An Incremental Learning-Based Mechanism to Deploying Radio Map Estimation Models

Radio map estimation is a key technology in wireless communications and is critical for network planning, location-based services, and resource management. However, the deployment of deep learning based radio map estimation models in new data-sparse scenarios faces challenges of data scarcity and poor continuous model adaptation. To address these issues, this paper proposes a three-step deployment framework based on Incremental Learning. The framework first provides a priori knowledge for the deep learning model through offline training using data generated by the experience propagation model. Then, the pre-trained model is deployed online in a new scenario for initial prediction; and, the model is efficiently and continuously optimized and adaptively tuned in the feedback updating phase using a small amount of real-world measurement data continuously arriving in the new scenario through our proposed incremental learning mechanism. Our mechanism is designed to efficiently incorporate new knowledge and retain old knowledge, thus mitigating the catastrophic forgetting problem. We evaluate the performance of the framework on the public dataset. The experimental results show that, the incremental learning method proposed in this paper can reduce the RMSE from 6.2 dB to 5.1 dB.