whereismyai
Hangzhou Mahjong
Hangzhou Mahjong is a Hangzhou Mahjong training and evaluation framework with public-information Transformers and paired branch reinforcement learning.
This project provides a Hangzhou Mahjong agent, its training and evaluation framework, and results from a frozen internal opponent pool. The 52.01-million-parameter policy combines a tile-relation Transformer, a causal event encoder, legal-action queries, public tactical successor features, and compact opponent memory. The reported checkpoint inherits an already trained policy, receives supervised action fine-tuning, and continues with paired branch reinforcement learning. Each eligible decision on a reference trajectory produces five exploratory continuations, and their realized scores are compared with a reference. Evaluation rotates complete tables through all four seats and reports a fixed table-length-weighted net score. The repository contains the training framework, game simulator, evaluator, and the selected r2600 weights.
Hangzhou Mahjong at a glance
Key features
- Tile-relation Transformer policy
- Causal event encoder
- Legal-action queries
- Public tactical successor features
- Compact opponent memory
- Paired short-horizon exploration
- Game simulator and evaluator
Use it for
- Run the released Hangzhou Mahjong policy
- Evaluate user-supplied model checkpoints
- Simulate four-player local Hangzhou Mahjong games
- Execute training and evaluation frameworks
Alternatives to Hangzhou Mahjong
- Catan Advisor — Measures winning strategies across thousands of simulated games.
- Kungfu — Maintains work state and context when switching between different AI agents.
- ART — Enhances LLM agent performance and reliability using reinforcement learning techniques.
- Beeuni — Agents retain context and improve skills over time through a controlled review process.
- Pub Street Fighter AI — It trains a convolutional neural network on all 12 arcade levels using RGB game pixels.
- OmniAgent — It evolves skills in real-time and dynamically hardens security during task execution.
- MuscleMemory — It cuts input tokens, tool calls, and execution time while maintaining task success.
- Klavis — Provides live environments for training agents with verifiable reward data.
- Harness2 — It recursively refines agent execution structures using parallel and sequential search.
- Rax — Exploration compiles into code that runs up to 241 times faster.
- Reflexio — It autonomously extracts lessons from interactions to permanently correct agent behavior.
- AgentGarten — Runs executable code worlds with real-time neural rendering for agents.
All Hangzhou Mahjong alternatives