Deep reinforcement learning: from playing video games to
generating natural text
PRESENTER:
Harry Guo
NRC
ABSTRACT:
I will first briefly talk about deep reinforcement learning and its
application to playing video games (with a live demo of playing Pong
game).
I will then re-present my paper, entitled Generating Text with Deep
Reinforcement Learning.
The paper proposed a novel schema for sequence to sequence learning, in
which a Deep Q-Network (DQN) is deployed to decode output sequence
iteratively. Through embracing an encoder-decoder Long Short-Term Memory
(LSTM) network to automatically create internal states features and
formulate potential actions from the input sequence, the DQN learns to
make decisions on which candidate actions will be selected to modify the
current decoded output sequence. Consequently, the decoder is able to
first tackle easier portions of the sequences, and then turn to cope with
difficult parts.