PaperlayerView exact revision on arXiv

Source-linked AI summary

RAG-CSI: Global Local Retrieval via Joint State and Dynamics Matching for CSI Prediction

Xinyue Jiang, Tong Li, Shiceng Zhang, Ke Chen, Zhu Xiao

arXiv:2610.36881v1eess.SPcs.LG

TL;DR

CSI prediction must forecast future channel states from historical observations, while existing deep learning predictors can be limited in representing diverse dynamics across scenes. RAG-CSI retrieves historical channel knowledge using a global–local store and joint state–dynamics retrieval, then fuses it with current CSI history. On a multi-scene dataset, it improves on the best baseline across all five metrics, including a 15.42% reduction in NMSE.

  • Problem

    Existing deep learning predictors mainly encode channel knowledge in model parameters, which can limit representation of diverse channel dynamics across scenes.

    Source: §I · p. 1

  • Method

    RAG-CSI combines global and local CSI stores with joint state–dynamics retrieval and cross-attention fusion in a Transformer predictor.

    Source: §I · pp. 1–2, A. Overview · p. 2

  • Results

    RAG-CSI reduces NMSE by 15.42% versus the best baseline and improves by more than 10% over the best baseline across all five metrics.

    Source: §I · p. 2, C. Overall Performance · p. 5

  • Takeaways & Limitations

    Retrieved historical CSI provides prediction context beyond the current observation window, reducing reliance on channel knowledge encoded solely in model parameters.

    Source: §I · p. 2, §VI · p. 6

Abstract

from arXiv · show

Channel state information (CSI) prediction forecasts future channel states from historical observations, supporting transceiver design and resource allocation with reduced measurement overhead. However, existing deep learning predictors mainly encode channel knowledge in model parameters, which can limit their ability to represent diverse patterns of channel dynamics across scenes. In this paper, we propose RAG-CSI, a retrieval-enhanced Transformer framework that augments prediction with relevant historical channel knowledge. Specifically, we construct a global--local CSI knowledge store containing representative observations across scenes and historical observations from the current serving scene. We further develop joint state--dynamics retrieval (JSDR), which jointly characterizes channel states, phase dynamics, and amplitude dynamics to retrieve relevant CSI sequences. A Transformer predictor then uses cross-attention to fuse the retrieved knowledge with the current CSI history and predict future CSI with limited additional model parameters. Experiments on a multi-scene CSI dataset show overall improvements exceeding 10\% across all five evaluation metrics over the best baseline for each metric, including a 15.42\% reduction in normalized mean squared error.

I. INTRODUCTION

CSI prediction uses historical observations to forecast future channel states, but parameter-based deep learning can struggle to represent diverse dynamics across scenes. RAG-CSI adds retrieved historical knowledge through a global–local store and joint state–dynamics retrieval, then fuses it with the input history.

I. INTRODUCTION

  • Parameter-based deep learning methods may have limited capacity to represent diverse channel dynamics across scenes, while scene differences and short observation windows further complicate prediction.CSI sequences with similar overall values can still differ in phase directions and amplitude trends.Source: §I · p. 1
  • RAG-CSI combines representative historical CSI across scenes in a global store with observations from the current serving scene in a local store.The global store is maintained on the user side and the local store at the serving base station.Source: §I · p. 1
  • JSDR retrieves references using channel states, phase dynamics, and amplitude dynamics, and a Transformer fuses retrieved knowledge with the input history through cross-attention.The framework supplements the current observation window with historical channel knowledge and aims to reduce reliance on patterns encoded solely in model parameters.Source: §I · p. 2
  • RAG-CSI outperforms state-of-the-art baselines by more than 10% across MSE, NMSE, MAE, NMAE, and cosine similarity ρ on a multi-scene CSI dataset.The reported comparison covers all five evaluation metrics.Source: §I · p. 2

A. Channel Model

The channel model represents each subcarrier's channel as a sum of propagation paths, with array response determined by frequency and departure direction. The problem definition frames prediction as estimating future CSI from historical observations, optionally augmented with retrieved channel knowledge.

A. Channel Model

  • The channel between the base station and one user is modeled as a superposition of P propagation paths, each with a complex gain, delay, and departure direction.The channel vector for each subcarrier is assembled from these path contributions and the array response.Source: A. Channel Model · p. 2
  • The array response depends on subcarrier frequency and departure direction; for a uniform linear array it reduces to a phase progression determined by antenna spacing and angle of departure.Stacking subcarrier channel vectors yields the complete CSI matrix.Source: A. Channel Model · p. 2

B. Problem Definition

  • CSI snapshots generated under specified configurations are ordered along simulated user trajectories, and prediction estimates future CSI from historical observations.The history and prediction lengths are denoted by L and Q, respectively.Source: B. Problem Definition · p. 2
  • A conventional predictor uses only observed history, whereas RAG-CSI additionally retrieves channel knowledge from a CSI memory and trains the predictor by minimizing prediction error.R(·) denotes retrieval, K_t the retrieved channel knowledge, and L(·, ·) the prediction error.Source: B. Problem Definition · p. 2

A. Overview

RAG-CSI augments Transformer-based CSI prediction with retrieved historical channel knowledge. Its global and local stores provide complementary references, which JSDR retrieves for the current history.

A. Overview

  • RAG-CSI retrieves relevant sequences from global and local stores, then fuses that knowledge with the historical representation for Transformer prediction.The Global store holds representative CSI across scenes; the Local store holds CSI from the current serving scene.Source: A. Overview · pp. 2–3

B. Retrieval-Augmented Generation (RAG) for CSI Prediction

  • RAG uses external memory to provide reusable historical channel patterns beyond the current observation window for CSI prediction.The current CSI history serves as the query, and retrieved references are incorporated into the Transformer.Source: B. Retrieval-Augmented Generation… · p. 3

1) Global–Local CSI Knowledge Store:

C. Joint State–Dynamics Retrieval (JSDR)

JSDR retrieves CSI knowledge by jointly representing channel state and dynamics. Its state descriptor temporally weights real and imaginary CSI components before normalization.

C. Joint State–Dynamics Retrieval (JSDR)

  • JSDR addresses the limits of direct cosine or Euclidean retrieval by jointly representing channel state, phase dynamics, and amplitude dynamics.The descriptors are normalized and combined into a unified retrieval key.Source: C. Joint State–Dynamics Retrieval (JSDR) · p. 3

1) CSI State Representation:

  • The CSI state descriptor vectorizes real and imaginary components of historical CSI and concatenates observations in chronological order with temporal weights.The descriptor is ℓ2-normalized; smaller γ emphasizes recent observations, while γ = 1 gives uniform weights.Source: 1) CSI State Representation · p. 3
  • The same state representation is applied to query and stored CSI sequences.Source: 1) CSI State Representation · p. 3

2) Phase-Dynamics Representation:

The dynamics descriptors encode changes over time separately from CSI state: phase changes use normalized conjugate products, while amplitude changes use logarithmic magnitude differences.

2) Phase-Dynamics Representation:

  • Normalized complex conjugate products at temporal lags retain relative phase changes while removing magnitudes, avoiding phase-wrapping discontinuities.The phase descriptor includes valid observation pairs at selected lags and applies temporal weighting, vectorization, concatenation, and ℓ2 normalization.Source: 2) Phase-Dynamics Representation · p. 4

3) Amplitude-Dynamics Representation:

  • Logarithmic magnitude differences between consecutive observations capture relative amplitude changes while reducing sensitivity to absolute channel magnitudes.The amplitude descriptor is temporally weighted, vectorized, concatenated, and ℓ2-normalized.Source: 3) Amplitude-Dynamics Representation · p. 4

4) Joint State–Dynamics Retrieval Key:

The retrieval key combines CSI state, phase dynamics, and amplitude dynamics, with nonnegative weights controlling their contributions. A lightweight Transformer uses retrieved sequences through cross-attention to predict future CSI.

4) Joint State–Dynamics Retrieval Key:

D. Retrieval-Augmented Prediction

  • Cross-attention fuses current CSI history with retrieved-sequence representations, and the Transformer models temporal dependencies to generate future-channel features.A linear projection and reshaping produce the predicted CSI sequence's real and imaginary components.Source: D. Retrieval-Augmented Prediction · p. 4

IV. EXPERIMENTS

The evaluation uses simulated CSI sequences from multiple propagation scenes, comparing RAG-CSI with six representative predictors using five performance metrics.

A. Experimental settings

  • DeepMIMO v4 samples cover 10 scenes, with 12 historical CSI observations used to predict one future step.The generated sequences are split into training and test sets at an 8:2 ratio; retrieval defaults are K = 10 and λ = 0.5.Source: A. Experimental settings · p. 4

B. Baselines and Evaluation Metrics

  • RAG-CSI is compared with six baselines, including recurrent, Transformer, channel-deduction, and wireless foundation-model predictors.The listed baselines are LSTM, LSTM-GRU, Transformer, RCDNet, WiFo, and PromptFM.Source: B. Baselines and Evaluation Metrics · p. 4
  • Performance is evaluated with MSE, NMSE, MAE, NMAE, and cosine correlation ρ.Lower error values and higher ρ indicate better prediction.Source: B. Baselines and Evaluation Metrics · p. 4

C. Overall Performance

RAG-CSI outperforms the best baseline across all five metrics, with gains across four scenes. Ablations and sensitivity tests examine retrieval components and Global–Local balance.

C. Overall Performance

  • 14.36% lower MSE, 15.42% lower NMSE, 10.56% lower MAE, 10.13% lower NMAE, and 11.72% higher cosine correlation ρ than the best baseline for each metric.RAG-CSI achieves the best results across all five metrics.Source: C. Overall Performance · p. 5

D. Performance Across Scenes

  • Across four scenes, RAG-CSI reduces NMSE by 4.17%–12.74% and NMAE by 4.93%–7.76%, while improving ρ by 4.06%–12.06% against the best baseline for each scene and metric.The largest gains are observed in Miami.Source: D. Performance Across Scenes · p. 5

E. Ablation Study

  • Removing retrieval raises NMSE from 0.6172 to 0.7344 and lowers ρ from 0.7949 to 0.6351; replacing JSDR raises NMSE to 0.6783 and lowers ρ to 0.6844.The variants remove CSI sequence retrieval or replace JSDR with cosine-correlation retrieval, respectively.Source: E. Ablation Study · p. 5

F. Hyperparameter Sensitivity

  • NMSE falls markedly as K increases to 8, with only marginal gains beyond 8; λ = 0.5 performs best, while shifting toward either store increases error.The sensitivity tests use λ = 0.5 when varying K and K = 10 when varying λ; jointly using Global and Local knowledge balances diverse patterns with current-scene relevance.Source: F. Hyperparameter Sensitivity · p. 5, Figure 3 · p. 5

G. Complexity Analysis

RAG-CSI pairs a compact prediction network with retrieval-augmented CSI context. The paper situates retrieval as a way to access channel knowledge beyond fixed model parameters.

G. Complexity Analysis

  • RAG-CSI uses 0.451 million parameters, 50.39% fewer than the Transformer baseline, and takes 0.3242 ms per sequence, 65.72% less time than that baseline.Its inference time is third shortest overall, behind LSTM and LSTM-GRU.Source: G. Complexity Analysis · p. 5

V. RELATED WORK

  • Prior CSI predictors primarily encode channel knowledge in fixed model parameters, while retrieval augmentation can provide historical patterns as additional prediction context.The related-work discussion notes that retrieval augmentation remains largely unexplored for CSI prediction.Source: §V · p. 5

VI. CONCLUSION

  • RAG-CSI combines a Global–Local knowledge store, JSDR, and cross-attention fusion; on a multi-scene dataset, it improves over the best baselines by more than 10% across all five metrics.JSDR jointly captures channel states, phase dynamics, and amplitude dynamics to retrieve informative CSI sequences.Source: §VI · pp. 5–6

Related papers

Loading 2610.36881v1…