Keywords: [ Reinforcement Learning and Planning ]

[
Abstract
]
Spotlight presentation:
Reinforcement Learning and Bandits

Wed 21 Jul 7 a.m. PDT — 8 a.m. PDT

[
Slides]

[ Paper ]

Wed 21 Jul 9 a.m. PDT — 11 a.m. PDT

Wed 21 Jul 7 a.m. PDT — 8 a.m. PDT

Abstract:
Q-learning, which seeks to learn the optimal Q-function of a Markov decision process (MDP) in a model-free fashion, lies at the heart of reinforcement learning. Focusing on the synchronous setting (such that independent samples for all state-action pairs are queried via a generative model in each iteration), substantial progress has been made recently towards understanding the sample efficiency of Q-learning. To yield an entrywise $\varepsilon$-accurate estimate of the optimal Q-function, state-of-the-art theory requires at least an order of $\frac{|S||A|}{(1-\gamma)^5\varepsilon^{2}}$ samples in the infinite-horizon $\gamma$-discounted setting. In this work, we sharpen the sample complexity of synchronous Q-learning to the order of $\frac{|S||A|}{(1-\gamma)^4\varepsilon^2}$ (up to some logarithmic factor) for any $0<\varepsilon <1$, leading to an order-wise improvement in $\frac{1}{1-\gamma}$. Analogous results are derived for finite-horizon MDPs as well. Notably, our sample complexity analysis unveils the effectiveness of vanilla Q-learning, which matches that of speedy Q-learning without requiring extra computation and storage. Our result is obtained by identifying novel error decompositions and recursion relations, which might shed light on how to study other variants of Q-learning.

Chat is not available.