Optimality and Approximation with Policy Gradient Methods in Markov Decision Processes
Offered By: Simons Institute via YouTube
Course Description
Overview
Explore the intricacies of policy gradient methods in Markov Decision Processes through this 55-minute lecture by Alekh Agarwal from Microsoft Research Redmond. Delve into optimality and approximation concepts as part of the "Emerging Challenges in Deep Learning" series at the Simons Institute. Examine MDP preliminaries, policy parameterizations, and the policy gradient algorithm, with a focus on softmax parameterization and entropy regularization. Analyze the convergence of entropy-regularized PGA, natural solutions, and proof ideas. Investigate restricted parameterizations, natural policy gradient updates, policy assumptions, and extensions to finite samples. Gain valuable insights into this crucial area of deep learning and reinforcement learning research.
Syllabus
Intro
Questions of interest
Main challenges
MDP Preliminaries
Policy parameterizations
Policy gradient algorithm
Policy gradient example: Softmax parameterization
Entropy regularization
Convergence of Entropy regularized PG
A natural solution
Proof ideas
Restricted parameterizations
A closer look at Natural Policy Gradient • NPG performs the update
Assumptions on policies
Extension to finite samples
Looking ahead
Taught by
Simons Institute
Related Courses
A Complete Reinforcement Learning System (Capstone)University of Alberta via Coursera Fundamentals of Deep Reinforcement Learning
Learn Ventures via edX Data Science Decisions in Time: Using Data Effectively
Johns Hopkins University via Coursera Reinforcement Learning with Gymnasium in Python
DataCamp Decision-Making for Autonomous Systems
Chalmers University of Technology via edX