Introduction

GUI for https://mjai.ekyu.moe

Table of Contents

• Introduction
• Table of Contents
• Dealin Rate
  • Overview of the algorithm
  • Riichi trap suji
• Notes for translators/developers
• Credits

Dealin Rate

The Mortal engine does not have a dealin rate output, but in Options you can opt-in to showing dealin rates approximated by simple heuristics.

Overview of the algorithm

• Only enabled for players that Riichi
• Estimates the odds for each simple wait type
  • Ryanmen
  • Kanchan
  • Penchan
  • Shanpon
  • Tanki
  • (Complex waits are not explicitly accounted for)
• Likelyhood of each wait has several factors:
  • Number of combinations available to create that wait, based on unseen tiles from the POV of a player deciding on a discard
    • A general way account for e.g. kabe, one-chance, expensive vs cheap suits
    • For example, at first the combinations available for a 45m ryanmen are:
      • 4*4 = 16 combinations
    • But if there is only one 5m left and two 4m left
      • 1*2 = 2 combinations
  • Multiplier for how likely players are to end up with each type of wait

    Type	Multiplier	Notes
    Ryanmen	3.5	Players tend to aim for and Riichi ryanmens!
    Kanchan	0.21
    Kanchan*	2.6	Suji trap Riichi cutting e.g. 5m from 531m
    Penchan	1.0
    Honor Tanki/Shanpon	1.7
    Other Tanki/Shanpon	1.0

  • Multiplier for some simple discard reading and dora

    Type	Multiplier	Description
    Dora involved	1.2	Any wait involving the dora
    Ura suji	1.3	Cut 5m from 578m
    Matagi suji	0.6	Cut 3m early, probably not from 233m or 334m
    Riichi Matagi suji	1.2	Cut 3m late, could be from 233m or 334m
    Red 5 discard	0.14	Probably not cutting 0m from 455m

  • Multipliers were tuned for best predictions on Tenhou Houou games
    • Credit The Hopeless Girl on the Path of Houou for games and analysis framework code (which they got from ApplySci? And maybe other people?)
• Examples of things not accounted for
  • Yaku reading
  • Assumes never furiten Riichi
  • The tiles they are waiting for are mostly dead
    • e.g. Kan of 5p, will still assume they might Riichi a 46p Kanchan
  • Point situation indicates higher than normal odds they riichi on a bad wait

Riichi trap suji

Here is are some example dealin rates of a situation where they Riichi on 5m

                     Total  Ryankan Kanchan  Tanki Shanpon 
No-suji          2s:  9.5%   7.5%    0.3%    0.4%   1.2%
Riichi trap suji 2m:  6.0%    -      4.3%    0.4%   1.2%
Normal suji      2p:  2.0%    -      0.3%    0.4%   1.2%

The odds of a 13m Kanchan wait are much higher than the other 13 waits. The 2m is still safer than no-suji 2s, but not by much.

Notes for translators/developers

The GUI is simple html+javascript, but you still must run a local webserver to load the json file. Use e.g. vscode liveserver or python3 -m http.server.

You also need a local json file. Run a review on the mjai site, and then you can use the utility script curl_mjai.py to download the json locally:

./curl_mjai.py -u https://mjai.ekyu.moe/killerducky/?data=/report/9572f900340af356.json

It will output some URLs for reference, the last one is an example URL to point your browser to locally. Modify the local IP/port for your system.

Credits

• Equim-chan for creating Mortal and providing the online service for free!
• Tiles https://github.com/FluffyStuff/riichi-mahjong-tiles
• Tenhou log stuff from https://github.com/Equim-chan/tensoul/blob/main/convert.js
• Ukeire, Shanten, etc utilities from https://github.com/Euophrys/mahjong-discord-bot
• Translations originally from Tenhou English UI Chrome extension
• Compression from https://github.com/pieroxy/lz-string