In this paper, we consider the task of learn-ing control policies for text-based games. In these games, all interactions in the vir-tual world are through text and the under-lying state is not observed. The resulting language barrier makes such environments challenging for automatic game players. We employ a deep reinforcement learning framework to jointly learn state represen-tations and action policies using game re-wards as feedback. This framework en-ables us to map text descriptions into vec-tor representations that capture the seman-tics of the game states. We evaluate our approach on two game worlds, compar-ing against a baseline with a bag-of-words state representation. Our algorithm out-performs the baseline on quest completion by 54% on a newly created world and by 14% on a pre-existing fantasy game.
Mendeley saves you time finding and organizing research
Choose a citation style from the tabs below