Zero-Shot Generalizable Task Offloading in Edge Computing: a Causal Reinforcement Learning Approach

Multi-Access Edge Computing (MEC) is a key solution to address the computational resource constraints of User Equipment (UE). To achieve generalization on unseen scenarios, existing gradient-based meta-reinforcement learning (Meta-RL) methods rely on fine-tuning model parameters via gradient descent during the testing phase. However, in real-world MEC environments, parameter fine-tuning is frequently impractical owing to device heterogeneity, computational resource constraints, and stringent real-time requirements. Moreover, even when fine-tuning is possible, the additional processing overhead can significantly undermine the latency advantages provided by edge computing. To overcome these limitations, we focus on the zero-shot generalization problem in MEC task offloading, which aims to ensure the model's generalization ability without performing parameter fine-tuning. To this end, we propose a Causal Reinforcement Learning (CRL) method that learns a distribution-invariant offloading policy to achieve zero-shot generalization. Our method dynamically identifies task distribution shifts as confounding factors via a self-attention-based Soft Mask. Then, we apply causal interventions using these confounders to disentangle their effect on the policy, allowing the model to learn an invariant policy across different task distributions. Simulation results demonstrate that the proposed method achieves strong generalization performance in zero-shot settings across multiple heterogeneous MEC task offloading scenarios.