AI Paper Talks: Episode Two – Reinforcement Learning
Discussion of the paper “Reinforcement Learning as One Big Sequence Modeling Problem” by Michael Janner, Qiyang Li, Sergey Levine
Join Phil Steitz, Chief Technology Officer at Nextiva, for an Artificial Intelligence Paper Talk to discuss the paper, “Reinforcement Learning as One Big Sequence Modeling Problem” by Michael Janner, Qiyang Li, Sergey Levine, from UC Berkeley.
This paper introduces a novel approach to reinforcement learning (RL). The basic idea is to apply natural language processing attention-based techniques to learn the behavior of action, state, rewards sequences. The analog of next sentence prediction can then be used directly for imitation learning via beam search and other techniques can be used to extend to other RL applications. Experiments in the paper show how this technique can be applied to long-horizon dynamics prediction, imitation learning, goal-conditioned RL, and offline RL.
Nextiva Artificial Intelligence Paper Talks are technical discussions based on recent research about Natural Language Processing (NLP) or Artificial Intelligence (AI). This event will include a 30-minute presentation and 15 minutes of open discussion.
Watch the episode one 👇👀📺