Training a recurrent policy about pytorch-rl HOT 4 OPEN

The value function also needs to be an RNN, or you can pass the rnn output from the policy to the value net.

from pytorch-rl.

Any help regarding how to use RNN? should it be used single step, or over an episode(specifying times steps as sequences)? and how will the backpropagation take place, all at once at the end of episode? Do you have any code regarding this available?

from pytorch-rl.

I don't have any code for RNN yet. But I can imagine how it can be done.

Suppose the agent collected a batch of three episodes of length 5,6,7, so the total length is 18. You need to pad these episodes to be of the same length 7. So you will have an input size 7 x 3 x d, where d is your input dim. Then you pass it through an LSTM, which will give you an output of size 7 x 3 x h, where h is the output dim. You reshape it into 21 x h, and find where the episode steps (5,6,7) correspond to by indexing operation, then you will get the right output of 18 x h. Then you can just pass this to an MLP or any other network you want. All these operations are achievable by pytorch.

from pytorch-rl.

Okay, but should the functions like get_log_probability which is used in ppo step, be updated? because they compute forward propagation, getting new hidden vector after every backpropagation. or should the new hidden vector be ignored, and old hidden vector be used(that was obtained during data collection by agent).

from pytorch-rl.