Are there some blocking points in adding LSTM to DQN?

I will probably change the DQN policy & Trainer to support LSTM (or at least RNN).

Do you know major problems preventing that?
I am asking since DQN doesn’t currently support the RNN and LSTM options in RLLib. I would like to be sure that the is no serious known issue preventing that.

@sven1977 @ericl could you chime in here?

Hey @Maxime_Riche , great question! :slight_smile:
We do have an R2D2 agent (since a few weeks ago) in the current master (and upcoming 1.3 release).
It’s basically a vanilla DQN that runs with LSTMs/RNNs by storing sequences (of len max_seq_len) in the buffer, then samples these sequences as-is from the buffer and runs them through the LSTM for learning updates. The paper we used is linked here:


@sven1977 ,

as far as I understand it that means then that a max_seq_len=1 would make the LSTM layer meaningless?