AlphaZero.jl icon indicating copy to clipboard operation
AlphaZero.jl copied to clipboard

[Feedback Requested] Wishlist for the next major release of AlphaZero.jl

Open jonathan-laurent opened this issue 1 year ago • 4 comments

I've been thinking for a while about doing an overhaul of AlphaZero.jl and this summer may be the right time to execute on it via the GSOC program. I would really appreciate it if anyone can share feedback regarding the pain points of the current framework and what new features you wish were added in priority.

I have already written a short redesign note with some ideas. Feel free to comment on them also.

jonathan-laurent avatar Apr 13 '23 08:04 jonathan-laurent

Wow! The code is already quite easy to understand, and I'm excited to see it become even more accessible. The hyperparameter-related aspects appear to be very useful as well. Looking forward to it!

solidcub avatar Apr 13 '23 10:04 solidcub

Excited too to anticipate new progress for AlphaZero.jl! I definitively "vote" for the following (taken from the redesign note), which could certainly help with adoption:

Batteries included

  • Provide a test suite for new environments
  • Check hyperparameters consistency
  • Provide standard hyperparameter tuning utilities
  • Provide profiling utilities

smart-fr avatar Apr 15 '23 15:04 smart-fr

One question that's been asked over and over in issues (latest example) is how to scale-down hyperparameters to avoid OOM errors when using GPUs with limited memory. More generally, I believe that an auto-tuning script that adjusts hyperparameters based on current hardware resources would be welcome (getting such things exactly right is a research problem but there are definitely low-hanging fruits to be picked here).

jonathan-laurent avatar Apr 27 '23 10:04 jonathan-laurent

This would be awesome!

smart-fr avatar Apr 27 '23 10:04 smart-fr