Arimaa

[hyperpape]

Who here plays? I'd heard of it before, but yesterday was the first time I played, and right now I feel addicted.

I looked at some high level games, and go to reexperience the feeling of a 30k looking at a professional game.

[daniel_the_smith]

Who here plays? I'd heard of it before, but yesterday was the first time I played, and right now I feel addicted.

I looked at some high level games, and go to reexperience the feeling of a 30k looking at a professional game.

I've played before. I think I got up to around 1500 or so on the rating system playing bots, but I don't know if I ever got as far as playing humans.

I even have a half-written arimaa bot sitting on my computer. Sigh, things I will never have time to complete...

[Mivo]

I played it for a while, a few years ago, but didn't get very far with it. I also didn't know anyone else who played and it was before they had a physical set available, so my interest in it kind of fizzled. Plus, my usual problem: I generally find it hard to get interested in any other strategy games ever since I learned (about) Go.

[hyperpape]

I just crossed 1400, the 1500 bots are still way too tough for me. And I've yet to play a human, which is just a wee bit ironic.

[Laman]

another beginner here, played 20 games in the last 2 weeks. it is nice to learn something completely new again, to suck so much that you improve just by playing and random messing around

but lack of human opponents is troublesome, i also played only one human so far

[Numsgil]

Looks interesting.

Are there any monte carlo bots? I'd imagine that would be the way to make the strongest computer player.

[daniel_the_smith]

Monte Carlo does pretty bad at Arimaa, actually! Random play moves the rabbits way too much. It's basically only our own game that MCTS has done really well at.

[EDIT: Last I heard, anyway. I don't know if anyone has tried the statistical learning techniques (other than RAVE, which IIRC did not work well for Arimaa) that work for go.]

[Mike Novack]

Looks interesting.

Are there any monte carlo bots? I'd imagine that would be the way to make the strongest computer player.

Not completely understanding the situation. The game "Arimaa" was devised with all currently known methods for computers to play games in mind with the intent that these be ineffective for Animaa.

Notice that before it was thought of and applied far from clear that MCTS would be the best currently known way for computers to play go. That was after many years of trying other things that weren't able to play a strong game. Well there may or may not be a way for computers to play a strong game of arimaa but hasn't been discovered yet.

[daniel_the_smith]

Yeah, Arimaa was designed before MCTS hit the big-time with go. MCTS just happens to not work well for Arimaa (just as it doesn't work well for chess, for much the same reasons).

[hyperpape]

It's not obvious to me that bot progress in Arimaa is worse than bot progress in Go. Of course it's complicated because both humans and bots are very new to the game, so both are making progress at the same time, and you don't have a fixed target to aim for.

And you can't imagine bots placing so highly on a list like this for Go (even given the disparity in number of players, imho) http://arimaa.com/arimaa/gameroom/topRatedPlayers.cgi

[daniel_the_smith]

I agree with that, too. Traditional chess engine techniques have worked fairly well in Arimaa, but not quite well enough to stay ahead of the humans. Arimaa is young and has few players, so it's actually a little surprising how well the humans have done!

[Numsgil]

Interesting. I'd have guessed Monte Carlo was the way to go.

Maybe Monte Carlo works well with go because the final board position isn't a binary win/fail, but a gradation of points for either side. Chess and Arimaa are both strictly win/lose (draw). I'm not sure why that'd matter, but that's the only thing I can think of that is really fundamentally different between the games. I'm sure there's some really interesting fundamental truth about Monte Carlo creeping in here somewhere, but I can't figure it out.

[daniel_the_smith]

One reason it works well in go (IIRC) is because random errors roughly cancel out. MCTS doesn't care about the margin of victory at the end, it aggregates binary win/lose events into a probability of winning.

[Numsgil]

One reason it works well in go (IIRC) is because random errors roughly cancel out.

That's more or less what I meant. Like, a random walk from a given even position proceeds to a slightly less even position, from which a random walk might walk back to the even position. At the end of the game there's a clear outcome but before then the game has this inherent gradation. Or to put it in terms of selection, the fitness landscape is smooth, with few steep gradients. Versus Chess and Arimaa where things could be going fine until they're not and you lose. So the fitness landscape has all these horrible cliffs and canyons everywhere.

I don't know if I actually believe that or not. Just an interesting thought.

[Mivo]

I actually wonder why the game isn't more popular. Thanks to this thread, I picked up the iOS Arimaa app and had a good time playing against the bot (last time I looked into Arimaa the only bot I found was David Fotland's). Is the pulling/pushing/freezing too complicated for people, or is it the stale presentation of the site? Or is it just hard for new board games without commercial backing or a very long history to take off?

It seems off that a game that is designed to be hard for computers has numerous bots in its list of the world's top 100 players.