Deep Reinforcement Learning with Stacked Hierarchical Attention for Text-based Games

摘要: We study reinforcement learning (RL) for text-based games, which are interactive simulations in the context of natural language. While different methods have been developed to represent environment information and language actions, existing RL agents not empowered with any reasoning capabilities deal textual games. In this work, we aim conduct explicit knowledge graphs decision making, so that actions an agent generated supported by interpretable inference procedure. propose a stacked hierarchical attention mechanism construct representation process exploiting structure graph. extensively evaluate our method on number man-made benchmark experimental results demonstrate performs better than agents.