Advances in Reinforcement Learning and Game Theory

The field of reinforcement learning and game theory is rapidly advancing, with a focus on developing more efficient and robust algorithms for complex decision-making problems. Recent research has explored the use of symmetry in Markov games, enabling players to compete without observing payoffs, and the development of novel frameworks for sparse tensor decomposition and nonlinear reinforcement learning. Additionally, there has been a growing interest in harnessing information in incentive design and analyzing the fundamental structure of reward functions to enable efficient sparse-reward learning. Noteworthy papers include: Playing Markov Games Without Observing Payoffs, which introduces a new class of zero-sum symmetric Markov games and shows that players can compete without observing payoffs. ReLATE: Learning Efficient Sparse Encoding for High-Performance Tensor Decomposition, which presents a novel learning-augmented method for automatic construction of efficient sparse tensor representations. The Geometry of Nonlinear Reinforcement Learning, which presents a unified geometric framework for viewing reward maximization, safe exploration, and intrinsic motivation as instances of a single optimization problem.

Sources

Playing Markov Games Without Observing Payoffs

ReLATE: Learning Efficient Sparse Encoding for High-Performance Tensor Decomposition

The Geometry of Nonlinear Reinforcement Learning

Succeed or Learn Slowly: Sample Efficient Off-Policy Reinforcement Learning for Mobile App Control

Harnessing Information in Incentive Design

Imitate Optimal Policy: Prevail and Induce Action Collapse in Policy Gradient

What Fundamental Structure in Reward Functions Enables Efficient Sparse-Reward Learning?

Shift Before You Learn: Enabling Low-Rank Representations in Reinforcement Learning

Reinforcement Learning with Anticipation: A Hierarchical Approach for Long-Horizon Tasks

Teaching Precommitted Agents: Model-Free Policy Evaluation and Control in Quasi-Hyperbolic Discounted MDPs

Physics-informed Value Learner for Offline Goal-Conditioned Reinforcement Learning

Built with on top of