Authors: Konka Kishan, Thuppathi Krishna Sree, Ramagiri Nissy Jasmine, Prathikantam Rakshitha
Abstract: Deep reinforcement learning (DRL) has excelled in video games but remains vulnerable to adversarial attacks. The project unveils Orthogonal Adversarial DRL (OADRL) to improve robustness in both discrete and continuous action spaces. OADRL integrates orthogonal regularization to limit overfitting and adversarial training to enhance resilience. The method assess against standard DRL models, measuring reward stability, adversarial robustness, and generalization. The project presents the OADRL reduces sensitivity to perturbations while maintaining high performance. OADRL improves robustness, ensuring smoother policies and greater resistance to adversarial noise. The insight highlight its potential for real-world applications like robotics and autonomous systems.
DOI: http://doi.org/
