Performance of ML-Based Channel Prediction Algorithms for URLLC: Channel Model Matters

Authors:

Abstract
Massive Multiple-Input Multiple-Output (MMIMO) is a key technology for boosting the performance of 5G systems. The efficiency of M-MIMO depends on the knowledge of Channel State Information (CSI), which quickly becomes outdated in high-mobility scenarios. To address this problem, many papers propose channel prediction algorithms that forecast CSI. Such algorithms are especially crucial for enabling Ultra-Reliable Low Latency Communications (URLLC). Unfortunately, many existing algorithms are validated using different channel models and assumptions, which complicates their analysis and comparison. The paper evaluates the performance of the modern machine learning-based channel prediction algorithms using the state-of-the-art geometry-based stochastic channel model with various approaches for modeling channel evolution. The results show that the used channel model does significantly affect the performance of the considered algorithms. For the simplified channel models, the considered algorithms demonstrate too optimistic results, while for more realistic channel models, the performance significantly degrades and even becomes worse than without prediction for high speeds. Among ML-based algorithms, autoregressive algorithms show much better performance than neural network-based algorithms because they use different approaches to update their internal parameters with time. Thus, to boost the performance of the neural network-based algorithms, new computationally effective and robust fine-tuning methods are required.

This publication has 26 references indexed in Scilit: