Google DeepMind's AlphaGo Artificial Intelligence Takes Game One Against Top Player

[cageymaru]

The current number one player in the world was narrowly defeated by the Google DeepMind artificial intelligence in the first of three scheduled matches based in China. The game went to a count and the DeepMind AI won by the closest possible margin of half a point. Demis Hassabis, founder and CEO of DeepMind, commented on Twitter that "Ke Jie is using the ideas #AlphaGo used in the master series of online games in January against AlphaGo intriguing to see what it will do." It will be interesting to see what approach Ke Jie uses against the AI for the second match and if it will be more successful.

"Ke Jie fought bravely and some wonderful moves were played," wrote Google DeepMind CEO Demis Hassabis on Twitter. He added: "[The] Game went to a count. #AlphaGo won by just half a point, the closest margin possible." Eric Schmidt, the former CEO of Google and the executive chairman of Alphabet, Google's parent company, has flown in for the event, as has David Silver, the lead programmer on AlphaGo, and Mustafa Suleyman, another one of the DeepMind founders.

 

[the-one1]

Did Google make the loser eat a bag of Go tiles?

 

[wootius]

The thing with AlphaGo is it isn't designed to win by highest score, but by taking the path with the largest margin of victory.

Earlier in the match it was up by like 14 points and an hour on the clock.

 

[Uvaman2]

Honestly at this point in my life I find this type of shit a pathetic showing of computer 'power'.
Its the human that is doing things that are against brain nature.

 

[sfsuphysics]

Honestly at this point in my life I find this type of shit a pathetic showing of computer 'power'.
Its the human that is doing things that are against brain nature.

Yeah, makes me wonder how much of an "AI" thing is going on versus brute force algorithms that can analyze every possible move like 100 moves deep.

 

[travisty]

Yeah, makes me wonder how much of an "AI" thing is going on versus brute force algorithms that can analyze every possible move like 100 moves deep.

LOL that works with checkers and chess. That cannot be done with Go.

Tic-Tac-Toe
Possible Moves: 10^4

Checkers
Possible Moves: 10^20

Chess
Possible Moves: 10^45 (it goes higher for entire games but with some intelligence behind the brute force this is more in line)

GO
Possible Moves: 10^172

To put that in better perspective, the estimated number of atoms in the universe is 10^80

 

[Uvaman2]

LOL that works with checkers and chess. That cannot be done with Go.

Tic-Tac-Toe
Possible Moves: 10^4

Checkers
Possible Moves: 10^20

Chess
Possible Moves: 10^45 (it goes higher for entire games but with some intelligence behind the brute force this is more in line)

GO
Possible Moves: 10^172

To put that in better perspective, the estimated number of atoms in the universe is 10^80

Then you upload as many games as possible and as many strategy as possible, again brute force, some outcome analysis and rough comparison analysis. Still not impressed for some reason.

 

[travisty]

Then you upload as many games as possible and as many strategy as possible, again brute force, some outcome analysis and rough comparison analysis. Still not impressed for some reason.

Wha??? Not even sure what you're saying. Brute force can be done in simpler games where tree-traversals can be performed. With Go that just isnt possible. Watch the video if you want a better understanding

 

[TheBuzzer]

is there is a finite amount of moves it should be possible for a computer to calculate.

 

[serpretetsky]

is there is a finite amount of moves it should be possible for a computer to calculate.

If there are a finite number of combinations it is theoretically possible to go through all of them with a computer. It is not practically possible given information storage limitations and processing time limitations for problems with a large amount of combinations.

 

[travisty]

is there is a finite amount of moves it should be possible for a computer to calculate.

As serpretetsky said, theoretically it's possible but are you willing to wait hundreds of years to get an answer? (though technically we, the entire human race, don't have enough storage capacity with all computers combined to hold 10^172 moves worth of data (each move probably taking somewhere in the range of at least a hundred bytes to store the state of the board - thinking 2 pointers for the next nodes and bit-wise operations for the chips on the board))

It's equivalent to building Deep Thought and asking it "What is the answer to life, the universe, and everything?" and then waiting seven and a half million years to get the answer
https://en.wikipedia.org/wiki/List_of_minor_The_Hitchhiker's_Guide_to_the_Galaxy_characters#Deep_Thought

 

[sfsuphysics]

LOL that works with checkers and chess. That cannot be done with Go.

Tic-Tac-Toe
Possible Moves: 10^4

Checkers
Possible Moves: 10^20

Chess
Possible Moves: 10^45 (it goes higher for entire games but with some intelligence behind the brute force this is more in line)

GO
Possible Moves: 10^172

To put that in better perspective, the estimated number of atoms in the universe is 10^80

Never said that it was to be done with every single variant possible for an entire game, I said 100 moves deep. Now I don't know enough about the game to know the actual statistics for how many moves that could entail, but if it's astronomical large, then fine go 20 moves deep, or even 10. Numbers can't be anywhere close to 10^172 then. Ditto with chess, you don't need to analyze every possible outcome from the beginning, just say 20 moves, and then recalculate based upon what pieces moved, and which are left, etc.

Bottom line, is take the total number of moves a human is able to think through in a given turn, and multiply that number by 10 and a computer can do it, then think about how many more can it do? 20? 100? and that can give the computer a distinct advantage

 

[serpretetsky]

Never said that it was to be done with every single variant possible for an entire game, I said 100 moves deep. Now I don't know enough about the game to know the actual statistics for how many moves that could entail, but if it's astronomical large, then fine go 20 moves deep, or even 10. Numbers can't be anywhere close to 10^172 then. Ditto with chess, you don't need to analyze every possible outcome from the beginning, just say 20 moves, and then recalculate based upon what pieces moved, and which are left, etc.

Bottom line, is take the total number of moves a human is able to think through in a given turn, and multiply that number by 10 and a computer can do it, then think about how many more can it do? 20? 100? and that can give the computer a distinct advantage

yes, what you are describing is the basis of most engines, chess or GO. The hard part is evaluating the position, deciding if a certain position does not merrit any more thinking (pruning), and choosing the best position. Humans cant check all of the combinations very quickly, but they can be amazing at evaluating positions, deciding which moves are even worth thinking about, and deciding if an overall long-term strategy is worth going for or not. Historically the evaluation of a position was preprogrammed in by what game masters told the programmers to program in (losing a rook is bad, losing a queen is worse, having a knight in this area of the board is good, etc etc).
https://www.chess.com/blog/zaifrun/creating-a-chess-engine-from-scratch-part-1
Programmers had trouble efficiently implementing this for GO. GO also has ALOT more possible moves for each step then chess. They had to use machine learning instead and let the computer learn itself how to play.
https://motherboard.vice.com/en_us/...ied-so-alphago-could-live-giraffe-matthew-lai

 

[Uvaman2]

yes, what you are describing is the basis of most engines, chess or GO. The hard part is evaluating the position, deciding if a certain position does not merrit any more thinking (pruning), and choosing the best position. Humans cant check all of the combinations very quickly, but they can be amazing at evaluating positions, deciding which moves are even worth thinking about, and deciding if an overall long-term strategy is worth going for or not. Historically the evaluation of a position was preprogrammed in by what game masters told the programmers to program in (losing a rook is bad, losing a queen is worse, having a knight in this area of the board is good, etc etc).
https://www.chess.com/blog/zaifrun/creating-a-chess-engine-from-scratch-part-1
Programmers had trouble efficiently implementing this for GO. GO also has ALOT more possible moves for each step then chess. They had to use machine learning instead and let the computer learn itself how to play.
https://motherboard.vice.com/en_us/...ied-so-alphago-could-live-giraffe-matthew-lai

In the end its models, statistics and probability, now adaptive!.
Machine 'learning' and adaptive algorithms, I am sure are 'impressive', just not as impressive to me anymore.. I am getting old I guess.. computers don't seem so game-changing to me.. more like game-fucking.

 

[Nytegard]

Well, AlphaGo has officially won the second game, and defeated the World's #1 Go player. Human's are now officially inferior at all games. I wonder if we can summon Death to play against it, as I heard he never loses.