A Deep Incremental Framework for Multi-Service Multi-Modal Devices in NextG AI-RAN Systems

TL;DR

This work tackles the challenge of supporting multi-service-modal UEs (MSMU) in NextG AI-RAN by jointly managing traffic forecasting, route selection, RAN slicing, service identification, and radio resource management under uncertainty. It introduces a two-scale decomposition: a long-term subproblem (L-SP) solved by a ReVIN-enabled Transformer for traffic-demand and routing forecasting, coupled with heuristic radio-slicing, and a short-term subproblem (S-SP) solved by a continual-learning LSTM for real-time service identification and resource allocation. A continual-learning mechanism with replay buffers mitigates forgetting when new service subclasses appear, enabling robust adaptation across seven tasks. Empirical results show substantial improvements in prediction accuracy (e.g., up to a 46.86% reduction in traffic-demand error), PRB and power estimation (26.70% and 18.79%), and near-perfect service/route identification (≈99%), with over 95% continual-learning accuracy, highlighting practical viability for dynamic NextG AI-RAN deployments.

Abstract

In this paper, we propose a deep incremental framework for efficient RAN management, introducing the Multi-Service-Modal UE (MSMU) system, which enables a single UE to handle eMBB and uRLLC services simultaneously. We formulate an optimization problem integrating traffic demand prediction, route optimization, RAN slicing, service identification, and radio resource management under uncertainty. We decompose it into long-term (L-SP) and short-term (S-SP) subproblems then propose a Transformer model for L-SP optimization, predicting eMBB and uRLLC traffic demands and optimizing routes for RAN slicing. To address non-stationary network traffic with evolving trends and scale variations, we integrate reversible instance normalization (ReVIN) into the forecasting pipeline. For the S-SP, we propose an LSTM model enabling real-time service type identification and resource management, utilizing L-SP predictions. We incorporate continual learning into the S-SP framework to adapt to new service types while preserving prior knowledge. Experimental results demonstrate that our proposed framework achieves up to 46.86% reduction in traffic demand prediction error, 26.70% and 18.79% improvement in PRBs and power estimation, 7.23% higher route selection accuracy, and 7.29% improvement in service identification over the baselines with 95% average accuracy in continual service identification across seven sequential tasks.

Figures (19)

• Figure 1: System model for multi-service-modal UE (MSMU) management within the AI-RAN framework. Each UE supports both eMBB and uRLLC services, with data buffered at the RU, DU, and CU in device-service-specific buffers. Besides AI capabilities in the RIC (as in O-RAN), the system model embeds AI functionalities directly into network components such as the CU and DU.
• Figure 2: proposed mixed numerology in frequency domain, multi-service-modal UE (MSMU) over service dedicated UE and proposed service switching mechanism. The proposed service-switching mechanism allows the same UE to dynamically switch between different services at each mini time frame, enabling support for multiple services in rapid succession.
• Figure 3: System model illustrating the continual arrival and adaptation of sub-services under eMBB and uRLLC superclasses. As new sub-services such as video streaming, autonomous driving, cloud gaming, remote surgery etc. emerge over time, they must be accurately classified under the appropriate superclass to enable proper network policy enforcement. Continual learning supports this process by adapting to new service patterns without forgetting previously learned classes.
• Figure 4: Proposed intelligent framework for managing multi-service-modal UEs (MSMUs) within the AI-RAN ecosystem. The L-SP is handled in the non-RT RIC through two rApps: rApp1 performs Transformer-based demand and route prediction, while rApp2 applies heuristic-based radio resource slicing. The S-SP is addressed in the near-RT RIC via xApp1, which uses a continual learning-enhanced LSTM model. Queue management decisions are executed at the CU, leveraging the embedded AI/ML capabilities of the AI-RAN architecture.
• Figure 5: (a) Overview of the process for computing single-head attention scores; (b) visualization of the single-head attention mechanism; (c) illustration of the multi-head attention architecture, adapted from advancing.