Page 2 of 3

Re: The State Of Play Tafl Online

Posted: Sat Feb 02, 2019 4:21 pm
by nath
unhandyandy wrote:
Fri Feb 01, 2019 5:24 pm
There is a database here.
Oops, I missed that - should be very useful. Is the code by Hagbard? Is it open source?
I'm talking mainly about the game archive, but there should be some older games around as well. I had some pleasant acommunication with Hagbard, he was helpful. I don't know what open source is supposed to mean in that context. Are we talking about the same kind of database? I'm just talking about the played games.
unhandyandy wrote:
Fri Feb 01, 2019 5:24 pm
Of course I've playing my engine continuously over the past couple weeks. I'm a beginner, so the fact that I can beat it doesn't look good. Have you tried playing it? Just open index.html in a browser.
My main observation is that this thing is really slow.

Advice regarding playing strength: If you are really interested in playing strength, test your engine against other engines. "Other engine" is a pretty flexible term. Test it in "selfplay" against an old version of itself. That already gives a rough estimate whether your last change was reasonable or not. And it's more accurate than me trying to play it.

Besides: I'm definitely no expert for Brandubh. I've just played a game against your engine. I've noticed that the throne doesn't seem to be hostile to white. You see even the rules are new to me. :D
The second and third game both resulted in stalemates where we both moved back and forth. I stopped after nearly 50 full moves, since the thing is really slow and I didn't want more time watching the "a web page is slowing down your browser".

Re: The State Of Play Tafl Online

Posted: Sat Feb 02, 2019 5:47 pm
by unhandyandy
I had some pleasant acommunication with Hagbard, he was helpful. I don't know what open source is supposed to mean in that context. Are we talking about the same kind of database?
Sorry, I wasn't clear - I meant the "web app" that plays through games. Is it server or client side? Maybe I could hack it to add editing capabilities.
My main observation is that this thing is really slow.
Yes. :oops:

It can take as much as a minute to move. But it remembers previous games and will play remembered lines instantly.
I've noticed that the throne doesn't seem to be hostile to white. You see even the rules are new to me. :D
When empty it's hostile to white pawns, but not the king. I wanted to use the simplest rules possible, so I started with Damian Walker's and simplified further.
Test it in "selfplay" against an old version of itself. That already gives a rough estimate whether your last change was reasonable or not. And it's more accurate than me trying to play it.
Yes, in fact I tried stealing(?) an idea from AlphaZero by playing 100000 games against itself to set the parameters. It helped a bit.

But the first line of QA is that I shouldn't be able to beat it easily. So far I can.
The second and third game both resulted in stalemates where we both moved back and forth. I stopped after nearly 50 full moves,
Hmmm, it's supposed to recognize repetitions, and give Black the win...
"a web page is slowing down your browser".
In Chrome you can't turn that message off unless you're in dev mode. In Firefox you can turn it off in about:config.

Re: The State Of Play Tafl Online

Posted: Sun Feb 03, 2019 3:23 am
by nath
unhandyandy wrote:
Sat Feb 02, 2019 5:47 pm
Sorry, I wasn't clear - I meant the "web app" that plays through games. Is it server or client side? Maybe I could hack it to add editing capabilities.
I think it's not only java, but plays awful. But I guess that is a question @Hagbard .
unhandyandy wrote:
Sat Feb 02, 2019 5:47 pm
Yes, in fact I tried stealing(?) an idea from AlphaZero by playing 100000 games against itself to set the parameters. It helped a bit.
That idea is way older than googles research on anything, yet alone alpha zero.
unhandyandy wrote:
Sat Feb 02, 2019 5:47 pm
Hmmm, it's supposed to recognize repetitions, and give Black the win...
If you are willing to do more and more and more full moves...maybe. I had something else to do today, it takes quite some time.

Re: The State Of Play Tafl Online

Posted: Sun Feb 03, 2019 5:43 pm
by unhandyandy
Sorry, I wasn't clear - I meant the "web app" that plays through games. Is it server or client side? Maybe I could hack it to add editing
capabilities.

I think it's not only java, but plays awful. But I guess that is a question @Hagbard .
I was referring to the game viewer app at

http://aagenielsen.dk/hnefatafl_replay_medklasser.php

I assume that's by @Hagbard also.
That idea is way older than googles research on anything, yet alone alpha zero.
Hmm, so there isn't that much about AZ that's actually original - Google just had the resources to combine the best elements from a multitude of sources.

Re: The State Of Play Tafl Online

Posted: Sun Feb 03, 2019 11:10 pm
by nath
unhandyandy wrote:
Sun Feb 03, 2019 5:43 pm
Hmm, so there isn't that much about AZ that's actually original - Google just had the resources to combine the best elements from a multitude of sources.
Compared to AlphaGo Zero, which brought the justification for the value and policy network, the novelty of alpha zero is slim. Alpha Zero is basically just AlphaGo Lee stripped away all the layers of complexity that were not really needed. Most of them were introduced to save resources.

The results are actually results of Deep Mind (not really Google). Deep Mind got happily acquired by google to have access to google resources. So in a sense it was to be expected.

Still it's quite a feast, both regarding software and hardware. I think an engine that just plays a variant that is roughly equivalent to Fetlar with shieldwall might be doable with resources of a hobby project.

Re: The State Of Play Tafl Online

Posted: Mon Feb 04, 2019 8:21 pm
by unhandyandy
AlphaGo Zero, which brought the justification for the value and policy network
I'm probably wrong, but I thought those ideas came from Monte Carlo search.
The results are actually results of Deep Mind (not really Google). Deep Mind got happily acquired by google to have access to google resources.
The results wouldn't have come so fast or been so spectacular without Google's resources.
I think an engine that just plays a variant that is roughly equivalent to Fetlar with shieldwall might be doable with resources of a hobby project.
Like LeelaZero? Do you think there are enough cpu cycles in the hnefatafl community?

Re: The State Of Play Tafl Online

Posted: Tue Feb 05, 2019 2:51 am
by nath
unhandyandy wrote:
Mon Feb 04, 2019 8:21 pm
I'm probably wrong, but I thought those ideas came from Monte Carlo search.
You are right on that one. What I meant was, that I don't know of any reasonable DNN experiment for the value network before. It's a very different method.
unhandyandy wrote:
Mon Feb 04, 2019 8:21 pm
The results wouldn't have come so fast or been so spectacular without Google's resources.
Nobody ever doubted that. The founder of Deep mind even said that that was the reason for his decision to sell the company to google in the first place.
unhandyandy wrote:
Mon Feb 04, 2019 8:21 pm
I think an engine that just plays a variant that is roughly equivalent to Fetlar with shieldwall might be doable with resources of a hobby project.
Like LeelaZero? Do you think there are enough cpu cycles in the hnefatafl community?
Well...enough for what? From the perspective of understanding the game we are off much worse with Hnefatafl than we were in chess. LeelaZero has eaten more than a thousand fold of the resources we'll ever have. I still suspect that Hnefatafl (at least besides the edge fort) is better suited for that approach than chess.

So I'd definitely say that should have enough resources to train something that can teach us something about the game. Whether we have enough resources to create something that could be able to beat me? No idea, but we'll never know unless we try.

Re: The State Of Play Tafl Online

Posted: Tue Feb 05, 2019 9:19 pm
by unhandyandy
I don't know of any reasonable DNN experiment for the value network before. It's a very different method.
Deep Mind's achievement is very impressive, but I don't think there was any real conceptual breakthrough that wouldn't have occurred to most people in the field.
So I'd definitely say that should have enough resources to train something that can teach us something about the game. Whether we have enough resources to create something that could be able to beat me? No idea, but we'll never know unless we try.
Well I'd be happy to offer my meager resources.

Re: The State Of Play Tafl Online

Posted: Wed Feb 06, 2019 5:59 am
by nath
unhandyandy wrote:
Tue Feb 05, 2019 9:19 pm
Deep Mind's achievement is very impressive, but I don't think there was any real conceptual breakthrough that wouldn't have occurred to most people in the field.
The thing that was impressive, was to collaboration of the ML-scientists, programmers and the data centers. Another interesting part was the usage of TPUs...so I should include hardware specialists in the list.
unhandyandy wrote:
Tue Feb 05, 2019 9:19 pm
So I'd definitely say that should have enough resources to train something that can teach us something about the game. Whether we have enough resources to create something that could be able to beat me? No idea, but we'll never know unless we try.
Well I'd be happy to offer my meager resources.
I currently thinking about the source. I don't mind investing some time for an appropriate the movegen (profiling it and cutting it to the purpose). Yet I'm not that keen to do more with tensorflow that I absolutely have to. Especially since I'd need to have a look at cffi. I was talking about this to Brench a month ago, but his schedule seems to be even tighter than mine.

Re: The State Of Play Tafl Online

Posted: Wed Feb 06, 2019 8:04 pm
by unhandyandy
I take it you've looked through the LeelaZero code?