Neural networks artificial intelligence

Post Reply
Hagbard
Posts: 435
Joined: Sat Mar 14, 2015 6:07 pm

Neural networks artificial intelligence

Post by Hagbard » Sun Feb 26, 2017 9:44 am

In internet archives is a ten year old research paper of a neural networks artificial intelligence project in Australia:

"Evolving Players for an Ancient Game: Hnefatafl", Philip Hingston, Edith Cowan University, Australia, 2007
https://pdfs.semanticscholar.org/ba25/8 ... 784d7c.pdf

The paper mentions the shieldwall rule, which we use in Copenhagen Hnefatafl.

Hagbard
Posts: 435
Joined: Sat Mar 14, 2015 6:07 pm

Re: Neural networks artificial intelligence

Post by Hagbard » Sun Feb 26, 2017 9:48 am

25.2.2017 I received this email:
Hello Sir,

I'm a data scientist that was looking to work on a new side project with some collaborators: creating an AI that can play Hnefatafl from either
side. This wouldn't be by conventional heuristic means, but by feeding it the boards and moves of previous games played by humans.

The model implemented is a deep learning
<https://en.wikipedia.org/wiki/Deep_learning> one which will attempt to learn from a corpus of game data. It doesn't even have to be any
"grandmaster"-caliber game data, as even amateurs make moves based mostly in their own self-interest (at least, in the short-term).

Such efforts have been successfully implemented
<https://github.com/imthexie/giraffe> and even published
<https://arxiv.org/pdf/1509.01549.pdf> for symmetrical games like chess, but I don't find any records of this being done for asymmetric strategy
games like Hnefatafl. Besides it being less globally popular, I suspect this is because there exists no massive corpus of game data.

And this is why I contact you tonight. I stumbled across your site, and I have noticed that you have a sizable archive of games and their moves.
Might I ask that you make available this data to me and my collaborators so that we can attempt such an effort? It's not likely to succeed, but I think
it would be rewarding to give it a go. We can work with any format you provide.

In any case, this is just a side-project of ours, and if you wish to keep your hard-earned and well-curated data to yourself, I completely understand.

Many thanks for your time and consideration,

Bryan Dannowitz

Hagbard
Posts: 435
Joined: Sat Mar 14, 2015 6:07 pm

Re: Neural networks artificial intelligence

Post by Hagbard » Sun Feb 26, 2017 9:59 am

Bryan Dannowitz wrote:
creating an AI that can play Hnefatafl from either
side. ... by feeding it the boards and moves of previous games played by humans.
The model implemented is a deep learning ... one which will attempt to learn from a corpus of game data.
Hello Bryan Dannowitz,
that's an interesting project. Could it result in a neural networks robot accessible for free as an opponent for tafl players on the internet?

I would recommend to try the project on Historical Hnefatafl 9x9 (Saami Tablut) and Historical Hnefatafl 11x11 (Welsh Tawlbwrdd).
The game rules are found here:
http://aagenielsen.dk/historical_hnefatafl_rules.php
Adam Bartley did a pdf-layout of the rules here:
http://aagenielsen.dk/TaflRulesEnglish.pdf

A full extract of games between strong players are put on the net here:
Tablut games
http://aagenielsen.dk/tafl_tabluts.txt
Tawlbwrdd games
http://aagenielsen.dk/tafl_tawlbwrdds.txt


A ten year old research paper of a perhaps similar project in Australia can be found in internet archives:
"Evolving Players for an Ancient Game: Hnefatafl", Philip Hingston, Edith Cowan University, Australia, 2007
https://pdfs.semanticscholar.org/ba25/8 ... 784d7c.pdf
but this old project is probably not based on as many game data.


Good luck with the project!
Aage Nielsen, Denmark

Hagbard
Posts: 435
Joined: Sat Mar 14, 2015 6:07 pm

Re: Neural networks artificial intelligence

Post by Hagbard » Sat Jan 27, 2018 7:42 pm

"cc dark" sent this mail:
Greetings!

You should totally use neural net to make this game complete.
I suggest to stay close with one version of the game as much as possible.
Keep a single set of rules and call it just simple Tafl.

Put aside all historical threads, who cares how they played it? All great games walked long roads of metamorphoses to become truly playable. Chess, Xiangqi, Shogi are great examples.

But Tafl is so unique, wonderful, asymmetric game with one king! It deserve to be one of the top boards games. We don't have centuries, but we have technology at our side to shape this game to its potential.

Also, "escape to the edge win" aesthetically is more beautiful for gameplay in my opinion.

Fishbreath
Posts: 14
Joined: Sun Apr 24, 2016 12:13 am

Re: Neural networks artificial intelligence

Post by Fishbreath » Tue Jan 30, 2018 8:40 pm

Oh, data! Excellent. One of these days, I'm hoping to compile a corpus of games in OpenTafl save-game format; I already have a converter for PlayTaflOnline, but still need to write one to handle games from here.

If you have more data in that format, I might go ahead and write a parser.

As far as neural networks go, I'd love to take a crack at that sometime. Even with the combined set of games from here and PlayTaflOnline, I doubt there's enough data for directed learning, though, and I certainly don't have sufficient computing resources to do undirected learning in the same sense as DeepMind.

Post Reply

Who is online

Users browsing this forum: No registered users and 0 guests