CASUS Paper Reading Club, Pia Hanfeld, CASUS

This short presentation will be a summary of the current results after the practical part of the thesis is almost finished. Three methods for creating adversarial perturbations have been implemented. One of the methods has not been tested on Atari before and introduces a new and efficient attack on reinforcement learning (RL). It could be shown that all methods successfully lowered the overall performance of the agent. The presentation will also summarize the current issues and present ideas to solve them.