Fine-Tuning Approach to Configuration and Data Selection for Path Loss Prediction in Different Geographical Environments

In Beyond 5G/6G, research is being conducted on wireless emulators that can enable low-cost, short-term evaluation and verification of wireless systems by emulating radio systems in cyberspace. To accurately simulate the behavior of wireless systems in various scenarios, radio propagation must be simulated with high accuracy according to the environment. We developed a site-specific path loss model by using machine learning in conjunction with spatial information data and environmental parameters related to propagation characteristics. However, when a learning model based on data obtained in a specific environment is applied to a new environment, the similarity of spatial information data is low because of differences in urban structures, and sufficient estimation accuracy cannot be obtained. In this paper, we propose a fine-tuning method that transfers a machine learning model constructed in a specific environment to a new environment and refines it through an effective data selection technique. Evaluating the measurement data at 2 GHz revealed that the proposed method achieves higher estimation accuracy than does the conventional method when a model pretrained in a specific environment in an urban area is applied to other urban and suburban areas, with only 1% of the test data required as additional training data.