Assessing the Robustness of Intelligence-Driven Reinforcement Learning
Lorenzo Nodari, Federico Cerutti · Zenodo (CERN European Organization for Nuclear Research) · 2023
In this work, we carry out a preliminary investigation into the robustness of reward machine-based reinforcement learning agents to noise in their labelling function output, a critical symbolic component that governs their functioning. By subjecting optimally trained agents to varying amount of randomly generated noise, we demonstrate their lack of robustness to modifications of the labelling function, thus highlighting the need for further analysis of the robustness of such approaches to more refined adversarial settings.