In fact, the core part of DeepMind's go AI 'AlphaGo' and the evolved version 'AlphaZero' has been quietly open sourced



Developed by AI development company DeepMind, the core part of Go AI ' AlphaGo ', which became a hot topic by defeating a top human Go player , and ' AlphaZero ', an evolved AI that can learn all board games. , In fact, it is pointed out that it is quietly open sourced on GitHub.



NVIDIA's AI researcher Jim Fan claims that 'the core part of AlphaGo and AlphaZero' is ' Mctx ', a library with a JAX native implementation published by DeepMind on GitHub. Mctx provides a JAX-native Monte Carlo tree search that runs in parallel and at lightning speed on the input, allowing the algorithm to operate on a parameterized learning environment model, taking full advantage of the accelerator.

GitHub - deepmind/mctx: Monte Carlo tree search in JAX
https://github.com/deepmind/mctx



Monte Carlo tree search is a heuristic algorithm that constructs a search tree based on random simulation results and produces an answer that is close to the correct answer to some extent in a short period of time.

Monte Carlo tree search is useful for deriving the best move in turn-based games such as Go and chess, Fan said, ``This is definitely the most complicated component of AlphaGo, and it is even more important to make it efficient. ' said.



DeepMind's Mctx library supports not only AlphaGo, but also AlphaZero, which supports board games other than Go such as shogi and chess, and MuZero , which allows you to get a high score in games such as 'Pac-Man' even if you don't know the rules. It is said that



Related Posts:

in Software, Posted by log1h_ik