Problem: it is not easy to switch between training on one or different grids. also loading from a checkpoint needs to know on which env was trained.

environments are linked to

ENV_NAME = "ReferenceModel-2-1"
CP_TRAINED_ON_ENV_NAME = "ReferenceModel-3-1"

new name should look like
ENV_NAME = "ReferenceModelMultiAgent"

and a separate list with named grids to use for training or evaluation, which correspond with the expected names for get_grid