Pre-trained Bert for Natural Language Guided Reinforcement Learning in Atari Game
Xin Li, Yu Zhang, Junren Luo, Yifeng Liu · 2022 34th Chinese Control and Decision Conference (CCDC) · 2022
Inspired by the dream of using natural language to guide an Agent, we propose a language understanding model that links natural language commands with particular target states, so that the agent can better understand the tasks expressed in natural language. The agent performs exploration in the environment according to human intentions through the decision-making method of reinforcement learning, which will improve the performance and efficiency of exploration. By using the most advanced text-based pre-trained language model Bert, the language understanding model is robust to human commands.