About
NAGA (Neural Architectural Game Agent) is a research-grade Mahjong AI developed by Dwango Media Village (DMV) using deep learning techniques. Rather than relying on traditional rule-based heuristics or Monte Carlo simulations, NAGA trains four specialized convolutional neural network (CNN) models — for tile discarding, calling, riichi, and kan — entirely from game records of Tenhou's highest-ranked human players. Mahjong presents a significant AI challenge: it is a four-player, incomplete-information game where opponents' tiles are hidden, future draws are uncertain, and long-term strategy must span multiple rounds. NAGA addresses these challenges by learning offensive and defensive behaviors implicitly from human gameplay rather than explicitly programming strategic logic. Since its debut on Tenhou in October 2018, NAGA rapidly climbed from the general table to Eight-Dan, the platform's highest rank. Its tile selection mechanism notably handles folding — a notoriously difficult judgment for traditional AIs — by learning safe tile selection and game-state awareness directly from training data. DMV provides an interactive web demo where users can explore NAGA's confidence levels during a real game, inspect the distribution of tile-discarding reasons, and understand its decision logic at each turn. This makes NAGA a valuable resource for AI researchers, game AI enthusiasts, students studying reinforcement and supervised learning, and developers interested in applying deep learning to complex strategic games.
Key Features
- Four Specialized CNN Models: Separate convolutional neural network models handle tile discarding, calling, riichi declaration, and kan decisions, each trained on high-level human gameplay data.
- Eight-Dan Rank on Tenhou: NAGA achieved Tenhou's highest rank by competing against human players in real-time, validating its strength in a highly competitive environment.
- Learned Folding Judgment: Unlike traditional AIs, NAGA implicitly learns when to fold or play defensively by studying human game records, without explicit rule-based programming.
- Interactive Decision Visualization Demo: A web-based demo lets users inspect NAGA's confidence levels and tile-discarding rationale at any point during a sample game.
- Long-Term Strategy Awareness: NAGA reflects multi-round strategic thinking — such as avoiding a win that would cause last place — learned directly from human behavior across full East and half-game sessions.
Use Cases
- Academic research into deep learning for incomplete-information multi-agent games
- Benchmarking AI decision-making against expert human players in competitive Mahjong
- Studying how CNNs can implicitly learn complex strategic behaviors like folding and long-term planning
- Educating students and developers on applying supervised learning to game AI without hand-crafted heuristics
- Exploring interactive AI decision visualization for game state analysis and explainability research
Pros
- Proven Competitive Performance: Reaching Eight-Dan on one of Japan's most rigorous Mahjong platforms demonstrates real-world strength against skilled human opponents.
- Research Transparency: DMV publishes methodology, model architecture details, and an interactive demo, making NAGA highly valuable for academic study and reproducibility.
- Implicit Strategy Learning: By training purely on human data, NAGA captures subtle nuances like folding and long-term positioning without hand-crafted rules.
Cons
- Not a Consumer Product: NAGA is a research project and demo — it is not packaged as an app or API that end users can integrate into their own projects or play against directly.
- Mahjong-Specific Scope: The models and insights are narrowly focused on Japanese Mahjong (Riichi), limiting transferability to other games or domains without significant retraining.
- Partial Language Barrier: Much of the surrounding documentation, tweets, and Tenhou platform resources are in Japanese, which may limit accessibility for non-Japanese researchers.
Frequently Asked Questions
What is NAGA?
NAGA (Neural Architectural Game Agent) is a deep learning Mahjong AI developed by Dwango Media Village. It uses convolutional neural networks trained on high-ranking human game records to make tile selection and call decisions.
How does NAGA decide which tile to discard?
NAGA uses a CNN-based policy model that takes the full game state as input — including the player's hand, all four players' discards, calls, scores, and unseen tile counts — and outputs a probability distribution over possible tile discards.
What rank did NAGA achieve on Tenhou?
NAGA achieved Eight-Dan, the highest rank on Tenhou, one of Japan's largest online Mahjong platforms, after starting from the general table in October 2018.
How was NAGA trained?
NAGA was trained using supervised learning on game records from Tenhou's highest-level tables. Each of its four models (tile discarding, calling, riichi, kan) learned from the actions taken by top-ranked human players in those records.
Can I try NAGA or play against it?
DMV provides an interactive web demo on their research article page where you can explore NAGA's confidence levels and tile-selection decisions throughout a sample game. Direct play against NAGA is not publicly available through this interface.
