882 points by chessmind 5 months ago flag hide 10 comments
john_carmack 5 months ago next
This is really impressive! I've been watching chess AI for years, but never tried creating one. Any tips on kicking off such a project?
jane_doe 5 months ago next
John, I'd start off with reading the AlphaZero paper and trying to understand the general architecture. From there, you can break it down into smaller tasks, like implementing a neural network and learning about tree search algorithms.
cooltechie 5 months ago next
@jane_doe, I plan to try building this over the weekend! I really enjoy the chess, so glad to find this thread! Thanks!
georgehotz 5 months ago prev next
I built a similar chess AI with my team last year. Dissecting AlphaZero's architecture is a massive undertaking, but doable. I'd recommend following the paper and studying tree search algorithms.
coder_super 5 months ago next
@georgehotz any open source repos with your implementation? Would be cool to see how y'alli approached the problem.
georgehotz 5 months ago next
I'm afraid not :( We couldn't open-source our implementation for legal reasons. We did write several blog posts highlighting a few key techniques. I'll share a link in a separate comment.
codingpractice101 5 months ago prev next
Really fascinating topic, George! I studied AlphaZero paper last year, but created a TF implementation of AlphaGo Zero. I'd be glad to share code snippets/ideas for chess AI noobs like me!
newbieai 5 months ago next
Hi @codingpractice101, I'm eager to see your AG Zero implementation. Have you tried moving that to chess? Or should I start from scratch?
curiousprogrammer 5 months ago prev next
I've been doing deep dive into deep learning and RL lately. Seeing a real-life example definitely motivates me to focus on a chess AI as a pet project. Thanks for the inspiration!
excellentcoder 5 months ago next
Happy to see you motivated, @curiousprogrammer. I've got a bunch of resources for both ML and chess that might be helpful.