Hello,
When I first got into RL I used matlab but it was incredibly slow compared to python so I switched. Fast forward awhile to today, my RLLib multiagent python env trains and learns using PPO, so I have a baseline. I recently set up a RLLib multiagent env that interfaces to matlab (using MATLAB python API package) and wanted to compare results and training speeds. Where the entire simulator is built in matlab, and python is just used to wrap it in a RLLib multiagent env. Therefor I have all the benefits of rllibs algos and tensorboard etc… I normally train with about half the logical processors I have available on my cpu, but if I try to do this with the matlab verson I get errors with matlab because I believe only one instance is being created.
So, I was curious if or how to create multiple instances of matlab and assign each specific rollout worker to one. Or vice versa, have each rollout worker create a separate instance of matlab.
Thanks!