check out my new ML Agent written completly from pytorch to c-sharp

https://github.com/Isawyou/Reinforcement-Learning-c-sharp-Unity-ppo-ddpg-dqn

PPO2 DDPG and DQN, it needs only to be improved for functions to pass parameters by reference for extra speed no more python S***.

 

you're Welcome!

 

Hey,

if i add the files to a Project with ML-Agents installed, I get many "no suitable method found to override" errors.
Could you give a little guide to Setup the project?

Greetings

 

Well this Repo have no relation between ML Agents you can use them both, Just simply Add an Agent Component, and edit propreties, or change reward code via script by respecting the hierarchy of the code. you error come from something else. tell us more info about you're error.

 

well, I've made updates, I added PPO discrete, now fully functional you can test that in the gridworld I've added Q learning too

 

Ball Balance

 

this is super cool. how do I use the repo onto my own scene/environment? Am trying to use DDPG in my own ML Agents environment but am quite lost

 

This is a great bit of work, I've been playing around with it, and can confirm its learning in a WebGL build no problem, so I'd assume cross compatibility is no problem (Android & IOS compatibility) - something no other machine learning library is currently capable of doing!

I forked your project and optimised a few Unity specific bits to try and improve performance. (https://github.com/makeplayhappy/Unity-CSharp-ML/)

I'm getting some big spikes as the ActorNet:Backward runs through the Net, making the batch sizes small does help a little.

I wonder about changing the multidimensional arrays into jagged ones, as I think the C# compiler may be better able to optimise them, if there are any DOTs programmers reading this, I think this would be an idea candidate for that type of architecture.

@isawyou_unity
I'd like to add in some save / load functionality and was wondering if you could point me at your original source material - I have looked through Pytorch but I can't see much similarity between it and the code here - did you work of an old / different branch?

 

Wow man this is stellar work you've done here! Thank you for sharing!
I only wish I had found this sooner.
This is definitely on my list of things to test out now.