Long Short-Term Memory Network with Reinforced Learning

Hi all,

I'm working with LSTM deep neural networks for a sequence-to-sequence forecasting.

I read papers about the fact that there is an "exposure bias" in LSTM networks trained with the previous steps of a sequence which you want to forecast, since when you let the model free to forecast it can't rely on real data inputs but on its own predicted steps.

In practise the model is trained with "solid" inputs but it is not trained to provide solid inputs for itself.

In the same paper, the reinforced learning was suggested as a solution, how can I combine the LSTM networks and RL with matlab toolboxes in order to perform sequence-to-sequence forecasting?