Mahjong Engines
University of Tokyo · Riichi

Bakuuchi 爆打

The strongest mahjong AI before deep learning took over. Bakuuchi predicts what opponents are waiting on, then simulates the rest of the hand to pick its discard.

Paper (PDF) ↗
Built by
Naoki Mizukami, University of Tokyo 13
Rules
Four-player Japanese riichi 1
Approach
Monte Carlo simulation + opponent models 1
Strength
Tenhou 9 dan peak, stable rank 6.59 2
Access
Research only

Who built it

Bakuuchi (爆打) was built by Naoki Mizukami while a graduate student at the University of Tokyo 3. The method was published with Yoshimasa Tsuruoka at the 2015 IEEE Conference on Computational Intelligence and Games (CIG) 12.

Rules

Bakuuchi plays four-player Japanese riichi on Tenhou 13.

How it works

Bakuuchi splits an opponent’s hidden state into three things and trains a model for each from expert game records 1:

  • Waiting. Is this opponent one tile from winning? (logistic regression, 6,888 features)
  • Winning tiles. Which tiles would they win on? (34 logistic regressions, one per tile type)
  • Score. How much would their hand be worth? (linear regression on the log of the score)

The training data was games from Tenhou’s Houou table, open only to the top 0.1% of players, played between February 2009 and December 2013 1. To choose a discard, Bakuuchi simulates the rest of the round for each candidate tile, with opponents acting on the probabilities from those models, and subtracts the expected loss from dealing into an opponent 1. A separate one-step expected-value check decides when to fold. Early in a hand, when simulation helps least, it falls back to a supervised model of hand-building moves 1. It does not use reinforcement learning 2.

Strength

In the 2015 paper, Bakuuchi’s method reached a Tenhou rating of 1718 over about 2,600 games, roughly the average player at the Joukyuu table, the third of Tenhou’s four tables 1. By February 2018 it was 7 dan, had reached 9 dan, and was playing about 50 hanchan a day 3. The Suphx paper later recorded Bakuuchi at 9 dan with a stable rank of 6.59 over 30,516 games, among the two strongest AIs before Suphx 2.

How to access it

We found no released code, model or playable service. The CIG 2015 paper 1 describes the method in enough detail to reimplement.

Sources

  1. Mizukami and Tsuruoka, “Building a Computer Mahjong Player Based on Monte Carlo Simulation and Opponent Models”, IEEE CIG 2015, pp. 275–283 (PDF)
  2. Li et al., “Suphx: Mastering Mahjong with Deep Reinforcement Learning”, arXiv:2003.13590 (2020)
  3. ITmedia NEWS, interview with Naoki Mizukami (23 February 2018, Japanese)

Facts checked against these sources on 24 October 2026.