MAPPO
We developed a light-weight, well-tuned and super-fast multi-agent PPO library, MAPPO, for academic use cases.
MAPPO achieves strong performances (SOTA or close-to-SOTA) on a collection of cooperative multi-agent benchmarks, including particle-world (MPE), Hanabi, StarCraft Multi-Agent Challenge (SMAC) and Google Football Research (GFR).
Check out our paper and BAIR blog for the most critical implementation factors.
MAPPO achieves strong performances (SOTA or close-to-SOTA) on a collection of cooperative multi-agent benchmarks, including particle-world (MPE), Hanabi, StarCraft Multi-Agent Challenge (SMAC) and Google Football Research (GFR).
Check out our paper and BAIR blog for the most critical implementation factors.
Multi-Agent Hide-and-Seek
|
We show that by enforcing simple hide-and-seek rules, deep reinforcement learning agents in an open physical world can learn extremely complex, human-interpretable strategies and counter-strategies.
These results are obtained via multi-agent self-play (paper). Check out our BLOG post with an extremely popular video demo. The environment code and trained policy can be found here. |
|
House3D Environment
House3D is an interactive environment specially designed for evaluating the generalization ability of a deep reinforcement learning agent. It is built on top of the SUNCG dataset, which contains 45K human designed 3D models of houses.
House3D is written in C++ and provides Python APIs for a variety of tasks.
Here is the Github Repo and the arxiv Paper.
MIT Tech Review article and some other blog post and article about House3D.
House3D is written in C++ and provides Python APIs for a variety of tasks.
Here is the Github Repo and the arxiv Paper.
MIT Tech Review article and some other blog post and article about House3D.
Compiler of Bayesian Logic (BLOG)
BLOG is one of the existing probabilistic modeling languages designed for making inferences about real-world objects that underlie some observed data.
Figaro Translator
A compiler translate a BLOG program (interpreted by possible world semantics) to a Scala program using an extended package Figaro (interpreted by random execution semantics)
Here is the Workshop Paper at NIPS 2014 about the translator.
C++ Compiler (Swift)
a C++ compiler that can compile a BLOG program to a C++ program. This project aims to accelerate the inference speed of the old Java inference engine of BLOG. This project is still underdevelopment. Till now, the running time is over 100 times faster.
Check the Github Repo and the Paper at IJCAI 2016 about the compiler.
A compiler translate a BLOG program (interpreted by possible world semantics) to a Scala program using an extended package Figaro (interpreted by random execution semantics)
Here is the Workshop Paper at NIPS 2014 about the translator.
C++ Compiler (Swift)
a C++ compiler that can compile a BLOG program to a C++ program. This project aims to accelerate the inference speed of the old Java inference engine of BLOG. This project is still underdevelopment. Till now, the running time is over 100 times faster.
Check the Github Repo and the Paper at IJCAI 2016 about the compiler.
