Q-Learning with Fine-Grained Gap-Dependent Regret
This work addresses a theoretical limitation in reinforcement learning for researchers, providing refined regret bounds that better capture suboptimality gaps, though it is incremental as it builds on prior algorithms like UCB-Hoeffding and AMB.
The paper tackles the problem of coarse gap-dependent regret bounds in model-free reinforcement learning for episodic tabular Markov Decision Processes, establishing fine-grained regret bounds for both UCB-based and non-UCB-based algorithms, with experiments showing improved performance over existing methods like AMB.
We study fine-grained gap-dependent regret bounds for model-free reinforcement learning in episodic tabular Markov Decision Processes. Existing model-free algorithms achieve minimax worst-case regret, but their gap-dependent bounds remain coarse and fail to fully capture the structure of suboptimality gaps. We address this limitation by establishing fine-grained gap-dependent regret bounds for both UCB-based and non-UCB-based algorithms. In the UCB-based setting, we develop a novel analytical framework that explicitly separates the analysis of optimal and suboptimal state-action pairs, yielding the first fine-grained regret upper bound for UCB-Hoeffding (Jin et al., 2018). To highlight the generality of this framework, we introduce ULCB-Hoeffding, a new UCB-based algorithm inspired by AMB (Xu et al.,2021) but with a simplified structure, which enjoys fine-grained regret guarantees and empirically outperforms AMB. In the non-UCB-based setting, we revisit the only known algorithm AMB, and identify two key issues in its algorithm design and analysis: improper truncation in the $Q$-updates and violation of the martingale difference condition in its concentration argument. We propose a refined version of AMB that addresses these issues, establishing the first rigorous fine-grained gap-dependent regret for a non-UCB-based method, with experiments demonstrating improved performance over AMB.