LGAISep 10, 2024

Double Successive Over-Relaxation Q-Learning with an Extension to Deep Reinforcement Learning

arXiv:2409.06356v21 citationsh-index: 1
AI Analysis

This work addresses convergence issues in reinforcement learning for practitioners, though it is incremental as it builds on existing SOR methods.

The paper tackles slow convergence in Q-learning by proposing a sample-based, model-free double SOR Q-learning algorithm, which reduces bias and is extended to deep RL, showing improved performance in tabular and large-scale environments.

Q-learning is a widely used algorithm in reinforcement learning (RL), but its convergence can be slow, especially when the discount factor is close to one. Successive Over-Relaxation (SOR) Q-learning, which introduces a relaxation factor to speed up convergence, addresses this issue but has two major limitations: In the tabular setting, the relaxation parameter depends on transition probability, making it not entirely model-free, and it suffers from overestimation bias. To overcome these limitations, we propose a sample-based, model-free double SOR Q-learning algorithm. Theoretically and empirically, this algorithm is shown to be less biased than SOR Q-learning. Further, in the tabular setting, the convergence analysis under boundedness assumptions on iterates is discussed. The proposed algorithm is extended to large-scale problems using deep RL. Finally, the tabular version of the proposed algorithm is compared using roulette and grid world environments, while the deep RL version is tested on a maximization bias example and OpenAI Gym environments.

Code Implementations1 repo
Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes