Poker-AI.org

Is anybody who trying to make mahjong bot?

Mahjonglogic is the most popular a real money mahjong network. I monitored activity on bodog88. Mahjong has a lot of game variations. Kansai Riichi (3 max game (standard mahjong — 4 max)) the most popular game on highstakes ($128/$256 max win $768). Average gain of $100. The duration of the game 10 minutes. You have access to all actions of the opponent in the history hands. I don't know anything about the winrate.

I read the rules and I thought "It should be interesting to mathematicians" Lot of combinations, lot of options which tile to discard, a lot of work for counting outs.

Never played it before myself, but reading the rules on wiki there seem to be a lot of things to consider. I really wonder what the game tree would look like drawn out, and how big it would be. How would you abstract it if you wanted to solve it game theoretic?

In comparison, poker is a somewhat simple game, albeit with a large domain space.

I sometimes play it with friends, it is similar to poker, but a big difference is obviously, that the tiles you discard can be used for your opponents draws, like in rommé.

4 handed mahjong is probably extremelly hard to bot, HU might be possible, but I dont know how popular it is online.

Thread moved to "Miscellaneous" (quality on-topic posts) — hah!

ron2.jp is the most famous mahjong room. It is very popular among the Japanese. Probably all the uber-top-regs play here. ron2.jp client has the bot (autoplay function, if you have lost internet connection).

A few words describing the bot:


One website evaluate the level of the opponents on mahjonglogic as extremely low (it has no the date of publication).

Maybe ron2.jp bot can beat the games on mahjonglogic.

What's the betting structure? Just put up your stake at the start, and collect when you win? If that's the case, I wouldn't expect deceptive play to be as important as poker. Is that true? If so, the bot doesn't have to worry about its own image, so you can use expectimax.

you stake reserved at the start, and you win part of all reserved funds. for example, if you play at $128/$256, than $256 will be reserved.

I don't understand the rest of the post. Do you mean, than mahjong haven't a psychological component? I don't know.

http://mahjongtime.com/mahjong-japanese-rules.html

A lot of rules seem to arise out of chance.

Found these:
http://link.springer.com/chapter/10.100 ... 28643-1_41
http://books.google.com/books?id=h0YUEP_uSCMC
http://code.google.com/p/ma-jiang/
http://www.cheatengine.org/forum/viewto ... 3bfc070743

Section (d) hints at player psychology, and self-image.
http://cs229.stanford.edu/proj2009/Loh.pdf

Seems there has been some research into mahjong AI, but it's still in it's infancy.

Bots:
http://games.brothersoft.com/four-winds-mah-jong.html
http://boardgamegeek.com/thread/287236/ ... -real-game

yes, I've seen good reviews about the Four Winds Mah Jong, but you cann't select a specific hand and get the answer. I've seen other similar solutions. I found a calculator for "in one step to win" situations. Is it a long way from the bot's client to the calc?

Your opponents use your tiles for their draws and you use theirs. When you discard certain tiles, your opponents will think, that you must have a draw, that couldnt improve with that tile. So you can sometimes try to discard a tile, that might have improved your draw a little, hoping that someone will deal a similar tile into you.

I'm not saying you ignore the opponent image and self image completely. I'm just suggesting it might be less important than in poker. And by less important, I mean the computational effort required to manage image does not dominate the solution as it does in poker.

The game has several rounds. You have to mix up your strategy depending on the amount of points. If you're the leader by a wide margin you must choose quickly collect a hand strategy. There are some nuances. For example, players are rarely discard a medium tiles in the beginning of the round.

I wonder about that. I see your point, from a single-state perspective, but if you consider the dynamics of the game (i.e. that what each player holds is constantly changing), image might be more important and harder to predict than in poker. <-- That statement doesn't necessarily equate to 'more computational effort,' though.

I wonder what the range update in an opponent model would look like after each action? Are tiles passed face down?

Put it another way. A poker strategy that doesn't bluff or slowplay itself and believes the opponent never bluffs or slowplays is crap. I contend that a rummy/mahjong strategy that discards everything that it needs to and believes opponent is doing the same isn't a very bad strategy. But hey it's just speculation and I've never played mahjong and haven't played rummy since I was a child.