A new paper on arXiv examines deep reinforcement learning (DRL) for job scheduling in high-performance computing. The authors compare six DRL treatments for learned job selection, but they caution that raw performance against heuristics is not enough to establish the approach's worth.
The central concern is action masking. In many scheduling setups, DRL agents are constrained to only choose among feasible actions, and the paper argues that gains attributed to the learned policy may actually come from this masking rather than from the reinforcement learning itself. Without isolating that effect, a DRL scheduler that beats a heuristic may simply be benefiting from a restricted action space.
The study calls for a statistical evaluation that separates these factors. By comparing six treatments, the authors aim to show where DRL genuinely contributes and where apparent improvements are an artifact of the evaluation setup. The paper does not dismiss DRL for HPC scheduling, but it insists on a more rigorous standard before claiming practical value.