Loading field map
Game theory, as applied to mahjong, is the study of strategic decision-making under conditions of uncertainty and competition. Mahjong is a tile-based game typically played by four players, where each player builds a hand of tiles to achieve a winning pattern. The game involves both chance—through the random draw of tiles—and skill, as players must decide which tiles to keep, discard, or claim from others. Game theory in this context seeks to understand optimal play: what decisions maximize a player's expected chances of winning, given incomplete information about opponents' hands and the stochastic nature of the tile wall.
The central questions of this subfield are practical and mathematical. How should a player balance the pursuit of their own winning hand against the risk of helping opponents? When is it correct to abandon a promising hand to play defensively, discarding tiles that are unlikely to let an opponent win? How does the scoring system—which varies across regional variants such as Japanese riichi, Chinese classical, or American mahjong—alter optimal strategy? These questions are not merely theoretical; they underpin the competitive play of skilled mahjong players, who often internalize game-theoretic principles without formalizing them.
Unlike perfect-information games such as chess, mahjong hides each player's hand from the others. A player knows only their own tiles, the discards made by all players, and the tiles they have claimed from others. This creates a game of imperfect information, where every decision is made under uncertainty about the state of opponents' hands. Game theory in mahjong therefore borrows heavily from the branch of mathematics dealing with such games, including concepts like expected value, Bayesian updating, and minimax reasoning.
A foundational concept is the expected value of a decision. For any possible discard, a player can estimate the probability that it will complete an opponent's hand, based on the tiles already visible. Similarly, a player can estimate the probability that keeping a tile will improve their own hand. Optimal play, in the simplest model, involves choosing the action with the highest net expected value—the benefit to oneself minus the expected cost of helping an opponent. This calculation is complicated by the fact that opponents also make strategic choices, so probabilities must be updated as the game progresses.
Defensive play emerges naturally from this framework. When a player's own hand is far from winning, the expected value of pursuing it may be negative, because the risk of dealing into an opponent's hand outweighs the chance of winning. Skilled players learn to "fold" such hands, discarding tiles that are statistically safe—those that have already been discarded by others, or that are unlikely to be needed by any opponent. This defensive strategy is a direct application of minimax thinking: minimize the maximum possible loss, even if it means forgoing a chance to win.
The field is not organized into rigid schools, but several distinct approaches have developed, each addressing different aspects of the game.
The oldest and most fundamental approach treats mahjong as a problem of probability. Early analyses, often published in Japanese strategy books, calculated the odds of drawing specific tiles, the likelihood of completing various hand patterns, and the expected value of different discards. This tradition remains central, especially in riichi mahjong, where the scoring system rewards complex hands and the "riichi" declaration—a bet on an incomplete hand—adds a layer of risk and reward.
This approach is empirical as well as theoretical. With the advent of computer simulations and large databases of recorded games, researchers have been able to estimate probabilities more accurately than was possible by hand calculation. For example, the probability that a given tile is safe to discard against a particular opponent can be estimated from the frequency with which that tile appears in winning hands, given the visible discards. These statistical models are not perfect—they cannot account for all strategic subtleties—but they provide a solid baseline for decision-making.
The limits of this approach are clear: it treats opponents as statistically average, not as adaptive agents. A human opponent may deviate from the most likely hand pattern, either through error or through deliberate deception. Probabilistic analysis therefore provides a foundation, but not a complete solution.
A more formal approach applies the tools of game theory proper, seeking Nash equilibria—sets of strategies where no player can improve their expected outcome by unilaterally changing their play. This is mathematically challenging because mahjong has a vast state space: the number of possible tile configurations is astronomically large, and the game includes stochastic elements (the draw) and hidden information.
Work in this vein often simplifies the game to make it tractable. Researchers might analyze a reduced version with fewer tiles or a simplified scoring system, or they might focus on a single decision point, such as whether to declare riichi or whether to claim a tile from another player's discard. These analyses yield insights into the structure of optimal play, even if they cannot produce a complete strategy for the full game.
A key finding from this approach is that mahjong is not a game of pure skill, nor pure luck, but a game where skill manifests in the management of risk. The equilibrium strategies often involve a mix of aggressive and defensive play, depending on the state of the game and the scores of all players. However, because the game is finite and stochastic, no strategy can guarantee a win; the best a player can do is maximize their expected score over many games.
In practice, most competitive players do not calculate exact probabilities or solve for equilibria. Instead, they rely on heuristics—rules of thumb that approximate optimal play. These heuristics are often derived from the probabilistic and game-theoretic analyses, but they are simplified for use in real time.
Examples include: "discard the tile that is least likely to be needed by an opponent," "keep a hand flexible by avoiding pairs and sequences that lock in a single pattern," and "if your hand is far from winning, play defensively." These rules are not always correct, but they are effective in most situations. The development of such heuristics is itself a form of research, often conducted by strong players who test and refine their rules through thousands of games.
The relationship between heuristics and formal analysis is bidirectional. Formal analysis can validate or refute a heuristic, while heuristics can suggest new questions for formal study. For example, the common advice to "discard from the honors first" (the wind and dragon tiles) is supported by probability calculations showing that these tiles are less useful for building sequences, but the advice is nuanced by the fact that honors can be valuable in certain hand patterns.
The most recent development is the use of artificial intelligence and machine learning to analyze mahjong. Computer programs have been built that play mahjong at a high level, using techniques such as Monte Carlo tree search, neural networks, and reinforcement learning. These programs do not explicitly solve the game; rather, they learn effective strategies by playing millions of games against themselves or against human opponents.
These AI systems have several uses. They can serve as training partners for human players, providing feedback on mistakes. They can also be used to test the validity of human heuristics, by comparing the AI's choices to the rules of thumb. In some cases, AI has discovered strategies that human players had not articulated, such as subtle defensive moves that are not obvious from probabilistic analysis alone.
The limits of AI approaches are important to note. A neural network that plays well does not necessarily provide an explanation of why its moves are good; it is a black box. Moreover, the AI is trained on a specific variant of mahjong, and its strategies may not transfer to other variants with different rules or scoring. Nevertheless, AI has become a valuable tool in the subfield, both as a practical aid and as a source of new hypotheses about optimal play.
Mahjong is not a single game but a family of games with shared roots. The major variants—Japanese riichi, Chinese classical, Hong Kong, Taiwanese, and American—differ in rules, scoring, and the number of tiles used. These differences have a profound impact on game theory.
In riichi mahjong, the scoring system heavily rewards complex hands and the riichi declaration, which adds a point for a ready hand but also restricts the player from changing their hand. This creates a tension between aggression (pursuing a high-scoring hand) and safety (avoiding dealing into an opponent's hand). The game-theoretic analysis of riichi therefore emphasizes risk management and the value of information, since the riichi declaration reveals information about the player's hand to opponents.
In Chinese classical mahjong, the scoring is simpler, and the game is often played with a "limit" on the maximum score. This reduces the incentive for complex hands and makes the game more about speed—winning quickly with a modest hand rather than slowly building a high-scoring one. The optimal strategy in this variant is therefore more aggressive, with less emphasis on defense.
American mahjong, with its use of jokers and a fixed set of winning patterns, is a different game altogether. The presence of jokers changes the probabilities dramatically, and the strategy revolves around the efficient use of these wild tiles. Game-theoretic analysis of American mahjong is less developed than for riichi, partly because the variant is less studied academically.
The existence of these variants means that game theory in mahjong is not a single unified field but a set of related analyses, each tailored to a specific rule set. A strategy that is optimal in riichi may be suboptimal in Chinese classical, and vice versa. This diversity is a strength of the subfield, as it allows for comparative studies of how rule changes affect strategic behavior.
The present state of game theory in mahjong is characterized by a productive interplay between formal analysis, empirical study, and practical play. Probabilistic models provide the foundation, game-theoretic equilibrium analysis offers insights into the structure of optimal play, heuristics translate these insights into usable rules, and AI systems push the boundaries of what is known.
There is no consensus on a single "correct" approach, and the field is not divided into rival schools. Instead, researchers and players draw on multiple methods, depending on the question at hand. A player might use a heuristic in the heat of the game, a statistical model to evaluate a specific situation, and a game-theoretic framework to understand the long-term consequences of their style.
The central questions remain open. No one has produced a complete solution to any major variant of mahjong, in the sense of a strategy that is provably optimal against all possible opponents. The game's complexity—its large state space, hidden information, and stochastic elements—makes such a solution unlikely in the foreseeable future. What game theory offers instead is a set of tools for improving play, understanding the game's structure, and appreciating the depth of strategic choice that lies beneath the surface of a seemingly simple tile game.
For the educated newcomer, the most useful takeaway is that mahjong is a game of calculated risk. Every decision involves a trade-off between the chance of winning and the danger of helping an opponent. Game theory provides the language and the methods for making these trade-offs explicit, and the field's ongoing development—through probability, equilibrium analysis, heuristics, and AI—continues to refine our understanding of what it means to play well.