- It only accepts four-player log.
- It only works with standard games with common rules, like those in the ranked lobbies.
- This public service no longer supports Mahjong Soul logs, but you can still review it locally. Check the instructions on how to use akochan-reviewer with mjsoul logs.
According to critter, the creator of akochan who himself has two 9 Dan accounts, akochan is 7~9 Dan.[ref]
According to some Houou players, some of akochan's decisions are at 8 Dan level.
According to myself who made the reviewer tool, akochan is at least stable 6 Dan in Tokujou.
We did tested akochan in Tenhou's Ippan lobby, and it reached Tokujou condition within 400 Ippan games. Here are the details: (nodocchi link)
Unfortunately we cannot go further because games above Joukyuu requires approvement from Tenhou for AI to play.
In a previous study of Kurita et al.[ref], the proposed AI, which was actually the prototype of akochan, had defeated the same version of ⓝ爆打 that had achieved its highest grade and ratings (R2206) in Tenhou, with 3557 games of mahjong in Tonpuu rule.
However, akochan is more of a traditional AI unlike NAGA and Suphx that use deep learning, so I personally doubt if akochan can defeat them.
Rumor has it that Tenhou doesn't favor AI developed by individual developers; it only cooperates with companies.
If you are talking about why akochan doesn't choose the tile with the lowest deal-in rate to discard when it is folding and has little chance of winning this round, then akochan does have the knowledge of 差し込み (sashikomi, intentional deal in) and アシスト (assist, intentionally let someone meld). By the way, it also knows 見逃し (minogashi, ignore ron) and 絞り (shibori, the opposite of assist).
Again, akochan takes the "final" pt/placement EV as its sole goal, so it simply makes decisions at any cost as long as it serves that goal.
You can, just contact me and perhaps ask 2 more friends in Tenhou.
No, that's another game.
It usually takes about 10~30 minutes to review one hanchan log and publish the result page. This is a public service, so please do not expect it to run that fast.
It is recommended to run the reviewer on your local machine. You can download prebuilt Windows binary here.
Result pages older than 14 days will be deleted on the server; save them if you need.
❤️ by Equim: GitHub | Twitter | 天鳳