Deep Reinforcement Learning Enhanced by Graph Neural Networks and Meta-Learning for Dynamic Model Partitioning and Resource Allo

With the advent of the 6G era, the \\textbf{Space\u2013Air\u2013Ground Integrated Network (SAGIN)} is considered a key architecture to achieve ubiquitous coverage and intelligent services. Existing research on collaborative inference primarily focuses on deterministic tasks and fixed partitioning strategies, which are insufficient to meet the stringent requirements of multi-type Deep Neural Networks (DNNs) for resource allocation and latency in dynamic environments . To address this issue, this paper proposes a \\textbf{deep reinforcement learning method enhanced by Graph Neural Networks (GNNs) and meta-learning}, enabling dynamic model partitioning and collaborative resource optimization for multi-task DNN inference. The proposed method extracts the topological relationships of the space\u2013air\u2013ground network and the coupling features among tasks through GNNs, introduces a meta-learning mechanism to quickly adapt to environmental changes, and integrates an improved Proximal Policy Optimization (PPO) framework to output hybrid discrete\u2013continuous actions. Experimental results demonstrate that the proposed approach significantly outperforms seven state-of-the-art comparison methods in terms of energy consumption, latency, and task completion rate, verifying its superiority and robustness. The main contribution of this paper lies in integrating GNNs, meta-learning, and reinforcement learning to solve the dynamic optimization problem of collaborative inference in SAGIN, providing new insights for future research.