Enhancing cloud resource management: leveraging adversarial reinforcement learning for resilient optimization

Abstract

This paper introduces the first adversarial reinforcement learning (ARL) framework for resilient cloud resource optimization under dynamic and adversarial conditions. While traditional reinforcement learning (RL) methods improve adaptability, they fail when faced with sudden workload surges, security threats, or system failures. To address this, we propose an ARL-based approach that trains RL agents using simulated adversarial perturbations, such as workload spikes and resource drops, enabling them to develop robust allocation policies. The framework is evaluated using synthetic and real-world Google Cluster traces within an OpenAI Gym-based simulator. Results show that the ARL model achieves 82% resource utilization and a 180 ms response time under adversarial scenarios, outperforming static policies and conventional RL by up to 12% in terms of cost-effectiveness. Statistical validation (p<0.05) confirms significant improvements in resilience. This work demonstrates the potential of ARL for self-healing cloud schedulers in production environments.