Deep Actor-Critics with Tight Risk Certificates
Bahareh Tasdighi, Manuel Haussmann, Yi-Shan Wu, Andres R. Masegosa, Melih Kandemir
TL;DR
This work tackles the risk of unsafe generalization in deep actor-critic reinforcement learning by deriving tight risk certificates via a Recursive PAC-Bayes framework. By splitting evaluation data, forming data-informed priors, and bounding excess loss across multiple partitions, the approach yields high-probability certificates that predict test-time returns from minimal evaluation rollouts. The authors instantiate the framework with Bayesian neural networks and apply it to PPO, SAC, and REDQ across MuJoCo and related benchmarks, showing that deeper recursion and data-informed priors substantially tighten the certificates and improve correlation with actual performance. They also demonstrate the benefits of the local reparameterization trick in improving bound tightness. The results suggest that practical, reliable risk certificates can enable safer deployment and broader adoption of RL in safety-critical, real-world systems, while outlining clear directions for future work in online certification and physical-system validation.
Abstract
Deep actor-critic algorithms have reached a level where they influence everyday life. They are a driving force behind continual improvement of large language models through user feedback. However, their deployment in physical systems is not yet widely adopted, mainly because no validation scheme fully quantifies their risk of malfunction. We demonstrate that it is possible to develop tight risk certificates for deep actor-critic algorithms that predict generalization performance from validation-time observations. Our key insight centers on the effectiveness of minimal evaluation data. A small feasible set of evaluation roll-outs collected from a pretrained policy suffices to produce accurate risk certificates when combined with a simple adaptation of PAC-Bayes theory. Specifically, we adopt a recently introduced recursive PAC-Bayes approach, which splits validation data into portions and recursively builds PAC-Bayes bounds on the excess loss of each portion's predictor, using the predictor from the previous portion as a data-informed prior. Our empirical results across multiple locomotion tasks, actor-critic methods, and policy expertise levels demonstrate risk certificates tight enough to be considered for practical use.