An Adaptive Cache Replacement Policy for Heterogeneous Workloads Using Machine Learning

This paper presents Adaptive Replacement via Policy Selection (ARPS), a novel, lightweight machine learning-based cache management framework. ARPS dynamically selects the most appropriate replacement policy (LRU, FIFO, or Random) at runtime based on the detected characteristics of the ongoing workload. I implement a low-overhead neural network predictor that monitors a minimal set of runtime performance counters (e.g., miss rate trends, access interval variance) to classify the workload's access pattern. The core insight is that no single policy is optimal for all scenarios, but a strategic switch between them can closely approximate optimal behaviour.