Perceiving the World: Question-guided Reinforcement Learning for Text-based Games

• DOI: 10.18653/v1/2022.acl-long.41
• Publication Date: 2022-06-03T01:34:53Z
• Authors: Yunqiu Xu; Meng Fang; Ling Chen; Yali Du; Joey Zhou; Chengqi Zhang
• Keywords: Robustness; Sample (material); Training set

Text-based games provide an interactive way to study natural language processing. While deep reinforcement learning has shown effectiveness in developing the game playing agent, low sample efficiency and large action space remain be two major challenges that hinder DRL from being applied real world. In this paper, we address by introducing world-perceiving modules, which automatically decompose tasks prune actions answering questions about environment. We then propose a two-phase training framework decouple learning, further improves efficiency. The experimental results show proposed method significantly performance Besides, it shows robustness against compound error limited pre-training data.