It is currently Thu Aug 13, 2020 4:50 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 39 posts ]  Go to page 1, 2  Next
Author Message
Offline
 Post subject: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #1 Posted: Mon Dec 11, 2017 5:51 am 
Lives in sente

Posts: 727
Liked others: 44
Was liked: 218
GD Posts: 10
Is this the tool you expect from DeepMind?

https://alphagoteach.deepmind.com

From Aja Huang's facebook

大家好,很高興向大家宣布,AlphaGo教學工具上線了。
這是一個AlphaGo教學的開局庫,相關細節如下:
1. 本教學工具總共收錄了約6000個近代圍棋史上主要的開局變化,從23萬個人類棋譜中收集而來。
2. 所有盤面都有AlphaGo評估的勝率,以及AlphaGo推薦的下法。
3. 所有AlphaGo的勝率與推薦下法,AlphaGo都思考將近10分鐘(1000萬次模擬)。
4. 每一個開局變化,AlphaGo都固定延伸20步棋。加上AlphaGo的下法,整個教學工具約有2萬個分支變化,37萬個盤面。
受限於投稿《自然》雜誌的時程,本教學工具使用的版本是AlphaGo Master。希望大家享受AlphaGo教學工具中的創新下法,也能從中有所收穫

Translation provided by Reddit user AngelLeliel
Hello everyone, I am happy to announce that AlphaGo teaching tool is online. This is a opening library of AlphaGo teaching, details are below: The teaching tool has collected about 6,000 major opening variations in the history of modern Go, collected from 230,000 human plays. All boards have the AlphaGo estimated winning rates, as well as AlphaGo's recommended moves. All estimated winning rates and recommend moves are evaluated by AlphaGo for about 10 minutes (10 million simulations). AlphaGo also plays 20 more moves on every opening variation. With AlphaGo's moves, the entire teaching tools have about 20,000 branches, 370,000 game state. Due to the timeline for submission of the Nature, AlphaGo Master is the version used in this educational tool. Hopefully, everyone will enjoy these innovative moves from AlphaGo, and also benefit from it.


Last edited by pookpooi on Mon Dec 11, 2017 10:58 am, edited 3 times in total.

This post by pookpooi was liked by: Bonobo
Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion
Post #2 Posted: Mon Dec 11, 2017 6:01 am 
Lives with ko

Posts: 271
Location: Hong Kong
Liked others: 2
Was liked: 49
Rank: IGS 7d
IGS: goeye
pookpooi wrote:
Is this the tool you expect from DeepMind?

https://alphagoteach.deepmind.com


A little disappointed.


This post by GoEye2012 was liked by: Charlie
Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion
Post #3 Posted: Mon Dec 11, 2017 6:10 am 
Gosei

Posts: 1450
Liked others: 412
Was liked: 412
Rank: 5d
GD Posts: 1000
Early in my go career I played through many pro games. I began to imitate pro moves in my own games but being 15 ranks weaker, of course I didn't understand many of the moves I was imitating. My imporession of AlphaGoTeach is that it doesn't teach in the sense of explaining the moves it recommends so it seems to put us in the position I was in may years ago, imitating things I didn't understand. It isn't clear whether this will even work for weaker players more than a few moves into the game. Would AlphaGoTeach work with opening moves of double digit kyu players? Presumably these games would include moves AlphaGo would not have seen in its database. Also, what is the meaning of the probability estimates in games of weak players. Does it mean the probability of a weak black player winning or the probability of AlphaGo winning from that point on?

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #4 Posted: Mon Dec 11, 2017 6:34 am 
Gosei

Posts: 1564
Location: Earth
Liked others: 588
Was liked: 261
I for one, love it.

It is working fine here with firefox, some problems with google chrome :lol:

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #5 Posted: Mon Dec 11, 2017 6:49 am 
Lives in sente

Posts: 839
Location: Littleton, CO
Liked others: 286
Was liked: 264
Rank: KGS 3k
Universal go server handle: jeromie
I believe Deepmind made an announcement that Ke Jie was going to work with them when they were developing this tool. (I’ll have to search and see if I can find that quote.) I wonder if he has had early access to it, and if that has contributed to his changing style?

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #6 Posted: Mon Dec 11, 2017 7:00 am 
Lives in gote

Posts: 311
Liked others: 0
Was liked: 44
Rank: 2d
Gomoto wrote:
It is working fine here with firefox, some problems with google chrome :lol:
I cannot seem to explore variations in Chrome, even though I see they are there when I load the book into another program...

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #7 Posted: Mon Dec 11, 2017 7:01 am 
Oza
User avatar

Posts: 2204
Location: Germany
Liked others: 8130
Was liked: 908
Rank: OGS DDK again
Universal go server handle: trohde
Thanks, pookpooi — I’d have wished for some sort of Nuremberg Funnel, of course :D
________

Gomoto wrote:
some problems with google chrome
Works fine here with Google Chrome version 63.0.3239.84 on macOS Sierra 10.12.6 … takes a while to load, though.

__________________________________________________________________

As for me and my dwindling learning abilities …
Image

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS? :)


This post by Bonobo was liked by: Baywa
Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #8 Posted: Mon Dec 11, 2017 7:01 am 
Lives in sente

Posts: 727
Liked others: 44
Was liked: 218
GD Posts: 10
jeromie wrote:
I believe Deepmind made an announcement that Ke Joe was going to work with them when they were developing this tool. (I’ll have to search and see if I can find that quote.) I wonder if he has had early access to it, and if that has contributed to his changing style?


We’re also working on a teaching tool - one of the top requests we’ve received throughout this week. The tool will show AlphaGo’s analysis of Go positions, providing an insight into how the program thinks, and hopefully giving all players and fans the opportunity to see the game through the lens of AlphaGo. We’re particularly honoured that our first collaborator in this effort will be the great Ke Jie, who has agreed to work with us on a study of his match with AlphaGo. We’re excited to hear his insights into these amazing games, and to have the chance to share some of AlphaGo’s own analysis too.

https://deepmind.com/blog/alphagos-next-move/

Maybe his recent extreme AlphaGo style is because he's absorbing so much information from this collaboration?

Top
 Profile  
 
Offline
 Post subject:
Post #9 Posted: Mon Dec 11, 2017 7:48 am 
Honinbo
User avatar

Posts: 8803
Location: Santa Barbara, CA
Liked others: 339
Was liked: 2059
GD Posts: 312
Quote:
1. The teaching tool has collected a total of about 6,000 major changes in the history of modern Go, collected from 230,000 human chess records.
1. This teaching tool has compiled about 6,000 mainstream opening variations in modern Go, gleaned from 230,000 human games.
Quote:
as well as AlphaGo's recommended downside.
...as well as AlphaGo's own continuations.
Quote:
All AlphaGo wins and recommends the next law, AlphaGo are thinking about 10 minutes (10 million simulations).
All of AlphaGo's win rates and variations are based on a thinking time of about 10 minutes per move (10 million simulations).
Quote:
Every opening change, AlphaGo are fixed extended 20 chess.
For each opening variation, AlphaGo provides a fixed continuation of 20 moves.
Quote:
Coupled with the AlphaGo method, the entire teaching tools about 20,000 branch changes, 370,000 disk.
Together with AlphaGo's moves, the teaching tool includes a total of about 20,000 variations and 370,000 full-board positions.


This post by EdLee was liked by 2 people: Bonobo, fireproof
Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #10 Posted: Mon Dec 11, 2017 8:02 am 
Judan

Posts: 6540
Location: Cambridge, UK
Liked others: 398
Was liked: 3609
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
moha wrote:
Gomoto wrote:
It is working fine here with firefox, some problems with google chrome :lol:
I cannot seem to explore variations in Chrome, even though I see they are there when I load the book into another program...

Yup, I've told a friend who works at DeepMind. Fails in Chrome on Win 7 for me, but works in Firefox. Funny how building a superhuman go bot is easier than making a cross-browser web UI :lol:

P.S. I actually have a bit of a concern in that rather encouraging creativity, this could make Go more like chess in terms of having a rigid opening book (this giant variation file, 7 MB sgf), will people want to keep playing openings that AlphaGo said was bad?

Top
 Profile  
 
Offline
 Post subject:
Post #11 Posted: Mon Dec 11, 2017 8:20 am 
Honinbo
User avatar

Posts: 8803
Location: Santa Barbara, CA
Liked others: 339
Was liked: 2059
GD Posts: 312
Quote:
will people want to keep playing openings that AlphaGo said was bad?
Given an AG-evaluated "bad" opening (say, sanrensei), if it requires 3 stones beyond top human level to exploit this 'mistake', then it makes little difference against a human opponent, so people are happy to continue to play it (against humans).

OTOH, if AG exposes an established human opening as bad, and humans subsequently figure out how to exploit it, then some people may refrain from it.

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #12 Posted: Mon Dec 11, 2017 8:55 am 
Lives in sente

Posts: 839
Location: Littleton, CO
Liked others: 286
Was liked: 264
Rank: KGS 3k
Universal go server handle: jeromie
Uberdude wrote:
I actually have a bit of a concern in that rather encouraging creativity, this could make Go more like chess in terms of having a rigid opening book (this giant variation file, 7 MB sgf), will people want to keep playing openings that AlphaGo said was bad?


I don't think that will be a problem (yet), for a few reasons:
  • From my (very brief, obviously) time playing with this tool, it looks like the common human moves are only about one percentage point worse than AlphaGo's top choice. For all but the very top players, that's a negligible difference.
  • This tool is based on AlphaGo Master, and we've already seen that AlphaGo Zero makes different choices in the opening. That leaves a degree of uncertainty in the evaluation that I believe will keep this from being rigidly followed.
  • This tool only takes us through the fuseki, so you have to be able to follow up from the final position in an AlphaGo like manner for it to be useful.
  • Because the tool is based entirely on whole board positions (which makes sense based on how AlphaGo evaluates things), if one player deviates from the moves AlphaGo shows the tool can't show how to refute it.

That may be more of a concern when we have general access to a computer program that can play at AlphaGo's level, though. At the very least, people will be able to hunt for refutations to common joseki / fuseki, and it's likely that some will fall out of favor. But that's been true of every era of Go.

Top
 Profile  
 
Online
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #13 Posted: Mon Dec 11, 2017 9:26 am 
Honinbo

Posts: 10107
Liked others: 3384
Was liked: 3273
Uberdude wrote:
Funny how building a superhuman go bot is easier than making a cross-browser web UI :lol:


Just wait for AlphaChrome Zero. :mrgreen:

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

My two main guides in life:
My mother and my wife. :)

Everything with love. Stay safe.


This post by Bill Spight was liked by 3 people: ez4u, pnprog, Waylon
Top
 Profile  
 
Online
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #14 Posted: Mon Dec 11, 2017 9:29 am 
Honinbo

Posts: 10107
Liked others: 3384
Was liked: 3273
jeromie wrote:
From my (very brief, obviously) time playing with this tool, it looks like the common human moves are only about one percentage point worse than AlphaGo's top choice. For all but the very top players, that's a negligible difference.


From what I have seen so far, which, admittedly, isn't much, 4% pts. is within the margin of error.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

My two main guides in life:
My mother and my wife. :)

Everything with love. Stay safe.

Top
 Profile  
 
Online
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #15 Posted: Mon Dec 11, 2017 10:00 am 
Honinbo

Posts: 10107
Liked others: 3384
Was liked: 3273
From the web site:

Quote:
each move’s winning probability was computed by running an independent search of 10 million simulations from that position.


Simulations of what, pray tell?

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

My two main guides in life:
My mother and my wife. :)

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #16 Posted: Mon Dec 11, 2017 10:32 am 
Lives in gote

Posts: 545
Liked others: 90
Was liked: 594
Rank: maybe 2d
I actually think that for mid to high dan level amateurs, there is some usefulness here. I'm below the necessary strength to derive a lot of value out of these variations, but I browsed through and still learned some things about openings that I actually encounter from both sides - the chinese, the mini chinese, the kobayashi, san-ren-sei. Moves that I wouldn't have even considered that I can experiment with now that I know that they're possible, or some moves that I would have considered and rejected as bad instead being evaluated as not-bad by AlphaGo with a simple and clear variation that disproves my misconception.

I also think that at the extremes the evaluations are useful at dan level. Mostly when they get noticeably outside the 40%-60% range I find that I'm more likely than not to myself feel dissatisfied with the side that supposedly fell behind - and that gives me some confidence that the evals, once they get that extreme, are at the right level to be at least somewhat informative. I wouldn't put huge value on it at amateur dan, but at minimum you can still use it as another good hundred datapoints for improving one's overall sense of direction of play and whole-board judgment, one more drop in the bucket.


This post by lightvector was liked by: johnsmith
Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #17 Posted: Mon Dec 11, 2017 12:33 pm 
Lives in gote

Posts: 311
Liked others: 0
Was liked: 44
Rank: 2d
Bill Spight wrote:
From the web site:
Quote:
each move’s winning probability was computed by running an independent search of 10 million simulations from that position.
Simulations of what, pray tell?
I think they simply mean the MCTS effort / iterations / node expands. EDIT: Actually this was Master still so rollouts as well.

Top
 Profile  
 
Online
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #18 Posted: Mon Dec 11, 2017 12:46 pm 
Honinbo

Posts: 10107
Liked others: 3384
Was liked: 3273
moha wrote:
Bill Spight wrote:
From the web site:
Quote:
each move’s winning probability was computed by running an independent search of 10 million simulations from that position.
Simulations of what, pray tell?
I think they simply mean the MCTS effort / iterations / node expands. EDIT: Actually this was Master still so rollouts as well.


If so, the win rates should be perfect for SDKs, eh?

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

My two main guides in life:
My mother and my wife. :)

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #19 Posted: Mon Dec 11, 2017 12:53 pm 
Oza

Posts: 2618
Liked others: 16
Was liked: 3752
The wording of the announcement is unclear as regards selection of "major" variations or indeed what is meant by "opening" (fuseki or joseki - I infer the former).

But insofar as I can make sense of it, the results don't seem to equate with pro tournament practice and the human moves shown must therefore be from a mix of pro and ama games but mostly ama.

If I follow the first two recommended moves (R16 then Q4) DeepMind gives 11 human moves: C17, C16, C15, C4, D17, D16, D4, D3, E17, O3 and P17.

But doing a similar exercise on a pro tournament database (approx. 90,000 games) gives a quite different result with 21 moves tried by humans, Human pros did not try the O3 given above, but did try the rest plus C5, C3, D15, D14, D5, E16, E4, E3, O17, P16 and R6.

The AG win rates are potentially interesting (but how much different from pro win rates as shown by Kombilo?), but if they include many wins against amateurs (and/or internet blitz games) they may be somewhat spurious. It would be nice if Nature could be cajoled into including a properly versed go player as one of their referees.

Just offering users the chance to find plays they have never considered before and thus be "creative" (sounds like an AI definition of the word) is not much of a selling point given the wealth of pro games around, and specifically those that were experimental such as New Fuseki, that already do that.

Top
 Profile  
 
Offline
 Post subject: Re: AlphaGo Teach discussion (Go Tool from DeepMind)
Post #20 Posted: Mon Dec 11, 2017 1:24 pm 
Lives in gote

Posts: 311
Liked others: 0
Was liked: 44
Rank: 2d
Bill Spight wrote:
moha wrote:
I think they simply mean the MCTS effort / iterations / node expands. EDIT: Actually this was Master still so rollouts as well.
If so, the win rates should be perfect for SDKs, eh?
Master itself did OK, looking only at these numbers, didn't it? :)

John Fairbairn wrote:
The AG win rates are potentially interesting (but how much different from pro win rates as shown by Kombilo?), but if they include many wins against amateurs (and/or internet blitz games)
Again, I think these are only probabilites, estimated by AG with a deep search (and rollouts with the reduced policy net). So there should be some correlations to actual pro data, but significant differences as well.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 39 posts ]  Go to page 1, 2  Next

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: Bill Spight and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group