On a Hybrid Scheduling Approach Combining Whale Optimization and Deep Reinforcement Learning for DAG-Structured Machine Learning Workloads

In RF computing systems, devices with limited resources work together to run Machine Learning (ML) training and inference. In these systems, efficient task scheduling is key to meeting time and energy limits. The workloads are often represented as Directed Acyclic Graphs (DAGs), where each task depends on others. Scheduling DAG tasks across different accelerators and wireless links is challenging because of limits in RF channel bandwidth, device differences, and task data dependencies. This problem can be modeled as a Nonlinear Integer Programming (NIP) problem and is NP-complete. Based on our earlier work showing a clear relationship between Scheduling Plan Distance (SPD) and Finish Time Gap (FTG), we design WORL-RTGS, a hybrid method. It combines the global search ability of the Whale Optimization Algorithm (WOA) with the adaptive learning of Double Deep Q-Networks (DDQN). Experiments with real Alibaba ML workload traces in RF-based GPU systems show that WORL-RTGS improves WOA’s stability and reduces completion time by up to 66.56% compared with five state-of-the-art baselines.