Beyond its much publicized success in attaining superhuman level at games such as Chess and Go, DeepMind's AlphaZero algorithm illustrates a more general methodology of combining learning and search to explore large combinatorial spaces effectively. We believe that this methodology can have exciting applications in many different research areas. Because AlphaZero is resource-hungry, successful open-source implementations (such as Leela Zero) are written in low-level languages (such as C++) and optimized for highly distributed computing environments. This makes them hardly accessible for students, researchers and hackers. Many simple Python implementations can be found on Github, but none of them is able to beat a reasonable baseline on games such as Othello or Connect Four. As an illustration, the benchmark in the README of the most popular of them only features a random baseline, along with a greedy baseline that does not appear to be significantly stronger.
Features
- The core algorithm is only 2,000 lines of pure, hackable Julia code
- Generic interfaces make it easy to add support for new games or new learning frameworks
- Between one and two orders of magnitude faster than its Python alternatives
- This implementation enables solving nontrivial games on a standard desktop computer with a GPU
- The same agent can be trained on a cluster of machines as easily as on a single computer and without modifying a single line of code
- Asynchronous simulation mechanism that enables batching requests to the neural network across several simulation threads
Project Samples
