Neural networks artificial intelligence

Ytreza
Posts: 20
Joined: Tue Aug 27, 2019 5:56 pm

Re: Neural networks artificial intelligence

Post by Ytreza » Wed Sep 25, 2019 8:32 am

I may have access to a supercomputing center if needed (not sure if it must be dedicated to my pro projects or if there is some tolerance...) In any case I agree it's good to have an optimized code in the first place.

Yeah something I find very representative of the flaws in our understanding (and a bit scary if I think about applying AI to real life stuff) is the problem of the ladder in go, that the AI wasn't able to learn by itself. It resulted in an AI suddenly playing as a beginner. Now I wonder if this is still true with unsupervised learning? I guess not, as far as I understand the idea behind it.

Yeah python is still useful because there are so many available librairies!

For the question of the asymmetry, you mean that the output should be a law of probabilities on the set (all possible moves)x(3 pieces) ?

Thank you for answering my thoughts about strategy. That's very interesting. The goal of white in middle game would then be to attack corners in order to reduce the influence of black in edges, either by capturing attackers of by destroying their position. The goal of black would be to defend the corners trying to keep some control on edges.

Should we start with a much simpler game using the same setup? For instance, the king starts at the center with 4 defenders attached to him, while 4 attackers stand in the middle of the edges (facing the defenders). The goal for white is to escape at corners, the goal for black is to capture the king by placing an attacker (only one) on an adjacent square. There is no capture of pawns.

Can you send me the movegen you already have so that I can have a look at it?

nath
Posts: 65
Joined: Mon Oct 09, 2017 9:34 pm

Re: Neural networks artificial intelligence

Post by nath » Sat Sep 28, 2019 10:42 pm

Ytreza wrote:
Wed Sep 25, 2019 8:32 am
I may have access to a supercomputing center if needed (not sure if it must be dedicated to my pro projects or if there is some tolerance...)
Supercomputers are build to have crazy ram with crazy contentedness. We don't need need that, we need a lot of games being played to learn from that. That is a very simple task that can be done decentralized, so we don't need special hardware for that.
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
In any case I agree it's good to have an optimized code in the first place.
The DNN stuff of tf is pretty optimized. The movegen and the win evaluation are a different story. Just look at the bottleneck.
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
Yeah something I find very representative of the flaws in our understanding (and a bit scary if I think about applying AI to real life stuff) is the problem of the ladder in go, that the AI wasn't able to learn by itself. It resulted in an AI suddenly playing as a beginner. Now I wonder if this is still true with unsupervised learning? I guess not, as far as I understand the idea behind it.
Sort of everything can be learned if you attach *enough* computation power. It's just for a lot of problems we tend to hit hardware limits...
Learning the ladder isn't even that hard, since you punish the move. It's just that the punishment is pretty far away, which makes you require more examples.

We will never know the difficulties unless we try.
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
Yeah python is still useful because there are so many available librairies!
Mainly tf/keras. They seem to me like the real reason to use python.
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
For the question of the asymmetry, you mean that the output should be a law of probabilities on the set (all possible moves)x(3 pieces) ?
What do you mean by "law of probabilities"? I just wanted to say that we should include the piece when we define the move for the purpose of the input layer of the network.
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
The goal of white in middle game would then be to attack corners in order to reduce the influence of black in edges, either by capturing attackers of by destroying their position. The goal of black would be to defend the corners trying to keep some control on edges.
Maybe, but maybe there is a real strong corner attack nobody found yet. I don't expect that, but I can't rule it out completely. I'm pretty convinced casshern could win the wtf world cup, even he wasn't allowed to use edge forts.
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
Should we start with a much simpler game using the same setup? For instance, the king starts at the center with 4 defenders attached to him, while 4 attackers stand in the middle of the edges (facing the defenders). The goal for white is to escape at corners, the goal for black is to capture the king by placing an attacker (only one) on an adjacent square. There is no capture of pawns.
That game has a totally different dynamic. If you want to study that, go ahead. But be aware that no result carries over. If you have ever played this, you know that the board size changes the game more than drastically.

I'm only helping for specific 11x11er variants (Copenhagen, Fetlar). Fetlar is much easier than Copenhagen. I'd suggest to start with Fetlar + different move limits (maybe even the 100 moves you suggested) + shieldwall (that's still pretty simple).
Ytreza wrote:
Wed Sep 25, 2019 8:32 am
Can you send me the movegen you already have so that I can have a look at it?
I sent you a mail.

Regards
nath

Michel
Posts: 1
Joined: Tue Oct 22, 2019 9:06 am

Re: Neural networks artificial intelligence

Post by Michel » Tue Oct 22, 2019 8:31 pm

If anyone is interested, i have a collection of 70.000 games of tablut.

It was generated by computer self-play, and only uses 4 ply searches. I did some tests with automated learning from the database, but it looks like it is not trivial; one complication may be that the interactions between the pieces are long-distance, and my hand-tuned evaluation function outperforms the learned one.

Post Reply

Who is online

Users browsing this forum: No registered users and 2 guests