An Enhanced Discrete Hummingbird Algorithm With Reinforcement Learning for Efficient Hybrid Flow Shop Scheduling

ABSTRACT To overcome the limitations of the artificial hummingbird algorithm (AHA), such as slow convergence and its inability to address discrete optimization problems, this paper proposes an improved Q‐learning‐based discrete artificial hummingbird algorithm (QIDAHA). The original foraging strategies of AHA are discretized, and a tailored model for the hybrid flow‐shop scheduling problem (HFSP) is constructed, with specifically designed encoding and decoding mechanisms. By integrating reinforcement learning, Q‐learning is employed to adaptively select the optimal flight directions of hummingbirds according to environmental states and feedback, while an ‐convergence factor balances exploration and exploitation during action selection. In addition, an elite neighborhood optimization strategy is introduced in the later search stages to accelerate convergence and enhance solution accuracy. The effectiveness of QIDAHA is validated on three small‐scale and 10 large‐scale HFSP benchmark instances, and its performance is compared against several state‐of‐the‐art discrete algorithms. Experimental results show that QIDAHA achieves superior scheduling performance, with an average optimal makespan of 604.6 on large‐scale problems, outperforming SHPSO (613.5), PSO (632.5), DWOA (645.1), IDDE (631.3), DSSA (623.4), and DAHA (613.2). This study not only extends the applicability of AHA to discrete scheduling but also provides an efficient and reliable approach to solving HFSP. The proposed algorithm demonstrates both theoretical significance and practical value for optimizing complex manufacturing systems. Finally, the structure of this paper is organized as follows: Chapter 1 introduces the overall background; Chapter 2 presents the traditional Artificial Hummingbird Algorithm (AHA); Chapter 3 describes its discrete version; Chapter 4 analyzes the shortcomings of the discrete AHA and introduces the Qlearning–enhanced improvements; Chapter 5 reports the experimental studies; and Chapter 6 provides the overall conclusions.