Nils Jansen has received an ERC Proof of Concept Grant for the project Pro-DEUCE: A Platform for Safe Reinforcement Learning and Verification with Industrial Applications.
AI agents trained by reinforcement learning now steer robots, telecommunication networks and infrastructure. These agents are vulnerable to adversarial attacks. Even small, targeted manipulations of their sensor input can cause them to act unsafely. Pro-DEUCE builds VeriGym, an open-source platform that formally verifies an agent stays safe under such manipulations. Several industrial partners from the Netherlands, Germany, the United Kingdom and Sweden have agreed to collaborate on the project. Pro-DEUCE builds on Jansen’s ERC Starting Grant DEUCE (Data-Driven Verification and Learning Under Uncertainty).
General note: In case of using gender-assigning attributes we include all those who consider themselves in this gender regardless of their own biological sex.