-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
AlphaGo round 2
- Thread starter post-S
- Start date
chillybright
Banned
Ke Jie gg!
Monochromaticx
Member
> Well Played!
> Ez Mid
> Ez Mid
Yea, at least that I understand!
Lee Ha-Jin is a boss. Love her youtube channel.
Aruarian Reflection
Chauffeur de la gdlk
Game 2 is on now
https://www.youtube.com/watch?v=1U1p4Mwis60
https://www.youtube.com/watch?v=1U1p4Mwis60
Prototype Viktor
Member
Wasn't even close he got mollywhopped
I used to find this interesting, and as an engineer I can appreciate the technical complexity, but I don't feel the events are very compelling anymore.
Even last time it wasn't remotely close, and at this point it really feels like Go is a solved game in terms of human ability, and putting human players through this seems akin to some sort of psychological humiliation.
Even last time it wasn't remotely close, and at this point it really feels like Go is a solved game in terms of human ability, and putting human players through this seems akin to some sort of psychological humiliation.
Prototype Viktor
Member
AlphaGo is SKT if it was 5 Fakers
It sounds like the second game was close up until a certain point. Looks interesting to study.
Apparently Ke Jie tried a gamble and it didn't pan out? The blog post doesn't make it sound as bad as it sounds here.
https://deepmind.com/research/alphago/alphago-china/
Apparently Ke Jie tried a gamble and it didn't pan out? The blog post doesn't make it sound as bad as it sounds here.
https://deepmind.com/research/alphago/alphago-china/
http://www.npr.org/sections/thetwo-...against-humanitys-last-best-hope-to-win-at-go
I wonder if the five GMs working together can win. Ke Jie said he got excited in the mid-game and possibly made mistakes.
Aruarian Reflection
Chauffeur de la gdlk
I found it interesting that AlphaGo had assessed Ke Jie was playing a flawless game for the first 50 moves
Prototype Viktor
Member
This PairGo matchbox a little weird, white AlphaGo seems to be taking a lot of time
OpinionatedCyborg
Member
It would be cool if they could roll the game back to where ke jie made a mistkae and let him play it again. Basically, save scum alpha go and see if you could beat it.
Best part about the Team vs AlphaGO match was this: (Stolen from reddit)
After AlphaGO did its thing where it knows it's gonna win with 100% certainty anyway so it stops giving a fuck about gaining/keeping points. I know that it's silly to anthropomorphize AlphaGO into a person, but it just seems hilarious to me.
After AlphaGO did its thing where it knows it's gonna win with 100% certainty anyway so it stops giving a fuck about gaining/keeping points. I know that it's silly to anthropomorphize AlphaGO into a person, but it just seems hilarious to me.
Prototype Viktor
Member
The 3rd match he was doing seemingly experimental strategy in hopes of throwing the AI off
The human players sometimes seem like they're attempting to confuse the AI or "gamble".
The program is incredibly calculating. I don't really understand why gambling or confusion would work. It's not like it attempts to understand human reasoning, it just seeks optimal winning odds.
The program is incredibly calculating. I don't really understand why gambling or confusion would work. It's not like it attempts to understand human reasoning, it just seeks optimal winning odds.
Ke Jie beat Fine Art something like four times in a row by exploiting a flaw in its reasoning. Tencent fixed that issue but there's precedent that AI can be beaten by strange moves (not to mention the game Lee Sedol won against AG by confusing it).
It's become pretty clear that there's no beating AG in its current state. There's really no flaw in its calculations that we can exploit.
The program can't calculate everything though, just as we know it is clearly superior to human players. It's pretty basic game theory to attempt to add randomness to increase the variance of the potential outcomes when you know your opponent is superior.
Best chance of winning vs AlphaGo seems to be to make the entire game into one life or death fight and hope that the computer makes some fatal error early on in the the fight when the outcome is still impossible to read out. Since the computer has more 'read depth' than the human, the human has more chances to make such a mistake, but it is not impossible that the human can luck out a win in such a game. It's how Lee Sedol took a game off alpha go, creating a super complicated fight that would decide the whole game, and AlphaGo made an error early in the sequence that it didn't recognize until it was too late. If AlphaGo has improved it's depth of reading, or its ability to prevent such all or nothing fights, such a strategy will be come less and less likely to work.
Seems like a bad gamble.
They posted the first 10 (of 50 they're planning to release) samples of full strength self-play games that AlphaGo did during training, and they're nuts. Some of the pros are saying the moves almost look alien, incomprehensible.
https://deepmind.com/research/alphago/alphago-vs-alphago-self-play-games/
No shit, but playing the kind of spread out game of incremental small gains that alphago prefers is a guaranteed loss.
High variance chaotic games where you have a 1 in 10 chance of winning is better than no chance of winning. Depending on how far the program has advanced since the lee sedol matches, the chances may not even be that high anymore, but it's still the best chance a human has.