Today, if you use "nb_env > 1" in the train function of all agents inheriting from DeepQAgent it is not clear at all that the environment provided should be an instance of MultiEnvironment and not an instance of the natural environment itself Environment.
This is confusing and should be clarified in this version, and the redundancy should be removed in future major release:
remove the "nb_env" keyword argument of the "train" function
initialize it with 1 if an Environment is passed as env argument of the train function
initialize it with the appropriate number if this is a MultiEnvironment
This should solve the issue, and will be properly documented in the doc (l2rpn-baselines + grid2op and in the getting_started notebook of grid2op)
Today, if you use "nb_env > 1" in the
train
function of all agents inheriting fromDeepQAgent
it is not clear at all that the environment provided should be an instance ofMultiEnvironment
and not an instance of the natural environment itselfEnvironment
.This is confusing and should be clarified in this version, and the redundancy should be removed in future major release:
Environment
is passed asenv
argument of the train functionMultiEnvironment
This should solve the issue, and will be properly documented in the doc (l2rpn-baselines + grid2op and in the
getting_started
notebook of grid2op)