Meaning of keyword abbreviations

[zichunxx]

Hi!

I'm new to model-based reinforcement learning and thanks for your contribution to Pytorch users.

I tried to read your code to understand the logistics of dreamerv3 but found some details are not mentioned in the original paper, especially the abbreviations of some keywords.

For example,

dreamerv3-torch/networks.py

Lines 174 to 179 in 4e50f30

	def obs_step(self, prev_state, prev_action, embed, is_first, sample=True):
	# initialize all prev_state
	if prev_state == None or torch.sum(is_first) == len(is_first):
	prev_state = self.initial(len(is_first))
	prev_action = torch.zeros((len(is_first), self._num_actions)).to(
	self._device

the prev_state is a dict including three keys, i.e., logit, stoch, and deter.

What do they mean and where can I find a more specific explanation?

Thanks for your time!

[NM512]

Thanks for asking!

Each of them means the following

• deter: deterministic state that is mentioned as h_t in the paper
• logit: stoch before processing with softmax function
• stoch: stochastic discrete state that is mentioned as z_t in the paper

If you have any questions, please ask more!