Utilizing textual information for reinforcement learning

Abstract

In this paper, we research the effectiveness of utilizing textual information for reinforcement learning agent while conducting reinforcement learning. We design a hierarchical policy architecture, and make the low-level policy to take on fine-grained control of the agent and let the high-level policy to order the low-level policy how to navigate, dividing the process of game control into high-level planning and low-level control. We give the manual text as an input to high-level policy to utilize the high-level planning information contained in the text. We experiment with Montezuma's Revenge on gym environment and compared performance of the model using textual information with the model that doesn't use textual information.