Mahjax: A GPU-Accelerated Mahjong Simulator for Reinforcement Learning in JAX

2026-05-20 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Mahjax is a new GPU-accelerated Riichi Mahjong simulator implemented in JAX, designed to facilitate reinforcement learning research, particularly for tabula rasa learning. This fully vectorized environment enables large-scale rollout parallelization on Graphics Processing Units, addressing the challenges of multi-player, imperfect-information games characterized by stochasticity and high-dimensional state spaces. Mahjax aims to support algorithms capable of learning from scratch, mirroring the success of the AlphaZero lineage, rather than relying solely on supervised learning from human play logs. The simulator achieves high throughputs, reaching up to 2 million steps per second under no-red rules and 1 million steps per second under red rules, utilizing eight NVIDIA A100 GPUs. It also includes a high-quality visualization tool for debugging and agent interaction, and its utility has been validated by effectively training agents to improve their rank against baseline policies.

Key takeaway

For Machine Learning Engineers developing reinforcement learning agents for complex, imperfect-information games, Mahjax provides a critical tool. You should consider integrating this JAX-based, GPU-accelerated simulator to achieve high-throughput tabula rasa training. This enables faster iteration and exploration of learning algorithms without reliance on human play data, potentially accelerating your research into generalizable AI for challenging stochastic environments.

Key insights

Mahjax provides a GPU-accelerated JAX environment for tabula rasa reinforcement learning in complex imperfect-information games.

Principles

• Imperfect-information games challenge RL.
• Tabula rasa learning offers broad applicability.
• GPU parallelization accelerates complex simulations.

Method

The article describes implementing a fully vectorized Riichi Mahjong environment in JAX for GPU-accelerated, large-scale rollout parallelization. It includes a visualization tool.

In practice

• Train RL agents from scratch.
• Debug complex game AI efficiently.
• Benchmark RL algorithms on Mahjong.

Topics

• Riichi Mahjong
• Reinforcement Learning
• JAX Framework
• GPU Acceleration
• Imperfect Information Games
• Tabula Rasa Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Mahjax: A GPU-Accelerated Mahjong Simulator for Reinforcement Learning in JAX — Mahjax provides a high-throughput, GPU-accelerated JAX environment for *tabula rasa* reinforcement learning in complex imperfect-information games.
• daVinci-kernel: Co-Evolving Skill Selection, Summarization, and Utilization via RL for GPU Kernel Optimization — A reinforcement learning framework co-evolves skill selection, summarization, and utilization for GPU kernel optimization.
• LogNEO: A GPT-Neo Reinforcement Learning Framework for Accurate Real-Time Log Anomaly Detection — LogNEO leverages GPT-Neo and a novel position-aware RL reward scheme to achieve high-accuracy, real-time log anomaly detection.
• OS-Themis: A Scalable Critic Framework for Generalist GUI Rewards — OS-Themis is a scalable multi-agent critic framework that improves GUI agent RL training through milestone-based reward decomposition and strict evidence auditing.
• Convex Markov Games and Beyond: New Proof of Existence, Characterization and Learning Algorithms for Nash Equilibria — Nash equilibria in General Utility Markov Games coincide with fixed points of projected pseudo-gradient dynamics.
• Value Bonuses using Ensemble Errors for Exploration in Reinforcement Learning — VBE uses ensemble errors to create value bonuses for first-visit optimism and deep exploration in RL.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.