What is the use of the attribute 'state' in the RecEnv from the single item recommender tutorial?

[davera-017]

The environment class in this tutorial has an attribute called state which tracks the user history but apparently is never used. Here is its initialization code:

class RecEnv(Environment):
    def __init__(
        self, actions: List[torch.Tensor], model: nn.Module, history_length: int
    ) -> None:
        ...
        self.state: torch.Tensor = torch.zeros((self.history_length, 100)).to(device)
        ...

I was wondering if this is somehow used inside other classes (e.g. some History Summarization Module) or if it is not used at all.

[rodrigodesalvobraz]

Hi Daniel, Why do you say it is never used? When I search for self.state, I see: The state is initialized with zeros. It is used to form state_batch, which is used to determine the reward in

reward = self.model(state_batch, action_batch) * 3

It is also gradually updated, upon each step, with the most recent action. After history_length steps, the zeros will have been overwritten and self.state it will be a record of the history_length last taken actions. This is the information used by self.model to compute the reward. I hope that helps. Rodrigo