undefined | Better HN

0 pointsamitport5y ago0 comments

Yes, but AI search algorithms like MCTS are slow in general. Even with C, it will be very slow when you want the AI to be smart and consider many actions. IMO, you should train using python libs like [1] and move it to the browser with something like [2] OR run AI in the server. YES definitely writing an AI lib for the browser is a great goal, and as a programmer, it is super interesting. Still, it is tough, and time-to-market is much more crucial, as the entire codebase will change once it will interact with actual people.

[1] https://github.com/datamllab/rlcard

[2] https://onnx.ai/

0 comments

skybrian5y ago

Other than solved games like tic-tac-toe, game-playing bots can always use more performance because if you can search more efficiently, your bot gets smarter. Sometimes supposedly more sophisticated algorithms end up making things worse because they slow down the search.

It's an unusual area where functionality (what answer you get) and performance can't be separated. This is still true on the server.

amitportOP5y ago

To be clear, I'm not talking about the server implementation only about implementation on the browser.

And of course, Rust and C are faster than javascript, and it will be noticeable on search algorithms, but IMO, it will not cause user loss. It is smarter to get to the stage you have those users as fast as you can to validate this need.

So I would not go into implementing new AI libs in rust as part of a turn-based game engine just for the performance gain--mainly because it is a tougher project than building a turn-based game engine!

I would search for existing, viable solutions. As I said before, this includes server AI and server trained models running on something like tensorflow.js. Additionally, I will also try to research browser libs that may use compiled webassembly and webGL. In any case, I think it's not wise to build your own for this purpose.

j / k navigate · click thread line to collapse