Google's Deep Mind "Alpha Zero" - amazing AI concept prooved in GO and chess

To add comments, please log in or register
Amir Yacoby
1261
Amir Yacoby  

Computer chess evolved a lot since Deep Blue won against Kasparov in 1997. 
Now computers are 600 ELO points stronger then any human being and no point in matching between them and any world champion.
Computer chess contest are being held regularly, and the top chess softwares of last years are StockFish and Komodo, each about 3400 ELO points (human strongest is about 2900).

Those computer chess programs are programmed using a lot of chess theory incorporated into them from chess theories evolved over the last 200 years, and they include an opening library they use In the opening, meaning there is no real AI in them, just calculations and rules.

In 05 December 2017, few days ago, google's Deep Mind company published a paper regarding their general AI program that uses Reinforcement learning that last year defeated the GO world human champion.
Their modified program from GO to chess and thaught it the chess rules giving it nothing more then this.

Then they let it play against itself and learn for 4 hours and after that they matched it with StockFish - the greatest chess software up to now.
They played 100 games from which Deep Minds program Alpha Zero won 28, and the rest was drown (no loss).

The games played are absolutely amazing, as many chess rules thought to be concrete up to now, were broken by the program which play in a style that does not let the opponent program air to breath (it sort of fights against the opponent mobility, leaving it's pieces useless and not mobile).

This was achieved in just 4 hours of training against itself using reinforcement learing with no clue as to how to play given to it.

Absuloutely amazing.


Amir Yacoby
1261
Amir Yacoby  
Aimak:

Very interesting, but can DeepMind make an AI able to "manually" trade forex profitable?


I think it can, google bought it because it was general purpose AI. And in deed it proved to make it in two games by now, forex is another game. 
They intend to use it in medicin and research, but the idea is valid in trading in my opinion.

Amir Yacoby
1261
Amir Yacoby  
Aimak:

My point is that Go and chess are self-contained games, this is with no external inputs that can affect the game. What could be the performance of AlphaGo Zero if a number of stones get withdrawn or deposited from table random and infrequently.


The number of possibilities in GO exceeds those of chess by the number of atoms on earth - it is far more complex game. It did not learn all possibilities but really seem to understand something about the game logic.

So in this sense, the game is unexpected for a player - just like trading. 

From what I know about reinforcement learning, I think it can suit trading and it certainly will be explored in the near future.

This might result in a new way of trading that is not conceived today.

Amir Yacoby
1261
Amir Yacoby  
Aimak:

I think you are fully correct. I would love to see an AI trading the forex market. However, the power of such software development could give a gigantic advantage to the owner in comparison to the human and algo players. Do you think that such AI could skew the market?


I think it should be interesting to see. It will probably be ahead of the others including algorithms, but when it will become world spread and most traders will be AI's then we will have something else in the market.

Because, theoretically, a chess AI playing against itself should be draw in long run.

This will happen eventually in the markets, as AI trades againt other AI's no one will have an edge.

But for start, I am interested to see the first ones in the markets.

Marco vd Heijden
Moderator
8415
Marco vd Heijden  

This is actually, old news.

Not the paper from a few days ago, but the AI that beat the GO Player.

I know for a fact that there has already been some work done on (deepmind) Forex AI robots in the past.

I believe this was related to http://openai.com/ 

It's basically just a matter of time the AI has to be trained by trading and analyzing the outcome, learning and getting 'smarter' 

So that it makes better decisions next time.

I truly believe that these system will outperform human traders in the future (if they are not already, but hidden) 

Because i have seen one being trained on a rather simple task: reading handwritten language.

This was problematic at first, but after the training it achieved 99,8% correctness and is being used to sort mail by reading the handwritten addresses etc.

It just has to be programmed to run a trading terminal, the rest will be done by the AI itself.

OpenAI
OpenAI
  • 2017.12.06
  • openai.com
OpenAI is a non-profit AI research company, discovering and enacting the path to safe artificial general intelligence.
bonnytu436
14
bonnytu436  

hi guys, very interesting this information, but I think forex is something more than this.

Amir Yacoby
1261
Amir Yacoby  
Marco vd Heijden:

This is actually, old news.

Not the paper from a few days ago, but the AI that beat the GO Player.

I know for a fact that there has already been some work done on (deepmind) Forex AI robots in the past.

I believe this was related to http://openai.com/ 

It's basically just a matter of time the AI has to be trained by trading and analyzing the outcome, learning and getting 'smarter' 

So that it makes better decisions next time.

I truly believe that these system will outperform human traders in the future (if they are not already, but hidden) 

Because i have seen one being trained on a rather simple task: reading handwritten language.

This was problematic at first, but after the training it achieved 99,8% correctness and is being used to sort mail by reading the handwritten addresses etc.

It just has to be programmed to run a trading terminal, the rest will be done by the AI itself.

You are right, the GO is from last year. But I think for the chess they gave it a different approach, just gave it the rules and nothing more.

Anyway, as a chess fan, who knows and owns the leading softwares today, I was really impressed by the chess skills it performed after only 4 hours of training by itself.
The way it plays is really a new thing in chess. Up to now, chess masters played and trained using leading softwares, and tried to play like the software. 
But now, the whole concept and play style including openings that are part of chess theory are being challenged by a fresh new thinker.

I didn't know about forex trials, interesting. It's reasonable that somewhere someone tries these methods already.

Alain Verleyen
36571
Alain Verleyen  
Amir Yacoby:

...


In 05 December 2017, few days ago, google's Deep Mind company published a paper regarding their general AI program that uses Reinforcement learning that last year defeated the GO world human champion.
Their modified program from GO to chess and thaught it the chess rules giving it nothing more then this.

Then they let it play against itself and learn for 4 hours and after that they matched it with StockFish - the greatest chess software up to now.
They played 100 games from which Deep Minds program Alpha Zero won 28, and the rest was drown (no loss).

The games played are absolutely amazing, as many chess rules thought to be concrete up to now, were broken by the program which play in a style that does not let the opponent program air to breath (it sort of fights against the opponent mobility, leaving it's pieces useless and not mobile).


Interesting.

It may depends about what definition you give to intelligence, but in my opinion there is no intelligence at all in AI in general or Alpha Zero AI, it's a marketing term. The intelligence is in the underlying code. All the rest is a matter of input, output, probability and very high speed. Of course it's amazing and very interesting nonetheless. But everyone should continue to think.


DeepMind co-founder and CEO Demis Hassabis said the programme was so powerful because it was "no longer constrained by the limits of human knowledge".

He believes that if applied to big health problems, such as defeating Alzhiemer’s, it could, in a matter of weeks, come up with cures that would have taken humans hundreds of years to find.
Seriously ? Now tell us why it didn't do it ?
Amir Yacoby
1261
Amir Yacoby  
Alain Verleyen:

Interesting.

It may depends about what definition you give to intelligence, but in my opinion there is no intelligence at all in AI in general or Alpha Zero AI, it's a marketing term. The intelligence is in the underlying code. All the rest is a matter of input, output, probability and very high speed. Of course it's amazing and very interesting nonetheless. But everyone should continue to think.


DeepMind co-founder and CEO Demis Hassabis said the programme was so powerful because it was "no longer constrained by the limits of human knowledge".

Seriously ? Now tell us why it didn't do it ?
Yeh, but the underlying code does not tell the machine how to do it. It finds a way by itself. Moreover, the developers don't have a way to know what the machine learned.
There is a known story about first days of AI development, when US army trained a tank AI to recognize other tanks around it.
They trained it using a series of images that some contained tank and some not, and for each image in the train set they told the AI the right answer - if
it contains a tank or not.
Then they tested it with a test set, and the AI was giving strange results, it did not succeed.
They found out later that each image they gave it that was cloudy sky, the AI said "Yes" - there is a tank.
They returned to the train set and saw that all the images of the tanks were taken on a cloudy day.

This is the beauty of modern AI. Using the training it finds the closest multi dimensional functions that describe what it is set to find. While it's true that it's not real intelligence, it is still very close to intelligent behaviour.

Amir Yacoby
1261
Amir Yacoby  
By the way, about the speed:
Stockfish, the non-AI chess software - calculates 200 million positions per second (in the contest between them, given the hardware).
Alpha-Zero calculated 70 thousand position per second. It is 2000 times less which is a huge differrence.

The reason is it does not rely on brute force at all. The AI is giving it a kind of "intuition", and it only calculates as little as to make sure that it's "intuition" is correct.
mrluck1
926
mrluck1  

Basically in this topic you are asking what's the difference between chess, dota2, go, and the markets, because on those games, AI already beat any human badly, but you are forgetting that they are single player, i would like to see an AI working with humans to beat the other side, it works well only if it's playing alone, the market is fairly more complex because it's


- a multiplayer "game", because there are several participants

- you don't know who are your partners,actually you are playing alone, but with friends ( the one who are on your trade side) that you can't define if they're your allies or not, because they can change their side


That way nobody can never measure AI vs human on the markets because that's NOT a game and also:

1- just because some AI algorithms have  beaten human players in some games does not mean it's more intelligent, because

a- it's very hard to define what intelligence is

b- until today, nobody knows exactly the algorithm of the brain to learn and use this knowledge learned


But in any game, eventually an AI will have the upper hand, we are doomed about this

123
To add comments, please log in or register