capture those white stones at some point to live. Can someone explain?

[@a_ghoul]

This is the new kind of tournetment

[@PasseScience]

Thumbnail still missing for this one :P

[@Deus_Auto]

Starts @ 3:02

[@ClarkPotter]

How strong is Chris Garlock?

[@adfaklsdjf]

Is he using CGoban to review the game?

[@64standardtrickyness]

is there anywhere I can find the kie jie vs alphago commentary that got cut out of the deepmind video?

[@Maharani1991]

+

[@sandybarnes887]

kzbin.info/www/bejne/kF6rfWikpdV1pK8

[@christophermartin5613]

I am sure that you guys have covered this before, but I have only recently been reading comments. What is the rank of Mr Garlock? Wouldn't it it be better if another pro accompanied Redmond? The problem is Redmond is so outspoken, it would be difficult to find someone who is a good counter to him. There are a number of points where I wished Garlock would have spoken up, but he lacks the strength to add anything meaningful.

[@ConsciousBreaks]

Chris Garlock is 3 dan.

[@nicokent9648]

23:26 lollll

[@oathofall]

Can anyone explain to me the difference between AlphaGo Zero and Master (as in the last commented game)?

[@ryuuguu01]

The two main differences are AG Zero is more advanced and did not see any human or other games when training. It just got the rules and played itself lots. AG Lee first saw a large number of human games before training against itself. Also AG Lee was an earlier design and is less advanced internally.

[@oathofall]

So Master is more related to Lee?

[@ryuuguu01]

None of the AlphaGos are trained to give up. They are only trained to win. They are not trained that giving up early is any different then losing by 1/4 point (Chinese counting).

[@benjaminschooley3108]

Zero also plays other games like chess and some competetive computer games, did it need any extra programming to "learn" to play these games?

[@ryuuguu01]

Those are really other programs built on the same architecture as the go playing program. So they also would start from a blank slate except for the rules and play against them themselves till they became superhuman. Neural Net programming is done in two parts. a) the network is designed, this includes including the game rules in the design. b) the network is trained by running it millions of times. In the Zero versions it trains by playing itself only. In the previous versions of Alphago (master ,lee etc.) initial training was done by looking at human games then playing against itself. So for part a) Chess Zero version would have very similar design with changes for things like board size and rules. Then for part b) Chess Zero would play Chess Zero millions of times, it would not play Go Zero.

[@undertyped1]

If every human on earth played 100 games against alphago zero, i wonder if humanity could win once.

[@thatchessguy7072]

Nah

[@bermudachan6292]

what is the uhm count today

[@dtracers]

Even though he does not have Alphago he does have Leela Zero and Elf Both of which are much stronger than 20 block Alpha zero. he should ask questions of those.

[@Maharani1991]

Source re: much stronger?

[@ConsciousBreaks]

GerSHAK There is no source because it's not true.

[@dtracers]

ConsciousBreaks the source is that alphago Lee loses to pros. While Elf gives handicaps to top pros.

[@Maharani1991]

dtracers What you said above was that these other bots were (much!) stronger than AG Zero 20 Blocks. I dont think anyone doubts that theyre stronger than AG Lee.

[@ConsciousBreaks]

Yes, I think dtracers probably meant AG Lee and not AG Zero (20 blocks).

[@bermudachan6292]

its gonna reduce a lot of territory and uhm and uhm

[@nomukun1138]

lol it's spelled um what a noob

[@bermudachan6292]

uhm uhm uhm uhm uhm uhm uhm

[@bermudachan6292]

uhm uhm uhm uhm uhm uhm uhm