WCSAC: Worst-Case Soft Actor Critic for Safety-Constrained Reinforcement Learning

Abstract

Safe exploration is regarded as a key priority area for reinforcement learning research. With separate reward and safety signals, it is natural to cast it as constrained reinforcement learning, where expected long-term costs of policies are constrained. However, it can be hazardous to set constraints on the expected safety signal without considering the tail of the distribution. For instance, in safety-critical domains, worst-case analysis is required to avoid disastrous results. We present a novel reinforcement learning algorithm called Worst-Case Soft Actor Critic, which extends the Soft Actor Critic algorithm with a safety critic to achieve risk control. More specifically, a certain level of conditional Value-at-Risk from the distribution is regarded as a safety measure to judge the constraint satisfaction, which guides the change of adaptive safety weights to achieve a trade-off between reward and safety. As a result, we can optimize policies under the premise that their worst-case performance satisfies the constraints. The empirical analysis shows that our algorithm attains better risk control compared to expectation-based methods.

Trustworthy Reinforcement Learning Against Intrinsic Vulnerabilities: Robustness, Safety, and Generalizability

“…where we denote 𝜃 as the policy parameters. Alternating between the maximizing over 𝜃 via any unconstrained reinforcement learning algorithms and minimizing over the Lagrange multiplier 𝜆 yields a series of Lagrangian-based methods to solve the safe deployment problem [208]. Chow et al [31] propose PDO to update both primal parameters and dual variables by performing gradient descent based on on-policy estimations of the reward and cost value functions 𝑉 𝜋 𝜃 𝑟 (𝜇 0 ) and 𝐽 𝑐 (𝜋 𝜃 ).…”

Multi-task safe reinforcement learning for navigating intersections in dense traffic

“…Safe RL. Constrained optimization techniques are usually adopted to solve safe RL problems (Garcıa & Fernández, 2015;Sootla et al, 2022;Yang et al, 2021;Flet-Berliac & Basu, 2022). Lagrangian-based methods use a multiplier to penalize constraint violations (Chow et al, 2017;Tessler et al, 2018;Stooke et al, 2020;Chen et al, 2021b).…”

Reducing Safety Interventions in Provably Safe Reinforcement Learning

“…The Lagrangian method [20] is a popular way to address constrained RL problems by converting them to a dual problem, following constrained optimization theory [23,Chapter 5] and optimizing the Lagrangian multiplier in conjunction with the RL policy. More recent works, such as constrained policy optimization [24], constrained RL with a PID-controlled Lagrange multiplier (PID-Lagrangian) [25], and worst-case soft actor-critic [26] build on the Lagrangian method and make it applicable to deep RL. A drawback of constrained RL is that it cannot guarantee safety in a provable way, as the learned behavior is not formalized or proven.…”