G-Wang / text-world-player

Framework and model code for the paper "Language Understanding for Text-based Games using Deep Reinforcement Learning", EMNLP 2015

Prerequisites

Torch
Also make sure you've setup the game environment using these instructions.

Lua Package requirements

1. luasocket
2. rnn
3. underscore

Most lua packages can be installed using luarocks install <pkg>

Runtime options

run_cpu has the list of user-defined settings used by the program. You can run the script as ./run_cpu <game_num> where <game_num> would be 1,2,3,etc. This allows us to choose the game server if multiple are running at the same time. See the script start.sh in the text-world code for details on starting multiple game servers.

The main options you should care about inside the run_cpu file are:

1. (important) text_world_location: Set this to the location of the text-world directory on your machine. You should have cloned the directory from here.
2. STEP_SIZE: This defines the number of steps taken by the agent in the game in an epoch.
3. max_steps: Maximum number of steps per episode of gameplay.
4. recurrent: Set this to 1 if using the LSTM for the Representation Generator.
5. bigram: Set this to 1 to use a bag-of-bigrams representation.
6. netfile: Choose the model to use for the Representation Generator. Remember to set the recurrent option appropriately.