Implement class functions mantle and arm in the CMA-ES evolution algo. mantle coordinates evolution by sending by sending information about population individuals to the arm subprocesses, which each run rollouts for specific individuals from the population before passing their fitness back to mantle.
This is in place, but actually the speed bottleneck is running ca steps on the GPU (at least on my machine) so running multiple instances with mpi4py does not typically speed anything up.
Implement class functions
mantle
andarm
in the CMA-ES evolution algo.mantle
coordinates evolution by sending by sending information about population individuals to thearm
subprocesses, which each run rollouts for specific individuals from the population before passing their fitness back tomantle
.