Domain knowledge is crucial for effective performance in autonomous controlsystems. Typically, human effort is required to encode this knowledge into acontrol algorithm. In this paper, we present an approach to language groundingwhich automatically interprets text in the context of a complex controlapplication, such as a game, and uses domain knowledge extracted from the textto improve control performance. Both text analysis and control strategies arelearned jointly using only a feedback signal inherent to the application. Toeffectively leverage textual information, our method automatically extracts thetext segment most relevant to the current game state, and labels it with atask-centric predicate structure. This labeled text is then used to bias anaction selection policy for the game, guiding it towards promising regions ofthe action space. We encode our model for text analysis and game playing in amulti-layer neural network, representing linguistic decisions via latentvariables in the hidden layers, and game action quality via the output layer.Operating within the Monte-Carlo Search framework, we estimate model parametersusing feedback from simulated games. We apply our approach to the complexstrategy game Civilization II using the official game manual as the text guide.Our results show that a linguistically-informed game-playing agentsignificantly outperforms its language-unaware counterpart, yielding a 34%absolute improvement and winning over 65% of games when playing against thebuilt-in AI of Civilization.
展开▼
