GPR-161 Meta-Reinforcement Learning with Discrete World Models for Adaptive Load Balancing

We present a novel integration of the RL^2 meta-reinforcement learning algorithm with discrete world models, employing the DreamerV3 architecture, to enhance load balancing in operating systems. This integration allows for rapid adaptation to dynamic workload distributions with minimal retraining. In experiments using the Park load balancing environment, our approach outperformed the traditional AC3 algorithm in both standard and adaptive trials. Additionally, it exhibited strong resilience to catastrophic forgetting, maintaining high performance despite continuous variations in workload distribution and size. These results demonstrate the effectiveness of combining recurrent policy networks with discrete world models, offering a significant advancement in meta-learning capabilities for dynamic operating system environments. This work has important implications for improving resource management and performance in modern operating systems, addressing the challenges posed by increasingly dynamic and heterogeneous workloads.