ricky1203

@markovyao alternative implement to @ppwwyyxx solution: 1. create matrix vars to store LSTM hidden states 2. every agent assign an unique agent_index 3. use ```tf.gather``` to select the hidden states...

@etienne87 check the ```def _create_rnn_from_cell()``` in [model.py](https://gist.github.com/ricky1203/279898553f8ea16166b3382f3a4c782b) note: for hidden states stored in model, agent should predict/train in one model(GPU device) during one episode