Near-Optimal Differentially Private Reinforcement Learning
This work addresses privacy concerns in sensitive applications like personalized healthcare by enabling no-regret reinforcement learning with differential privacy, offering a significant improvement over prior methods.
The paper tackles the problem of online exploration in reinforcement learning with differential privacy constraints, achieving near-optimal regret that matches the information-theoretic lower bound for non-private learning under certain conditions and providing the first algorithm with privacy that becomes asymptotically free as the number of steps increases.
Motivated by personalized healthcare and other applications involving sensitive data, we study online exploration in reinforcement learning with differential privacy (DP) constraints. Existing work on this problem established that no-regret learning is possible under joint differential privacy (JDP) and local differential privacy (LDP) but did not provide an algorithm with optimal regret. We close this gap for the JDP case by designing an $ε$-JDP algorithm with a regret of $\widetilde{O}(\sqrt{SAH^2T}+S^2AH^3/ε)$ which matches the information-theoretic lower bound of non-private learning for all choices of $ε> S^{1.5}A^{0.5} H^2/\sqrt{T}$. In the above, $S$, $A$ denote the number of states and actions, $H$ denotes the planning horizon, and $T$ is the number of steps. To the best of our knowledge, this is the first private RL algorithm that achieves \emph{privacy for free} asymptotically as $T\rightarrow \infty$. Our techniques -- which could be of independent interest -- include privately releasing Bernstein-type exploration bonuses and an improved method for releasing visitation statistics. The same techniques also imply a slightly improved regret bound for the LDP case.