Abstract:Millimeter-wave (mmWave) communication is promising for next-generation wireless networks but suffers from significant path loss, requiring extensive antenna arrays and frequent beam training. Traditional deep learning models, such as long short-term memory (LSTM), enhance beam tracking accuracy however are limited by poor robustness and generalization. In this letter, we use large language models (LLMs) to improve the robustness of beam prediction. By converting time series data into text-based representations and employing the Prompt-as-Prefix (PaP) technique for contextual enrichment, our approach unleashes the strength of LLMs for time series forecasting. Simulation results demonstrate that our LLM-based method offers superior robustness and generalization compared to LSTM-based models, showcasing the potential of LLMs in wireless communications.
What problem does this paper attempt to address?
This paper attempts to solve the beam prediction problem in millimeter - wave (mmWave) communications. Specifically, although millimeter - wave communications offer great potential for bandwidth in next - generation wireless networks, their significant path loss requires large - scale antenna arrays and frequent beam training. Traditional deep - learning models such as Long - Short - Term Memory networks (LSTM) can improve the accuracy of beam tracking, but have poor robustness and generalization ability in different wireless environments.
To overcome these challenges, this paper proposes a beam prediction method based on large - language models (LLMs). By converting time - series data into text - based representations and using Prompt - as - Prefix (PaP) technology for context enhancement, this method can fully utilize the advantages of LLMs in time - series prediction. Experimental results show that the LLMs - based method is superior to the LSTM - based model in terms of robustness and generalization ability, demonstrating the potential of LLMs in the field of wireless communications.
### Key Formulas
1. **Channel Vector Expression**:
\[
h_n=\sum_{l = 1}^{L_n}\frac{s}{\rho_{n,l}}\alpha_{n,l}a^*(\varphi_{n,l})
\]
where \(L_n\) represents the number of paths in the \(n\) - th time slot, \(\rho_{n,l}\) is the path loss, \(\alpha_{n,l}\) is the complex gain, \(\varphi_{n,l}\) is the Angle of Departure (AoD), and \(a(\varphi)\) is the antenna response vector.
2. **Antenna Response Vector**:
\[
a(\varphi)=\left[1, e^{j2\pi d\sin\varphi / \lambda},\ldots,e^{j\pi(M - 1)d\sin\varphi / \lambda}\right]^T
\]
where \(\varphi\) is the AoD, and \(d\) and \(\lambda\) are the antenna spacing and wavelength respectively.
3. **Candidate Transmit Beams**:
\[
f^{(q)}=\sqrt{\frac{1}{M}}\left[1, e^{j2\pi q / Q},\ldots,e^{j2\pi(M - 1)q / Q}\right]^T
\]
4. **Optimal Beam Selection**:
\[
q^*=\arg\max_{q\in\{0, 1, 2,\ldots,Q - 1\}}\|h^T f^{(q)}\|^2
\]
5. **Maximization of Normalized Beamforming Gain**:
\[
\max\frac{|h^T f(\hat{q}^*)|^2}{|h^T f(q^*)|^2}
\]
6. **Loss Function**:
\[
L=\frac{1}{H}\sum_{h = 1}^H\|\hat{Y}_h - Y_h\|^2_2
\]
Through these formulas, the paper elaborates on how to use LLMs for beam prediction, thereby improving the robustness and generalization ability of millimeter - wave communication systems.