Multi-source strategy adaptation network for time-triggered flow scheduling in unknown environments

Traditional Ethernet, based on event-triggered scheduling, often suffers from resource contention, queuing delays, and jitter, which can degrade the efficiency and sustainability of modern engineering systems under dynamic environmental conditions. Time-triggered (TT) mechanisms–such as those in Time-Triggered Ethernet (TTE) and Time-Sensitive Networking (TSN)–address these limitations by enabling precise scheduling of data flows, thereby ensuring high predictability and real-time performance. These capabilities are essential in safety-critical domains including aerospace, smart grids, and industrial automation. However, generating a static schedule table (SST) for TT flows is an NP-hard problem. While Deep Reinforcement Learning (DRL) has shown promise in handling large-scale TT scheduling, its high training overhead limits applicability in time-sensitive and resource-constrained environments. To overcome this challenge, we propose the Multi-Source Strategy Adaptation Network (MSAN), a soft computing framework that integrates multiple DRL agents for collaborative decision-making and adaptive exploration. MSAN accelerates training convergence, balances exploration and exploitation more effectively, and demonstrates improved scalability–with training time growing nearly linearly with the number of TT flows. Extensive experiments across diverse network topologies and unseen test environments demonstrate that MSAN finds feasible scheduling solutions significantly faster than conventional DRL approaches. The results highlight its potential for rapid adaptation, reduced computational cost, and enhanced support for sustainable engineering practices through intelligent soft computing. Our codes are available at: https://github.com/CodeZenTao/MSAN.