jonathan-laurent / AlphaZero.jl

A generic, simple and fast implementation of Deepmind's AlphaZero algorithm.
https://jonathan-laurent.github.io/AlphaZero.jl/stable/
MIT License
1.24k stars 140 forks source link

Can I save time on benchmarks? #167

Closed smart-fr closed 1 year ago

smart-fr commented 1 year ago

It appears that for my game and a given params.jl, most of the time of a training session is spent running benchmarks. If my goal is to obtain a trained agent without precisely tracking its progress over iterations, can I run iterations with only self-play and loss optimization, skipping benchmarks altogether, without losing training quality?

jonathan-laurent commented 1 year ago

It is weird that most of the time is spent on benchmarks since the number of games played during benchmarks should be less than the number of games played for training.

Anyway, isn't it possible to just send an empty benchmark list?

In general though, removing then benchmark may be ill-advised. Indeed, getting AlphaZero to work on any game may require a bit of tuning and benchmarks are an essential tool to diagnose problems.

smart-fr commented 1 year ago

Thank you for your answer. I tweaked my parameters and found a better balance. I was going to close this issue with the comment "RTFM". 😅