Effective task scheduling in Human–Cyber–Physical Systems is critically important for enterprises to achieve optimized energy consumption and performance. Solving task scheduling problems, particularly large-scale instances, poses a high computational burden. While heuristics can be used to find near-optimal solutions, fast heuristics often suffer from limited performance. To mitigate a trade-off between computational burden and solution effectiveness, we integrate two machine learning techniques with evolutionary algorithms. Specifically, a long short-term memory-based autoencoder is incorporated into evolutionary algorithms to facilitate the latter’s global search ability. An operator selection mechanism based on deep reinforcement learning is developed to enhance their optimization capabilities.