continuous action space about option-critic-pytorch

[lweitkamp]

There is no support for continuous action spaces currently, but there is no reason why it wouldn't be possible to augment the code to work with continuous action spaces. Each option essentially works as a policy gradient so you would have to change the following lines

option-critic-pytorch/option_critic.py

Lines 143 to 144 in 102436c

	self.options_W = nn.Parameter(torch.zeros(num_options, 64, num_actions))
	self.options_b = nn.Parameter(torch.zeros(num_options, num_actions))

where num_actions should be the dimensionality of the mean and diagonal covariance. Also change

option-critic-pytorch/option_critic.py

Line 172 in 102436c

action_dist = Categorical(action_dist)

to be a Gaussian distribution.