search

tspooner / rsrl

A fast, safe and easy to use reinforcement learning framework in Rust.
https://crates.io/crates/rsrl
MIT License
178 stars 13 forks source link

Hyper parameter optimisation as a native feature #38

Open tspooner opened 6 years ago

Ploppz commented 5 years ago

What about making another crate for hyper parameter optimization?

tspooner commented 5 years ago

For sure, that's almost certainly what I'd do as with the function approximation (lfa), ML generics (spaces) and probability distributions (dists). It's really just a question of finding the time!

Ploppz commented 5 years ago

Alright. When I need it, I might make it as well, we'll see!

tspooner commented 5 years ago

Haha good times. Well let me know if you do and when the time comes I'll look into any potential integrations.

Cheers, Tom

Ploppz commented 5 years ago

I started something here: https://github.com/Ploppz/blackbox Look at the provided simple example. So far I have only implemented random search. What do you think about this approach? Suggestions are welcome!

tspooner commented 5 years ago

I'm a big fan of the direction you've taken already. Procedural macros is super cool feature in Rust which wasn't available when I started unfortunately, though I intend to incorporate these techniques in the framework (https://github.com/tspooner/rsrl/issues/42). That being said, I do also think it's important to make the interface simple to use if the user opts out of using macros. I would think that you could include more traits too in order to make it easier to extend and generalise, but I realise you're only 2 commits in!

I'll watch your project, and please keep me up to date on this project. If/when it gets to a stable release phase I'll definitely look into how we can incorporate some of the ideas and/or code etc.

Cheers, Tom

Ploppz commented 5 years ago

I agree with your suggestion, I think I will make it configurable without macros too.