ELF OpenGo is an AI bot from Facebook AI Research (FAIR) that has defeated world champion professional Go players. Both the trained model and code used to create ELF OpenGo are available to the community to inspire others to think about new applications and research directions for this technology.
ELF OpenGo is a reimplementation of AlphaGoZero / AlphaZero that was trained on 2,000 GPUs over a 15 day period, and has achieved super-human performance. With only a single GPU, the ELF OpenGo bot played with four top 30 professional players and won 20-0 in slow games that impose no constraints on time spent for human players.
Using ELF OpenGo, we analyzed historical professional games from the past few centuries, and developed interactive tools to visualize the results.
ELF OpenGo public binary, allowing anyone with a CUDA-enabled GPU to play against the final ELF OpenGo model.
We have open sourced the code used to train ELF OpenGo and released the models and data.
Install Caffe2 with CUDA support. If you already have Caffe2 installed, make sure to update it to a version that includes the Detectron module.
Install Python dependences and the COCO API.
Clone the Detectron repository and set up Python modules.
Run inference using pretrained Detectron models, or install datasets and follow advanced directions to train on your own data.
OpenGo: An Analysis and Open Reimplementation of AlphaZero, Yuandong Tian, Jerry Ma, Qucheng Gong, Shubho Sengupta, Zhuoyuan Chen, James Pinkerton, C. Lawrence Zitnick, arXiv 2019.