Learning-Driven Portfolio–Island Decoding for Adaptive Scheduling Problems

Integrated production–transport scheduling with mobile robots tightly couples manufacturing operations and internal logistics, resulting in complex decision interactions that challenge conventional optimization approaches. Many existing evolutionary frameworks rely on a single decoding strategy throughout the search, limiting their ability to adapt to heterogeneous instance structures and dynamic scheduling characteristics. This paper introduces a learning-driven portfolio–island framework that shifts the focus from designing new operators to learning which decoding strategies are most effective during the search process. Multiple heterogeneous decoders evolve in parallel on separate islands, while an upper confidence bound multi-armed bandit continuously evaluates their contributions to makespan improvement and dynamically reallocates population resources to balance exploration and exploitation. The proposed framework is evaluated under different scheduling settings, including pure makespan minimization and extended production–transport coordination scenarios. Computational experiments on benchmark datasets demonstrate that the portfolio–learning mechanism consistently improves solution quality and convergence stability compared with single-decoder approaches, particularly for large and structurally diverse instances. These results highlight adaptive decoder selection as a powerful mechanism for robust and scalable scheduling in mobile-robot-supported manufacturing systems, providing a foundation for intelligent decision-making in production environments.