If you’ve been fascinated with DeepMind’s AlphaGo program, there’s good news for you. A few Go enthusiasts have replicated the results of the AlphaGo Zero paper, using a few resources provided by Google.
The developers are keen to stress that this project is in no way associated with the official AlphaGo program by DeepMind. It’s an independent effort that is inspired by AlphaGo, just not affiliated to it.
According to the developers, Minigo “is a pure Python implementation of a neural-network based Go AI, using TensorFlow”. It adds a few features and architecture changes to the “Mastering the game of Go without human knowledge” paper. Very recently, this architecture was extended further in the “Mastering Chess and Shogi by self-play with a general reinforcement learning algorithm” paper.
The goals of this project, as described by the authors, are listed below:
- Provide a clear set of learning examples using Tensorflow, Kubernetes, and Google Cloud Platform for establishing Reinforcement Learning pipelines on various hardware accelerators.
- Reproduce the methods of the original DeepMind AlphaGo papers as faithfully as possible, through an open-source implementation and open-source pipeline tools.
- Provide our data, results, and discoveries in the open to benefit the Go, machine learning, and Kubernetes communities.
If you’re interested in doing this on your machine, you can access Minigo’s source code, and other resources, here.
Just a note here that you will need the following before you can get started:
Our take on this
The developers repeatedly mention that this is not a DeepMind project and is explicitly not meant to compete with AlphaGo. They just wanted other developers in the community to understand (and perhaps replicate or improve) how the Go model works. It’s definitely something to keep your eye on as more progress is made in this study.