AlphaZero.jl is a generic implementation of DeepMind's algorithm written in Julia. It is consistently 20x to 100x faster than existing Python implementations while being equally simple and flexible. For the first time, it becomes possible for researchers to solve nontrivial problems with AlphaZero on any desktop computer with a GPU without leaving the assurance and comfort of a high-level language. AlphaZero.jl can be used on any MDP or zero-sum game with finite action space, on a personal computer as well as on a cluster of machines.
AlphaZero.jl is being developed in the context of the DARPA Assured Autonomy program, where it is used as a basis to develop tools for neural-guided proof search and symbolic planning.
Keywords: AlphaZero, Reinforcement Learning, Decision Making