Real-time multiobjective scheduling in hybrid flow shop base on D3QN-KAN

Rational and efficient resource scheduling in hybrid flow shop(HFS) is crucial for enhancing manufacturing efficiency. This study addresses two challenges in solving scheduling problems with deep reinforcement learning: neural network parameter redundancy, and dynamic balance optimization of multiple objectives in multi-target scheduling scenarios. We propose a D3QN-KAN-based multi-objective real-time scheduling method. By embedding Kolmogorov-Arnold Networks (KAN) into a Double Dueling Deep Q-Network(D3QN), we utilize its spline-parameterized activation functions to achieve efficient characterization of high-dimensional state spaces. A dynamic segmented reward mechanism is designed to construct nonlinear adaptive reward surfaces, combined with quadratic penalty functions to realize collaborative optimization of four-dimensional objectives including makespan. Experimental results demonstrate that in mixed-flow workshop scenarios with 10 workpiece types and 10 devices, compared to baseline models like D3QN-EfficientNet, the proposed method achieves 5.3% reduction in maximum makespan, 5.1% decrease in average makespan, 3.6% reduction in total machine costs, and 3.9% reduction in total machine load, providing efficient dynamic decision support for complex manufacturing systems.