Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a 13-minute conference presentation introducing CAMP (Certified-rAdius-Maximizing Policy), a novel training paradigm that enhances the robustness of deep reinforcement learning agents against adversarial attacks and noisy observations. Learn how traditional DRL systems, despite their strong performance in dynamic environments, remain vulnerable to adversarial perturbations, and discover how existing certifiably robust approaches using simple Gaussian augmentations create suboptimal trade-offs between certified robustness and performance returns. Understand the theoretical foundation behind CAMP's approach, which leverages the insight that global certified radius can be derived from local certified radii based on training-time statistics, enabling the formulation of a surrogate loss function that optimizes policies for better utility without compromising provable robustness. Examine the implementation of policy imitation as a stabilization technique for CAMP training and review experimental results demonstrating significant improvements in the robustness-return trade-off across various control and decision-making tasks, with CAMP achieving up to twice the certified expected return compared to baseline methods. Gain insights into the mathematical formulations, algorithmic design choices, and practical applications of this approach to creating more reliable and secure reinforcement learning systems for real-world deployment scenarios.
