Retraining-Free Cross-Environment Adaptation in Mobile Networks via a Learned-Gating Mixture-ofExperts: Application to UL Throughput Prediction

Abstract

This paper proposes a learned-gating Mixture-of-Experts (MoE) framework that integrates multiple pretrained throughput prediction models without retraining for crossenvironment adaptation in commercial mobile networks. The framework retains global experts and cluster-specific experts trained in urban and suburban areas, and applies them to a residential area as the unseen domain. Based on a 14-dimensional mean feature vector extracted from each cellsegment prefix, the prediction error of each expert is estimated and the expert outputs are combined through softmax-based weighting. Evaluation using commercial uplink measurement data showed that the proposed method achieved overall MAEs of 1.61, 1.99, and 2.08 Mbps for 1-, 2-, and 3-s-ahead predictions, respectively, outperforming a single expert and fixed-weight schemes. Furthermore, the ablation study confirmed both the effectiveness of soft-gating and the practicality of the lightweight design using mean-only features. These results demonstrate that the proposed method is an effective adaptation mechanism for mitigating performance degradation caused by environmental differences in practical deployment scenarios where inferencetime retraining is difficult.