An Overview of Natural Language State Representation for Reinforcement Learning

  • 2020-07-19 20:15:55
  • Brielen Madureira, David Schlangen
  • 3

Abstract

A suitable state representation is a fundamental part of the learning processin Reinforcement Learning. In various tasks, the state can either be describedby natural language or be natural language itself. This survey outlines thestrategies used in the literature to build natural language staterepresentations. We appeal for more linguistically interpretable and groundedrepresentations, careful justification of design decisions and evaluation ofthe effectiveness of different approaches.

 

Quick Read (beta)

loading the full paper ...