Abstract:

Reinforcement learning is about learning to make sequential decisions through interactions with an environment. Traditional tabular reinforcement learning suffered from the curse of dimensionality and therefore was rarely employed in important applications. Recently reinforcement learning has been combined with deep learning to create Deep Reinforcement Learning (DRL), enabling researchers to break through the curse of dimensionality and obtain striking results in a wide range of diverse applications, including learning to play the Atari games from raw pixel inputs, learning to beat the grand masters at Go, learning to control tokamak plasmas for nuclear fusion, learning computationally efficient algorithms for matrix multiplication, and most recently making the responses in ChatGPT3 more human-like and accurate. In this talk, I will first briefly review some of the DRL breakthroughs, then discuss our own algorithmic research on DRL for high-dimensional state and action spaces, and its application to robotic locomotion.