In order to train a computer agent to play a text-based computer game, we
must represent each hidden state of the game. A Long Short-Term Memory (LSTM)
model running over observed texts is a common choice for state construction.
However, a normal Deep Q-learning Network (DQN) for such