The State Of Play Tafl Online

Post Reply
Posts: 30
Joined: Fri Mar 04, 2016 8:44 pm

The State Of Play Tafl Online

Post by Tuireann » Mon Dec 31, 2018 10:51 pm

On whats going on with PTO:

The PTO server crashed, I did manage to recover the database(which had apparently not been backing up automatically due to the reason the server crashed); However it happens to be in it's binary format thus a little work must be done to it for it to be usable again. The inherant complexity of PTO also adds time to a redeployment. Any emails should have bounced back lest they will have become one with the void as the PTO server also hosted the PTO email. Recently at my work our CEO passed away after running the company for 30+ years and the new CEO has been making a lot of big changes before the end of the year which has been time consuming to say the least. Fortunately that year just ended and so I plan on finishing the redeployment this weekend with the best case scenario being that it will be running by next week. So far the biggest issue with the original crash has been resolved(faulty snapshot setup which led to filesystem corruption; oh the irony) and the data has been moved to the new system. Any running games will probably be set to declined by time out as not to affect anyone's ELO. Sorry for the inconveniance but unfortunately my day job has to take priority in my life and with 6-7 day work weeks with 12+ hour days I haven't had the time or will to even work on getting PTO setup again.

On AI Mirror Games:

I saw some questions about what would stop someone from running a mirror game with one of the AIs that were part of PTO. Not to talk down the efforts of Fishbreath and myself but the resident PTO AI, Monty(written by me), used Monte Carlo simulation(A very very basic version of what AlphaGo uses to train itself) and it's unsuitable for Tafl because the game of Go is almost always moving towards it's end game as pieces are rarely removed from the board and that made it ineffective for a game like Tafl where judging the value of board positions is extremely difficult. OpenTafl(written by Fishbreath) on the other hand I would characterize as taking a more Chess-like approach and it was light-years more advanced than the Monty AI. It would attempt to brute-force a large number of possible moves and evaluate the value of each position based on where the pieces would likely be 5-7 moves ahead. Even given 30 seconds per move on PTO it would struggle with anything larger than a 9x9 board; In fact for Copenhagen it would only make it 4-5 moves deep at best and it would consume 2+GB of memory in the process which would increase exponentially the deeper it got. For it, in its current state and in my own personal opinion, to be useful in winning against a 1700+ rated player you would effectively need a computer with at least 1TB(1024GB) of memory(I work with 'super-computers' at my day job and our most powerful single node has only 512GB). The lack of an ability to computationally evalutate a board position mixed with the complexity of a game like Tafl makes it unlikely an AI will defeat a competent human at it any time soon.

Happy New Year,

Posts: 1
Joined: Sun Dec 09, 2018 11:59 am

Re: The State Of Play Tafl Online

Post by Steiger » Tue Jan 01, 2019 7:07 pm

Thanks for information, Tuireann!

Posts: 38
Joined: Mon Oct 09, 2017 9:34 pm

Re: The State Of Play Tafl Online

Post by nath » Mon Jan 14, 2019 7:39 am

Well, while the complexity of Hnefatafl is bigger than the complexity of chess, your comparison isn't fair. We don't have the same amount of preordering and razoring a modern chess engine has. No modern chess engine could calculate anything interesting without cutting (more than 80 % of the variants)...stockfish is way better at this.

I as a human razor a lot. The interesting part is, that we should be able to do a decent preordering on the moves, just based on the move (and moving piece) itself. I mean things like outer rows are probably important as well as king moves are almost always interesting.

OpenTafl is basically trying the impossible: uniting radically different board games with one engine. I'm not only talking about the different kings setting, but also about the size. Some of the constants aren't relevant for a lot of variants. The eval also seems to be mainly tested on boars with sizes of 9x9 and smaller.

While one can get 2 TB RAM machines with standard parts these days, I never had one of those in my hands. If one would want to archive something like that, the hardware limit these days would be NVMe SSD configured as swap space (one can easily get 8 TB that way), but as I wrote above the solution isn't bare metal, it's intelligent cutting.
My take on this: Just because I failed on my own four years back, doesn't mean it's impossible...and I'm talking about the average desktop today.
What I am confident about, is that the computer can give us new perspectives about the game, even when they still loose to top players.

Happy new year!

Posts: 8
Joined: Fri Dec 28, 2018 6:39 pm

Re: The State Of Play Tafl Online

Post by unhandyandy » Sat Jan 19, 2019 9:17 pm

I would think that Tafl might be easier to razor than chess, because in the chess the tactics are so sharp a game can seem to completely turn around with one bad (or brilliant) move; that seems to be less the case in tafl (but I'm a beginner).

For example, an important metric in tafl is how many moves away the king is from his goal. That should be relatively easy to measure.

Post Reply

Who is online

Users browsing this forum: No registered users and 1 guest