|
empty whippet box posted:Does it win by the smallest margin possible on purpose? Like, did he actually come close to beating it or is did it just play ultra conservatively to maintain the one point lead that he never really had a chance to get ahead of? It plays for maximum chance to win without regard to margin of victory. At times it defies the notion of "always play the biggest move" and changes it to "always maximize your chance to win". These may sound synonymous, but AlphaGo is teaching us otherwise.
|
# ? May 24, 2017 05:13 |
|
|
# ? May 13, 2024 00:10 |
|
Overwined posted:It plays for maximum chance to win without regard to margin of victory. At times it defies the notion of "always play the biggest move" and changes it to "always maximize your chance to win". These may sound synonymous, but AlphaGo is teaching us otherwise. The situation doesn't come up too much between players of equal skill, but human players do the same tactic of playing safe moves that reduce their opponent's potential for counterplay (even if doing so gives up a few immediate points) when they're far ahead enough that the only way they lose is if their opponent manages to blow the position wide open. I guess the big difference is that AlphaGo has a much better intuition for how the points are going to fall, so it makes those sort of moves even when human players would consider it too close to call and too risky to give up those points.
|
# ? May 24, 2017 12:15 |
|
Jabor is right, but there's a nuance in the late endgame too. Since AlphaGo is coldly AI/probability-of-winning based and doesn't have a special algorithm for late endgames. Therefore, when even the weakest dan level amateur would go "okay, this is simple enough that all my future moves will be perfect", AlphaGo is thinking "well, [the obvious biggest move that any pro would make] feels like a 99.99% chance of winning, but [move that gives away some points] feels like a 99.9999736% chance of winning, so I'll do that instead". I suspect AlphaGo has uncertainty when humans and other programs designed to play endgames perfectly would play bigger moves.Overwined posted:So far only that one wedge move my Lee Sedol has managed to "fluster" it. dirby fucked around with this message at 13:40 on May 24, 2017 |
# ? May 24, 2017 13:37 |
|
dirby posted:Remember that the version Lee Sedol flustered was not nearly as strong as the "Master" version, and almost certainly not as strong as this new version. Which is the scary part. I'd actually like to see Ke Jie play more desperate moves so that we can see how AlphaGo responds to fringe moves. Who know, we may discover new strategies and isn't that kind of the whole point?
|
# ? May 24, 2017 15:54 |
|
Overwined posted:Which is the scary part. I'd actually like to see Ke Jie play more desperate moves so that we can see how AlphaGo responds to fringe moves. Who know, we may discover new strategies and isn't that kind of the whole point? Well who wants to see it beat fringe (i.e. disfavoured, because they're likely bad) moves? I'd like to see it turn standard theory on its head like it has been. Arguably it did that already too with the corner plays in game one anyway. Deepmind will put out another paper after the event is over detailing what's new in this version. Apparently it's running inference on a single neural-net ASIC rather than 1200 or so gpus like last year. Also the master version had 3 stones on the Lee Sedol version, which had 3 on the Fan Hui version. I wonder how many more this version has. Presumably they are getting diminishing returns but who knows.
|
# ? May 24, 2017 23:18 |
|
They are getting what looks like diminishing returns because they are running out of ruler. How can you measure a thing using only itself as a measure? You can't use margin of victory because AlphaGo gives no fucks about the margin.
|
# ? May 24, 2017 23:58 |
|
Overwined posted:They are getting what looks like diminishing returns because they are running out of ruler. How can you measure a thing using only itself as a measure? You can't use margin of victory because AlphaGo gives no fucks about the margin. Actually that's not really the hard part. Each week you have the latest version play last week's version a thousand times and measure the win rate. 90% win rate is ~3 stones. Where this breaks down is if both copies have some sort of blind spot, that one can't exploit in the other since they both have it. Hence, the external games.
|
# ? May 25, 2017 00:21 |
|
Fuzzy Mammal posted:Actually that's not really the hard part. Each week you have the latest version play last week's version a thousand times and measure the win rate. 90% win rate is ~3 stones. Where this breaks down is if both copies have some sort of blind spot, that one can't exploit in the other since they both have it. Hence, the external games. Very interesting. I hadn't thought about it that way. Nonetheless, external games are now officially useless since AlphaGo is so much stronger than the top human. I wonder where they'll push DeepMind next. I think it's always been the plan for Alphabet to abandon the AlphaGo project in order to push DeepMind into broader horizons. I wonder if they will open source AlphaGo or at least hand it off to a reputable Go organization.
|
# ? May 25, 2017 00:36 |
|
c...can you add me to ITGO?? username: Bl1tz
|
# ? May 25, 2017 00:41 |
|
Fuzzy Mammal posted:Also the master version had 3 stones on the Lee Sedol version
|
# ? May 25, 2017 01:37 |
|
Apparently it was in a talk yesterday. I can't find the original source, which I think is also in chinese. Hopefully we learn a lot from the paper. I work at google and even I can't find much out about what they've been up to. https://www.reddit.com/r/baduk/comments/6cza2t/david_silver_reveals_new_details_of_alphago/ e: aha: https://www.facebook.com/GOking2007/videos/1364474096921048/ Fuzzy Mammal fucked around with this message at 01:50 on May 25, 2017 |
# ? May 25, 2017 01:48 |
|
Reminder: We have a Discord server.
Xom fucked around with this message at 23:45 on Jul 15, 2018 |
# ? May 25, 2017 02:13 |
|
I'd also enjoy an ITGO invite. Username is ayc.
|
# ? May 25, 2017 14:59 |
|
Invited.
|
# ? May 25, 2017 17:28 |
|
AlphaGo currently can't be open-sourced, or at least not in a way that would let you run it on commodity hardware. It runs on specialized, custom-manafactured computer chips designed to be thousands of times better at this kind of problem than general-purpose chips..
|
# ? May 25, 2017 21:03 |
|
There are other bots out there that do play at a high level (Wikipedia claims 1-5d for Facebook's go bot, depending on hardware) and are open-source, though. e: i'm told that those are amateur dans extrapolated from online go performance, so maybe not as impressive as it sounds? Reiterpallasch fucked around with this message at 21:09 on May 25, 2017 |
# ? May 25, 2017 21:07 |
|
Reiterpallasch posted:AlphaGo currently can't be open-sourced, or at least not in a way that would let you run it on commodity hardware. It runs on specialized, custom-manafactured computer chips designed to be thousands of times better at this kind of problem than general-purpose chips.. It's not as hopeless as all that. Google has open-sourced TensorFlow, their main machine-learning library. TensorFlow is intended to run on arbitrary hardware, though it's likely to run much better on specialized devices. Plus, Google lets you rent out specialized machine learning hardware on their cloud platform, so even if it does only work well with that, normal people could still run it. The two main factors preventing AlphaGo from getting open-sourced are how much it uses TensorFlow and whether Google's willing to put in the time to open it up. TensorFlow is new enough that AlphaGo might use different code in some places, which would make it hard to open-source. And open-sourcing a project that wasn't intended to be open source is a big undertaking; Google has all sorts of crufty internal libraries that they might not want to release to the public, and unless AlphaGo was written from the ground up with the intention of being open sourced, it probably uses them. But open-sourcing AlphaGo would be great publicity for TensorFlow, and Google's pushing the library hard, so they might still do it. Disclaimer: I'm an ex-Google employee, but don't know anything special about machine learning or AlphaGo.
|
# ? May 25, 2017 22:23 |
|
Xom posted:Invited. Thank you! Now to find time to actually jump on...
|
# ? May 26, 2017 00:01 |
|
Reiterpallasch posted:There are other bots out there that do play at a high level (Wikipedia claims 1-5d for Facebook's go bot, depending on hardware) and are open-source, though. facebook has a go bot?
|
# ? May 26, 2017 07:14 |
|
empty whippet box posted:facebook has a go bot? Yes, and when AlphaGo beat Lee Sedol they gave them a very snarky "compliment". If I remember correctly they implied that AlphaGo is only a huge library, which is patently false.
|
# ? May 26, 2017 14:51 |
|
A snarky false statement on Facebook? Get out of town! Is there a goon OGS ladder?
|
# ? May 26, 2017 15:41 |
|
There is now 10 new alphago vs alphago games published, and they are completely bizarre and amazing. https://deepmind.com/research/alphago/alphago-vs-alphago-self-play-games/
|
# ? May 27, 2017 10:50 |
Wow. These are surreal.
|
|
# ? May 27, 2017 17:28 |
|
Yeah, looking at those things, it's pretty obvious that humans are quite a bit behind alphago now. So many weird moves.
|
# ? May 27, 2017 22:11 |
|
By the way, one of the AlphaGo engineers was asked about potential blindspots when AlphaGo plays itself. He said the exact opposite is true and that AlphaGo, as you can see, plays non-traditional moves far more frequently against itself than against human opponents. This isn't to say that it's "experimenting" but more that humans are the ones with the blindspots playing along largely orthodox lines. Only AlphaGo can come up with moves that will challenge AlphaGo.
|
# ? May 28, 2017 00:14 |
I really wish the pros played alphago at 2h or 3h just because I'm curious.
|
|
# ? May 28, 2017 06:44 |
|
Looking at game 9 that they just released, move 52-56 - why? this seems like a genuine boneheaded mistake on black alphago's part. It just made me and I can't find a way to make it not seem like a bad series of moves.
|
# ? May 29, 2017 04:45 |
|
They added 40 more games to: https://deepmind.com/research/alphago/alphago-vs-alphago-self-play-games/ TL;DR: It is always good to invade at 3x3.
|
# ? May 29, 2017 08:32 |
|
I just recovered a hard drive that failed in 2010. I found this old gem and thought people might want to hear it.
|
# ? May 30, 2017 02:03 |
|
https://www.reverbnation.com/djsente/songs
|
# ? May 30, 2017 04:48 |
|
simply amazing DJ sente...the man...the legend...the loving Dan.
|
# ? May 30, 2017 06:57 |
|
My wifes magazines are go-server-shaming me:
|
# ? May 30, 2017 08:39 |
|
Anyone know of any places to buy go apparel? Like, for instance, a shirt that the whole pattern was a go board with a game being played on it. I know how incredibly, ridiculously dorky this sounds. I want it. Hell, I'd wear a hoody like this. Pants and shoes like this. *definitely* socks.
|
# ? Jun 18, 2017 07:05 |
|
I have read your OP and some posts here and there. I think I would like to try my hand at learning and playing Go. I downloaded an app, "Tsumego Pro" to practice problems on my phone. But my problem with the app is the reasons for a move being right or wrong are not always readily apparent. Are there any good phone apps for playing against AI or humans or problem sets? I play to figure out the KGS client later and do that route but some mobile way to play/learn would help me find the time for this I think.
|
# ? Jun 21, 2017 02:45 |
|
The whole point of Tsumego is that you have to work out why the right move is the right move. They're problems, not examples to teach by. The best advice remains to play as much go as you can. Start earning those 100 losses.
|
# ? Jun 21, 2017 06:59 |
|
tangy yet delightful posted:I have read your OP and some posts here and there. I think I would like to try my hand at learning and playing Go. I downloaded an app, "Tsumego Pro" to practice problems on my phone. But my problem with the app is the reasons for a move being right or wrong are not always readily apparent. Are there any good phone apps for playing against AI or humans or problem sets? Most "problem sets" will be way, way too tough for you at first. They are designed on the assumption that you generally know what you're doing and some of the techniques for killing or saving stones. If you want to do problems you should pick up the Elementary Go series books "Tesuji" and "Life and Death" which actually introduce the general concepts being used before testing you on them. They are also going to get tricky for a beginner very fast but you can work through them very slowly in between playing games. e: You can read these books through the "Go Books" app if you have an iphone. secret volcano lair fucked around with this message at 22:24 on Jun 21, 2017 |
# ? Jun 21, 2017 22:17 |
|
Reached 9k on IGS! Finally, single digits! Soon, the world!
|
# ? Jun 22, 2017 00:06 |
|
Out of curiosity, how often do you high sdk and dan players review your own games?
|
# ? Jun 24, 2017 22:39 |
|
Overwined posted:Out of curiosity, how often do you high sdk and dan players review your own games? I'm about 4k, and maybe once every year or so.
|
# ? Jun 26, 2017 08:48 |
|
|
# ? May 13, 2024 00:10 |
|
empty whippet box posted:Anyone know of any places to buy go apparel? Like, for instance, a shirt that the whole pattern was a go board with a game being played on it. I know how incredibly, ridiculously dorky this sounds. I want it. Hell, I'd wear a hoody like this. Pants and shoes like this. *definitely* socks. Cafepress and search for baduk. They are pretty much all garbage designs though, youd probably have better luck designing your own
|
# ? Jun 26, 2017 10:03 |