Danger Estimation (Deal-in Risk)

Relevant source files

• internal/domain/ai/danger.go
• internal/domain/ai/danger_scene.go
• internal/domain/ai/danger_test.go
• internal/domain/ai/danger_tiles.go
• internal/domain/ai/deps.go

The DecisionTreeDangerEstimator is responsible for calculating the hojuProb (deal-in probability) for a specific discard against a target opponent. It utilizes a pre-trained decision tree that evaluates the "danger" of a tile based on game features such as Suji, Kabe, and discard history.

Overview

Danger estimation in mjai-manue-go is driven by a binary decision tree. For every legal discard candidate, the AI evaluates the risk of dealing into any opponent who has declared Riichi or is otherwise suspected of being Tenpai.

The estimation process follows these steps:

1. Anpai Short-circuit: If the tile is in the opponent's SafeTiles (Genbutsu), the probability is immediately returned as 0 internal/domain/ai/danger.go36-38
2. Scene Construction: A dangerScene is built, capturing the relevant context such as visible tiles, target's discards, and Riichi status internal/domain/ai/danger_scene.go27-60
3. Tree Traversal: The estimator traverses the DecisionNode tree. At each node, it evaluates a specific feature (e.g., "is this tile Suji?") against the dangerScene internal/domain/ai/danger.go43-64
4. Leaf Probability: Once a leaf node is reached, the stored probability is returned as the deal-in risk internal/domain/ai/danger.go46-48

System Architecture

The following diagram shows the relationship between the estimation logic and the data structures.

Danger Estimation Component Map

Sources: internal/domain/ai/danger.go11-30 internal/domain/ai/danger_scene.go14-25 internal/domain/ai/danger.go43-64

The Danger Scene (dangerScene)

The dangerScene struct acts as a snapshot of the game state from the perspective of a specific "danger" evaluation (Self vs. Target). It encapsulates all data required by the feature DSL internal/domain/ai/danger_scene.go14-25

Field	Description
selfHand	Current tiles in the AI's hand, including the drawn tile internal/domain/ai/danger_scene.go30-35
safeTiles	Tiles that are safe against the target (Genbutsu) internal/domain/ai/danger_scene.go50
visibleTiles	All tiles visible on the board (discards, melds, Dora indicators) internal/domain/ai/danger_scene.go51
doras	Current Dora indicators internal/domain/ai/danger_scene.go52
preRiichiTiles	Tiles discarded by the target before declaring Riichi internal/domain/ai/danger_scene.go42
riichiDeclarationTiles	The specific tile discarded to declare Riichi internal/domain/ai/danger_scene.go43

Sources: internal/domain/ai/danger_scene.go14-60

Feature DSL and Evaluation

The dangerScene.evaluate method implements a Domain Specific Language (DSL) used by the decision tree to query the state of a discard internal/domain/ai/danger_scene.go62-184

Core Features

• Suji/Kabe: suji, weak_suji, reach_suji, chances<=n internal/domain/ai/danger_scene.go87-125
• Discard Patterns: urasuji, matagisuji, senkisuji, aida4ken internal/domain/ai/danger_scene.go95-117
• Tile Metadata: tsupai (honors), dora, fanpai (Yaku-giving honors) internal/domain/ai/danger_scene.go66-75
• Visibility: visible>=n, suji_visible<=n internal/domain/ai/danger_scene.go132-139
• Hand Context: in_tehais>=n, suji_in_tehais>=n internal/domain/ai/danger_scene.go146-154
• Dynamic Suffixes: _outer_prereach_sutehai and _inner_prereach_sutehai for distance-based discard analysis internal/domain/ai/danger_scene.go158-171

Implementation Flow

The traversal logic in estimateDangerTreeProb determines the path through the tree:

Tree Traversal Logic

Sources: internal/domain/ai/danger.go43-64 internal/domain/ai/danger_scene.go62-118

Tile Mathematics

The estimator relies on several helper functions in danger_tiles.go to perform Mahjong-specific calculations:

• isSujiOf: Checks if the target tile is a Suji (e.g., 1 or 7 for a 4) relative to a set of tiles (usually safe tiles or Riichi discards) internal/domain/ai/danger_tiles.go63-78
• isNChanceOrLess: Implements "Kabe" (Wall) logic. It checks if the number of visible tiles that could form a specific sequence (Shuntsu) with the target is low enough (4-n) to make the sequence unlikely internal/domain/ai/danger_tiles.go92-114
• isAida4Ken: Checks for "Between 4-interval discards" (e.g., discarding 1 and 6, making 3 and 4 dangerous) internal/domain/ai/danger_tiles.go229-248
• fanpaiValue: Calculates how many Yaku points a honor tile is worth for the target (e.g., Guest Wind = 0, Round Wind = 1, Dragon = 1, Double Wind = 2) internal/domain/ai/danger_tiles.go32-48
• possibleSujis: Identifies potential Suji intervals that are not yet safe, used for calculating urasuji, matagisuji, and senkisuji internal/domain/ai/danger_tiles.go116-138

Data Loading and Integration

The decision tree is stored as a JSON file and embedded into the binary.

1. Loading: The DangerTreeNode interface allows the estimator to be decoupled from the specific JSON implementation internal/domain/ai/danger.go11-16
2. Injection: The DecisionTreeDangerEstimator is initialized with a root node that satisfies this interface internal/domain/ai/danger.go22-24
3. Agent Dependency: The estimator is passed to the AI via ManueAgentDeps internal/domain/ai/deps.go10-13

Short-circuiting and Normalization

• Red Five Normalization: Before estimation, discard.RemoveRed() is called. The danger model treats red and black fives identically for deal-in probability internal/domain/ai/danger.go35
• Anpai Check: If state.SafeTiles(winner).ContainsSameSymbol(discard) is true, the function returns 0 immediately, bypassing the tree traversal internal/domain/ai/danger.go36-38

Sources:

• internal/domain/ai/danger.go
• internal/domain/ai/danger_scene.go
• internal/domain/ai/danger_tiles.go
• internal/domain/ai/deps.go
• internal/domain/ai/danger_test.go