Language as a Scaffold for Reinforcement Learning

Explore how language can enhance reinforcement learning in this 31-minute talk by Jacob Andreas from MIT. Delve into an NLP perspective on RL, examining concepts like instructions as observations, tasks and subtasks, and the options framework. Discover approaches for learning from demonstrations, fast adaptation, and using natural language options. Investigate language modeling and representation in the context of string-valued MDPs and text adventure games. Learn about the application of language-based techniques to real-world scenarios and policy sketches. Gain insights into experimental results and the potential of language as a scaffold for improving reinforcement learning algorithms.

Intro
An NLPer's view of RL
Instructions as observations
Tasks & subtasks
The options framework
Learning from demonstrations
The mini-craft task
The path-walking task
Fast adaptation
Natural language options
Language for goal inference
Experimental results
Language modeling and represente
The string-valued MDP
Text adventure games
From language to the real world
Learning from policy sketches