YoVDO

V-Learning - Simple, Efficient, Decentralized Algorithm for Multiagent RL

Offered By: Simons Institute via YouTube

Tags

Algorithm Design Courses Nash Equilibrium Courses

Course Description

Overview

Explore a groundbreaking lecture on V-Learning, a novel decentralized algorithm for multiagent reinforcement learning (MARL). Delve into the challenges of MARL, particularly the curse of multiagents, and discover how V-Learning overcomes the exponential scaling of joint action spaces. Learn about the algorithm's ability to efficiently learn Nash equilibria, correlated equilibria, and coarse correlated equilibria in episodic Markov games. Understand the key differences between V-Learning and classical Q-learning, and how V-Learning's focus on V-values enables superior performance in MARL settings. Examine topics such as adversarial bandits, duality gaps, and no-regret learning in the context of Normal Form Games. Gain insights into the algorithm's applications, its relationship to single-agent reinforcement learning, and its potential to revolutionize the field of multiagent learning.

Syllabus

Introduction
Problems
cursive multiagents
centralized vs decentralized
characterization
markup games
policy value
setting
learning
Single Engine Reinforcement
Challenges
Normal Form Games
adversarial bandit
duality gap
no regret learning
converge


Taught by

Simons Institute

Related Courses

Natural Language Processing
Columbia University via Coursera
Intro to Algorithms
Udacity
Conception et mise en œuvre d'algorithmes.
École Polytechnique via Coursera
Paradigms of Computer Programming
Université catholique de Louvain via edX
Data Structures and Algorithm Design Part I | 数据结构与算法设计(上)
Tsinghua University via edX