A human representative against the machines

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

A human representative against the machines

Post by Fishbreath » Tue May 24, 2016 4:18 pm

As some of you may know, I'm running a tournament for computer tafl players. It's traditional, in tournaments like this, for the winning computer to face off against a strong human player, and now that I have five confirmed participants, I figured I would put the word out and see if anyone is interested in representing the humans against the computers.

The rules will be a slight modification of the Fetlar rules, aimed at reducing the potential for long games and draws. The game will be timed, but the human player will have wide latitude in selecting the timing rules. (OpenTafl supports a standard chess-style game clock, with the addition of Fischer increment time and go-style overtimes.)

Tuireann
Posts: 29
Joined: Fri Mar 04, 2016 8:44 pm

Re: A human representative against the machines

Post by Tuireann » Tue May 24, 2016 7:32 pm

I've always liked byo-yomi as a time control mechanism. Would it not be better to use something like the Historical Tablut rules if you wish to reduce game length and draws? There's far more disparity between the length of a game and the side of the player in Fetlar and other corner escape variants versus the more even results we see here from edge escape variants.

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

Re: A human representative against the machines

Post by Fishbreath » Tue May 24, 2016 7:52 pm

Unfortunately, it's a little late for a change now—I'm not a huge fan of the Fetlar rules myself, but along with Copenhagen, they're a widely-accepted tournament standard, and relatively simple to implement. Beyond that, Fetlar and other 11x11 games more accurately capture the mathematical complexity of the tafl games. A 9x9 board is a little too easy to conquer with tried and tested, brute-force AI techniques, whereas the 11x11 board requires some extra thought, and, perhaps, innovation.

As regards timing, I wrote about that earlier this year. I think byo-yomi is an extremely elegant way to handle the 'running out of time' problem, too.

Tuireann
Posts: 29
Joined: Fri Mar 04, 2016 8:44 pm

Re: A human representative against the machines

Post by Tuireann » Tue May 24, 2016 8:08 pm

I have some questions/criticisms about your modifications to Fetlar.
1. No draws are allowed. Draw fort formations, where the king is free to move inside an invincible formation of defenders, have no special meaning.
Would this not just create an infinite game? A white player in a draw fort with infinite movement and no possibility to win would have no incentive to end the game. Would not an AI simply move forever creating an infinite an unending game since the players can no longer agree to a draw(as is customary in such a circumstance).

There is a qoute from Crust on here explaining the draw fort as a strategy that this rule modification would annul.
... white may as well try to break out and seize control of a corner, and then go for a draw if that fails, and black manages to complete the encirclement. The interesting part comes when black realises that white has started to do this, and must work out how to stop it - often black will have to risk exposing a corner in order to prevent white from achieving the draw position. If that happens, white may have another chance at a win. So white is using the threat of a draw to disturb black's position, and thus create chances for a win. ... I don't see anything wrong in using the threat of a draw in the mid-game to get some leverage over black, especially after black has completed the encirclement - I mean, what else is there for white to do, apart from waiting to die?
Completely removing draws would in theory remove one of the key strategies Fetlar has and would instead likely introduce infinite games.
2. If the defenders are fully surrounded (that is, if no defending piece, given an unlimited number of moves from the current position, with no responses from the attacker, can reach the edge), the defenders lose.
I assume this is just the surround rule that's employed on almost every variant here. It's a good rule as it ends the game for black slightly quicker.
3. Board positions may be repeated a maximum of two times. On the third repeated board position, the player who made the move to enter the repeated position wins. (Read another way, this means that the aggressor—the player who is forcing the threefold repetition—loses.)
This rule as it is here just seems to remove small draw forts. Which if allowing draws would slightly tilt the balance to the favor of black. But as you cannot do a draw fort this would only seem to end the game in situations in the rare instance black does not have enough pieces to prevent white from escaping without moving forever. I am not sure if its better or worse to favor black in these situations.

I don't think these modifications will allow you to achieve shorter games. I think by removing the draw fort strategy you cripple a large part of the Fetlar strategy and produce a completely new variant that won't have any predictable expectation of balance and so perhaps Fetlar is not particularly well suited for what you wish to accomplish with your tournament.

EDIT: just saw that you responded before I posted this. Well yes i suppose it may be a little late now :P but maybe something to take with a grain of salt for your next AI tournament!

Tuireann
Posts: 29
Joined: Fri Mar 04, 2016 8:44 pm

Re: A human representative against the machines

Post by Tuireann » Tue May 24, 2016 8:11 pm

A 9x9 board is a little too easy to conquer with tried and tested, brute-force AI techniques, whereas the 11x11 board requires some extra thought, and, perhaps, innovation.
I've played a number of 9 depth AI games with your AI and it has yet to beat me in 9x9 :P However my preference is what used to be called Tablut 11x11 here. It's pretty nice as far as complexity goes. Even with Copenhagen and Fetlar being used in tournaments I still think this place has made it pretty obvious that the 'Historical Hnefatafl' rules make for a better game.

Tuireann
Posts: 29
Joined: Fri Mar 04, 2016 8:44 pm

Re: A human representative against the machines

Post by Tuireann » Tue May 24, 2016 8:27 pm

I am interested in this idea of using the main time for the opening and then having larger overtime periods for mid-end game moves. Have you played any human games this way? It sounds like an interesting variation of how byo-yomi typically works. I will have to find some skilled human players to give it a try.

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

Re: A human representative against the machines

Post by Fishbreath » Tue May 24, 2016 8:37 pm

Shorter isn't necessarily the goal, so much as 'eventually decisive'. Threefold repetition, fatal surrounding, and no draw forts make for a game which is much more likely to end than a Fetlar game. They may in some cases take a long time, but threefold repetition catches all but the most elaborate draw forts, and I don't think the balance issue is especially critical with the model I have planned—namely, that each AI plays each other AI twice, once with each side. If you can win faster, you're probably better than your opponent.

It's also a relatively simple game to teach computers to play, and implementing the rules of tafl is less interesting from an AI perspective than are the specific AI techniques one uses to search the problem space. Next year, if this year is successful, I may move to a different ruleset. (At the same time, I like that the tafl family of games is still a family of games; I feel it would lose something if everyone agreed on one set of rules.)

As for OpenTafl's AI and its playing strength, you'll note I only said that 9x9 a little too easy to conquer—not that it's easy overall. :P The version distributed right now is something of a placeholder; I'm working on all the structural stuff for the tournament before I spend too much effort on improving it.

I haven't had a chance to play that way against a human yet. In the next few weeks, I should have OpenTafl's network play functionality more or less squared away, at which point I can begin to test against players of about my strength. (The doors will also be open to anyone else who wants to give it a try.)

Tuireann
Posts: 29
Joined: Fri Mar 04, 2016 8:44 pm

Re: A human representative against the machines

Post by Tuireann » Tue May 24, 2016 8:47 pm

namely, that each AI plays each other AI twice, once with each side. If you can win faster, you're probably better than your opponent.
I've always thought this sounded like a good way to approach Fetlar tournaments. Curious to see how it plays out.
Next year, if this year is successful, I may move to a different ruleset.
I hope it is. As a programmer if I had the time this year I would have entered. Unfortunately I'm working on too many other projects and trying to move out of country to throw an AI on top of it all. :(
As for OpenTafl's AI and its playing strength, you'll note I only said that 9x9 a little too easy to conquer—not that it's easy overall.
Ha yeah, twas a joke. I've been trying to keep my version up to date. Hoping for an AI to help hone my skills!
(The doors will also be open to anyone else who wants to give it a try.)
Let me know. I would be happy/eager to help test your AI and this method of time control.

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

Re: A human representative against the machines

Post by Fishbreath » Wed Jun 01, 2016 6:45 pm

I've posted a prerelease version of OpenTafl v0.3.0.0b, which has network play enabled. (The default server setting, intersect.manywords.press, is my personal webserver; there should be a server running there most of the time, except when it crashes.) There isn't really a way to tell if anyone else is around at present, which does strike me as a bit of a flaw, but I should be fixing that in the not-too-distant future. Probably best if you've organized a game ahead of time. <.<
Last edited by Fishbreath on Thu Jun 02, 2016 3:20 pm, edited 1 time in total.

Tuireann
Posts: 29
Joined: Fri Mar 04, 2016 8:44 pm

Re: A human representative against the machines

Post by Tuireann » Thu Jun 02, 2016 2:22 am

I see that you've implemented the time control. Anyone interested in testing the game with Fishbreath's style of time control?
More generally, I propose this mutation from a single main-time chess clock: for a given chess clock length in minutes (say, 60), divide by six to get the tafl main time (10, in this case). Divide that by ten to get the length of the overtime periods. Depending on the desired difficulty and clock sensitivity, set the number of overtimes to somewhere between two and six.
I'd also like to try something like 10 minutes of main time with 3 overtime periods of 5 minutes if anyone is willing. It might be interesting to try though the overtime periods may be a bit long.

Post Reply

Who is online

Users browsing this forum: No registered users and 2 guests