FLASH: A Joint Optimization Framework of Splitting, Pairing, and Offloading Based on DRL towards Efficient Wireless LAM

The emergence of the sixth generation (6G) networks signifies a paradigm shift in wireless communications. Wireless large artificial intelligence model (WLAM), as a pivotal technology, leverages deep semantic awareness together with intelligent inference and decision-making capabilities to drive the evolution of applications in artificial intelligent internet of things (AIoT). However, despite the unprecedented efficiency, native intelligence, and seamless connectivity of 6G, deploying WLAM in AIoT environment remains challenging due to limited computational resource of edge terminals, communication bottlenecks from billion-scale parameters, and privacy trade-off. In response, we propose a hierarchical split federated learning framework assisted by edge intelligence of unmanned aerial vehicles (UAVs), called federated learning via aerial splitting hierarchy (FLASH), to help computing offloading of client-side through aerial nodes. The framework jointly optimizes model splitting layers, client-UAV pairing, and computational resource allocation of UAVs. Then, we formulate a mixed-integer nonlinear programming problem with objectives of weighted averages of single-round training latency and client energy consumption, subject to UAV energy and computational resource constraints. To solve the complex joint optimization problem, we formalize this problem as a markov decision process (MDP) and design a distributed synchronous deep reinforcement learning algorithm based on proximal policy optimization, achieving efficient joint resource scheduling through parallel training and synchronous policy updates. The simulation results prove that the proposed solution outperforms existing benchmarks in both performance of latency and energy efficiency.