AlphaGo- Whatever You Do Is Wrong

Gomoto · **#21**

I think the type of games should not matter after a few generations of the network. They games are just a source for the (random) patterns. Even in a KGS game there is a good local move from time to time ;-)

. The network "learns" how to apply the available patterns optimal. If the number of available pro games would be big enough to use them exclusivly for training the network would just learn a little bit faster.

sybob · **#22**

mumps wrote:

Actually not, as far as we're aware. Prior to the original Nature paper they only trained it using KGS games.

Although they may have added pro games I doubt whether there are even one million of them available - GoGod doesn't even have 100,000.

Even if they have a large number it's a good question as to whether the number they have will be significant compared to the total KGS and other online games they may have acquired.

Jon

Haven't I read somewhere that AlphaGo ran 24/7 for some time already?
It played against itself many times, perhaps indeed millions of games.

Stefany93 · **#23**

mumps wrote:

Stefany93 wrote:

I suppose what they did was feed the programs millions of pro games and its gradually developed a style to beat them, as @dankenzon suggested.

Actually not, as far as we're aware. Prior to the original Nature paper they only trained it using KGS games.

Although they may have added pro games I doubt whether there are even one million of them available - GoGod doesn't even have 100,000.

Even if they have a large number it's a good question as to whether the number they have will be significant compared to the total KGS and other online games they may have acquired.

Jon

I guess they had to be pro games, even in KGS in order for the AlphaGo to get stronger.

Stefany93 · **#24**

sybob wrote:

mumps wrote:

Actually not, as far as we're aware. Prior to the original Nature paper they only trained it using KGS games.

Although they may have added pro games I doubt whether there are even one million of them available - GoGod doesn't even have 100,000.

Even if they have a large number it's a good question as to whether the number they have will be significant compared to the total KGS and other online games they may have acquired.

Jon

Haven't I read somewhere that AlphaGo ran 24/7 for some time already?
It played against itself many times, perhaps indeed millions of games.

That's what I was thinking - it will need millions of games to develop its neural networks to a point where it will be able to beat all the pros.

Even if it generates them itself, or gets them online.

dfan · **#25**

My assumption is that it got markedly stronger over the year through self-play, not through training its policy network with more pro games. For one thing, a lot of its best moves were unusual from a pro perspective.

Stefany93 · **#26**

dfan wrote:

My assumption is that it got markedly stronger over the year through self-play, not through training its policy network with more pro games. For one thing, a lot of its best moves were unusual from a pro perspective.

That the thing - in order for it to play unusual moves, it has to learn what "usual" moves the pros play to surprise them.

dfan · **#27**

It doesn't seem to me that that has to be the case. If it's playing the best moves, some of them will happen to be unusual. I don't think that playing unusual moves was a goal in itself.

dankenzon · **#28**

Stefany93 wrote:

Quote:

http://www.youtube.com/watch?v=qv6UVOQ0F44

Neat! (no pun intended) :lol:

I also found this one - http://senseis.xmp.net/?AlphaGo Second headline talks about the algorithm, soooooooo... Alpha Go, you are busted :lol:

Joking aside, since AlphaGo was built on the NEAT principle with neural networks and what not, it proves its superiority to chess - in which chess programs merely store billions of games and just put the move with the highest chance of success.

I suppose what they did was feed the programs millions of pro games and its gradually developed a style to beat them, as @dankenzon suggested.

Millions of pro games? I will be surprised if at all the history there are 500 thousand!

By the way, AlphaGo learning experience was feeded by amateur games instead. It was used as a reference model to learn the game and find its own solutions.

So, he didn't developed a style to beat them. It has the goal of finding the moves that give the higher % of a sure win: it doesn't care about big wins but for a 100% sure win even if it means winning by only half point.

dankenzon · **#29**

sybob wrote:

mumps wrote:

Actually not, as far as we're aware. Prior to the original Nature paper they only trained it using KGS games.

Although they may have added pro games I doubt whether there are even one million of them available - GoGod doesn't even have 100,000.

Even if they have a large number it's a good question as to whether the number they have will be significant compared to the total KGS and other online games they may have acquired.

Jon

Haven't I read somewhere that AlphaGo ran 24/7 for some time already?
It played against itself many times, perhaps indeed millions of games.

Right.

dankenzon · **#30**

Stefany93 wrote:

dfan wrote:

My assumption is that it got markedly stronger over the year through self-play, not through training its policy network with more pro games. For one thing, a lot of its best moves were unusual from a pro perspective.

That the thing - in order for it to play unusual moves, it has to learn what "usual" moves the pros play to surprise them.

Again: he didn't learn nothing from pro games. It was feeded by amateur games. It just practiced and practiced and practiced and compared results until it came to the conclusion that some moves are more prone to have a higher probability of winning that others.

It doesn't care about usual or not. It simply evaluates and practices possibilities without never getting tired and with unlimited chance to make tests.

dankenzon · **#31**

Stefany93 wrote:

Quote:

http://www.youtube.com/watch?v=qv6UVOQ0F44

Neat! (no pun intended) :lol:

I also found this one - http://senseis.xmp.net/?AlphaGo Second headline talks about the algorithm, soooooooo... Alpha Go, you are busted :lol:

Joking aside, since AlphaGo was built on the NEAT principle with neural networks and what not, it proves its superiority to chess - in which chess programs merely store billions of games and just put the move with the highest chance of success.

I suppose what they did was feed the programs millions of pro games and its gradually developed a style to beat them, as @dankenzon suggested.

Finally, that second header is innacurate in the sense that, there's no algorithm for selecting moves. What they mention there are the mechanism used to implement "artificial brain" able to learn and find solutions. That's what they did.

dankenzon · **#32**

Stefany93 wrote:

dankenzon wrote:

Stefany93 wrote:

I wonder if they made the algorithm public yet?

They won't. Because there's no algorith. AlphaGo is a virtual brain. It learned and developed its own style.

Yeah... If that was possible we would be dead already by SkyNet.

There we go...
https://deepmind.com/blog/deepmind-ai-r ... g-bill-40/

or TL;DR

"...which is why we are excited to share that by applying DeepMind’s machine learning to our own Google data centres, we’ve managed to reduce the amount of energy we use for cooling by up to 40 percent. In any large scale energy-consuming environment, this would be a huge improvement. Given how sophisticated Google’s data centres are already, it’s a phenomenal step forward."
They feed the engine, it learns and it develops better way to manage energy.

Anzu · **#33**

Not entirely true, actually :mrgreen:

gowan · **#34**

The statement "everything you do is wrong" prompted me to remember attending a session at a Go Congress where a 6p asked for game records to analyze. Someone handed up a game record and the pro looked it over and said he didn't want to discuss this game because almost every move was a mistake. And I also remembe rwhen I first stared playing go, some 50 years ago, and my first regular partner was an amateur shodan; I had a strong feeling that everything I did was wrong.

dankenzon · **#35**

gowan wrote:

The statement "everything you do is wrong" prompted me to remember attending a session at a Go Congress where a 6p asked for game records to analyze. Someone handed up a game record and the pro looked it over and said he didn't want to discuss this game because almost every move was a mistake. And I also remembe rwhen I first stared playing go, some 50 years ago, and my first regular partner was an amateur shodan; I had a strong feeling that everything I did was wrong.

It's sad to find people like that.

Once a I met a very strong player from europe in Japan... he asked me for a game and asked me how strong are you? I was barely 6k then. He just made a disdain face and asked me to move out from the goban while making gestures with his hands.

I just told him F%$# U#@

Ulquiorra · **#36**

I have also felt the "everything you do is wrong". The proverb to find bravery is "even a 9dan can't live with one eye". Maybe now the proverb is " even AlphaGo can't live with one eye".

wineandgolover · **#37**

Not sure where to post this, here or the "Who's playing in Tygem?" thread.

Anyway, I made another AlphaGo video in which I talk about a few things I appreciate about its play. I hope you enjoy it. I've also included some variations and ideas from Jennie Shen, 2P. Please let me know your thoughts.

AlphaGo- Whatever You Do Is Wrong

Who is online