Custom Rules & Variants

This article introduces the various Go rules supported by KataGo, board size variants, and how to customize rule settings.

KataGo supports many rulesets, including Chinese, Japanese, Korean, AGA, New Zealand, and Tromp-Taylor. In config.cfg you can customize the scoring method, komi, and board dimensions through parameters such as rules, komi, and boardXSize/boardYSize.

---

Ruleset Overview

Main Rules Comparison

Ruleset	Scoring Method	Komi	Suicide	Ko Repetition
Chinese	Area scoring	7.5	Forbidden	Forbidden
Japanese	Territory scoring	6.5	Forbidden	Forbidden
Korean	Territory scoring	6.5	Forbidden	Forbidden
AGA	Hybrid	7.5	Forbidden	Forbidden
New Zealand	Area scoring	7	Allowed	Forbidden
Tromp-Taylor	Area scoring	7.5	Allowed	Forbidden

KataGo Settings

# config.cfg
rules = chinese           # Ruleset
komi = 7.5               # Komi
boardXSize = 19          # Board width
boardYSize = 19          # Board height

---

Chinese Rules

Features

Scoring method: Area scoring (stones + territory)
Komi: 7.5 points
Suicide: Forbidden
Ko repetition: Forbidden (simple rule)

Area Scoring Explanation

Final score = Own stones + Own empty points

Example:
Black stones 120 + Black territory 65 = 185 points
White stones 100 + White territory 75 + Komi 7.5 = 182.5 points
Black wins by 2.5 points

KataGo Settings

rules = chinese
komi = 7.5

---

Japanese Rules

Features

Scoring method: Territory scoring (only empty points)
Komi: 6.5 points
Suicide: Forbidden
Ko repetition: Forbidden
Requires marking dead stones

Territory Scoring Explanation

Final score = Own empty points + Captured opponent stones

Example:
Black territory 65 + Captures 10 = 75 points
White territory 75 + Captures 5 + Komi 6.5 = 86.5 points
White wins by 11.5 points

Dead Stone Determination

Japanese rules require both players to agree on which stones are dead:

def is_dead_by_japanese_rules(group, game_state):
    """Determine dead stones under Japanese rules"""
    # Need to prove the group cannot make two eyes
    # This is the complexity of Japanese rules
    pass

KataGo Settings

rules = japanese
komi = 6.5

---

AGA Rules

Features

American Go Association (AGA) rules combine advantages of Chinese and Japanese rules:

Scoring method: Hybrid (area or territory, same result)
Komi: 7.5 points
Suicide: Forbidden
White must fill one stone to pass

Pass Rules

Black passes: No stone needed
White passes: Must give one stone to Black

This makes area and territory scoring results identical

KataGo Settings

rules = aga
komi = 7.5

---

Tromp-Taylor Rules

Features

The simplest Go rules, suitable for programming:

Scoring method: Area scoring
Komi: 7.5 points
Suicide: Allowed
Ko repetition: Super Ko (forbid any repeated position)
No dead stone determination needed

Super Ko

def is_superko_violation(new_state, history):
    """Check if Super Ko is violated"""
    for past_state in history:
        if new_state == past_state:
            return True
    return False

End Game Determination

No agreement on dead stones needed
Game continues until:
1. Both players pass consecutively
2. Then use search or actual play to determine territory

KataGo Settings

rules = tromp-taylor
komi = 7.5

---

Board Size Variants

Supported Sizes

KataGo supports multiple board sizes:

Size	Features	Recommended Use
9×9	~81 points	Beginners, quick games
13×13	~169 points	Advanced learning
19×19	361 points	Standard competition
Custom	Any	Research, testing

Configuration

# 9×9 board
boardXSize = 9
boardYSize = 9
komi = 5.5

# 13×13 board
boardXSize = 13
boardYSize = 13
komi = 6.5

# Non-square board
boardXSize = 19
boardYSize = 9

Komi Recommendations

Size	Chinese Rules	Japanese Rules
9×9	5.5	5.5
13×13	6.5	6.5
19×19	7.5	6.5

---

Handicap Settings

Handicap Games

Handicap is a way to adjust for strength differences:

# 2 stone handicap
handicap = 2

# 9 stone handicap
handicap = 9

Handicap Positions

HANDICAP_POSITIONS = {
    2: [(3, 15), (15, 3)],
    3: [(3, 15), (15, 3), (15, 15)],
    4: [(3, 15), (15, 3), (3, 3), (15, 15)],
    # 5-9 stones use star points + tengen
}

Komi with Handicap

# Traditional: No komi or half point with handicap
komi = 0.5

# Modern: Adjust based on handicap stones
# Each stone is worth ~10-15 points

---

Analysis Mode Rule Settings

GTP Commands

# Set rules
kata-set-rules chinese

# Set komi
komi 7.5

# Set board size
boardsize 19

Analysis API

{
  "id": "query1",
  "moves": [["B", "Q4"], ["W", "D4"]],
  "rules": "chinese",
  "komi": 7.5,
  "boardXSize": 19,
  "boardYSize": 19,
  "overrideSettings": {
    "maxVisits": 1000
  }
}

---

Advanced Rule Options

Suicide Settings

# Forbid suicide (default)
allowSuicide = false

# Allow suicide (Tromp-Taylor style)
allowSuicide = true

Ko Rules

# Simple Ko (only forbid immediate recapture)
koRule = SIMPLE

# Positional Super Ko (forbid any repeated position, regardless of player)
koRule = POSITIONAL

# Situational Super Ko (forbid repeated positions by same player)
koRule = SITUATIONAL

Scoring Rules

# Area scoring (Chinese, AGA)
scoringRule = AREA

# Territory scoring (Japanese, Korean)
scoringRule = TERRITORY

Tax Rules

Some rules have special scoring for seki areas:

# No tax
taxRule = NONE

# Seki has no points
taxRule = SEKI

# All eyes have no points
taxRule = ALL

---

Multi-Rule Training

KataGo's Advantage

KataGo uses a single model supporting multiple rules:

def encode_rules(rules):
    """Encode rules as neural network input"""
    features = np.zeros(RULE_FEATURE_SIZE)

    # Scoring method
    features[0] = 1.0 if rules.scoring == 'area' else 0.0

    # Suicide
    features[1] = 1.0 if rules.allow_suicide else 0.0

    # Ko rule
    features[2:5] = encode_ko_rule(rules.ko)

    # Komi (normalized)
    features[5] = rules.komi / 15.0

    return features

Rule-Aware Input

Neural network input contains:
- Board state (19×19×N)
- Rule feature vector (K dimensions)

This lets the same model understand different rules

---

Rule Switching Examples

Python Code

from katago import KataGo

engine = KataGo(model_path="kata.bin.gz")

# Chinese rules analysis
result_cn = engine.analyze(
    moves=[("B", "Q4"), ("W", "D4")],
    rules="chinese",
    komi=7.5
)

# Japanese rules analysis (same position)
result_jp = engine.analyze(
    moves=[("B", "Q4"), ("W", "D4")],
    rules="japanese",
    komi=6.5
)

# Compare differences
print(f"Chinese rules Black win rate: {result_cn['winrate']:.1%}")
print(f"Japanese rules Black win rate: {result_jp['winrate']:.1%}")

Rule Impact Analysis

def compare_rules_impact(position, rules_list):
    """Compare impact of different rules on position evaluation"""
    results = {}

    for rules in rules_list:
        analysis = engine.analyze(
            moves=position,
            rules=rules,
            komi=get_default_komi(rules)
        )
        results[rules] = {
            'winrate': analysis['winrate'],
            'score': analysis['scoreLead'],
            'best_move': analysis['moveInfos'][0]['move']
        }

    return results

---

Common Questions

Different Rules Leading to Different Results

Same game, different rules may lead to different results:
- Scoring differences between area and territory
- Handling of seki areas
- Impact of passes

Which Rules to Choose?

Scenario	Recommended Rules
Beginners	Chinese (intuitive, no disputes)
Online games	Platform default (usually Chinese)
Nihon Ki-in	Japanese
Programming	Tromp-Taylor (simplest)
Chinese pro tournaments	Chinese

Do Models Need Rule-Specific Training?

KataGo's multi-rule model is already strong. But if using only one ruleset, consider:

# Fixed rules training (may slightly improve strength for specific rules)
rules = chinese

---

Further Reading

• KataGo Training Mechanism — Multi-rule training implementation
• Integrate into Your Project — API usage examples
• Evaluation & Benchmarking — Strength testing under different rules