Is there currently an elegant way of adding an RNN or LSTM to an agent?

Question

Is there currently an elegant way of adding an RNN or LSTM to an agent?

chokosabe opened this issue 4 years ago · comments

chokosabe commented 4 years ago

Looking at custom networks and LSTMs and RNNs in particulat. It isn't very clear how these might be added to an agent.

It may just be a case of additions being made to the docs.

Christian Hidber commented 4 years ago

:-)

Christian Hidber · Answer 1 · Mon Sep 07 2020 02:23:03 GMT+0800 (China Standard Time)

Actually we were contemplating about an extension to EasyAgents where you could easily provide an "arbitrary" nn, eg defined in keras or so. However it wasn't obvious to us how this might be done in an elegant fashion while still supporting all the rl libraries EasyAgents sits on top of (see also #25 ). So currently we only support fully connected networks. Did you try to integrate rnns or lstms ? In our uses cases so far the fully connected architecture was sufficient. Do you have a real life uses case which benefits from LSTMs or RNNs ? We are always very grateful for new, concrete applications.

chokosabe · Answer 2 · Thu Sep 10 2020 16:25:00 GMT+0800 (China Standard Time)

Thanks for the reply. For a real life use case I was thinking about about some language/text and time series problems (where there are a few LSTM papers claiming performance). Haven't tried to integrate any with easyagents so far but very impressed with all the work on it so far, prob the easiest library out there for this.