Machine learning-based multi-criteria optimization for adaptive hybrid flow shop scheduling

Abstract

The Adaptive Hybrid Flow Shop Scheduling Problem (AHFSSP) is a cornerstone of modern manufacturing, involving the coordination of jobs through multiple processing stages, each equipped with parallel machines, under conflicting objectives such as minimizing makespan, tardiness, and energy utilization. As an NP-hard problem, HFSSP poses significant challenges in dynamic industrial environments where uncertainties like fluctuating job arrivals, and varying priorities disrupt traditional scheduling methods. This thesis proposes an innovative hybrid methodology that integrates unsupervised machine learning, evolutionary optimization, and reinforcement learning to address these challenges. Specifically, k-means clustering is employed to group jobs based on critical features, simplifying the scheduling problem by prioritizing high-impact tasks. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) generates a Pareto-optimal set of solutions, effectively balancing the four conflicting objectives to produce diverse, high-quality schedules. Proximal Policy Optimization (PPO), a reinforcement learning algorithm, refines these schedules by shaping rewards and updating policies that respond to real-time changes, such as new job arrivals. The methodology is validated using a real-world dataset from a Fast-Moving Consumer Goods (FMCG) company, comprising 100 jobs across four stages with 12–17 machines per stage. The hybrid model achieved superior performance, with a makespan of 312.9 minutes, tardiness of 94.88 minutes, and a computation time of 0.0203 seconds, outperforming the benchmarks. Statistical validation through ANOVA and T-tests confirmed significant performance improvements, with on-time completion rates of 61%–74% across priority clusters. Comparative analysis against an exact method (Branch and Bound) demonstrated near-optimal solutions with a computation time over 100 times faster, highlighting the model’s suitability for real-time applications. By combining clustering for task prioritization, NSGA-II for global optimization, and PPO for dynamic adaptation, this research delivers a scalable, intelligent, and robust solution for adaptive HFSSP, advancing operational efficiency in complex manufacturing systems.