PaperlayerView exact revision on arXiv

Source-linked AI summary

Multimodal Large Language Model-guided Constrained Optimization for RAN Intelligent Control

Hyeonho Noh

arXiv:2609.06122v1eess.SYcs.SY

TL;DR

RAN controllers often assume predefined operating scenarios and optimization tasks, despite changes in network conditions and operator requirements after deployment. MLLM-coRIC uses a hierarchical controller to adapt the loss and execute resource and power allocation; evaluations report effective constrained solutions across different problems and loads with few refinement rounds.

  • Problem

    Many RAN controllers are designed for predefined operating scenarios and fixed optimization criteria, while deployed networks and operator requirements can change.

    Source: §I · p. 1, B. Challenges · p. 2

  • Method

    MLLM-coRIC uses operator requirements, predicted network context, and measured outcomes to refine a numerical loss for a Near-RT hybrid resource and power executor.

    Source: C. Contributions · p. 2, B. Non-RT rApp: Requirement… · p. 5, C. Near-RT xApp: Loss-Conditioned… · p. 7

  • Results

    The evaluation reports effective constrained solutions across different optimization problems and network loads with only a small number of MLLM-guided refinement rounds.

    Source: §V · p. 12

  • Takeaways & Limitations

    A single deployed controller can accommodate changing network conditions and operator-defined optimization problems without per-request retraining.

    Source: C. Contributions · p. 2

Abstract

from arXiv · show

Artificial intelligence (AI)-based radio access network (RAN) controllers are commonly designed for predefined operating scenarios and optimization tasks, limiting their adaptability when network conditions and operator requirements change after deployment. This paper proposes multimodal large language model (MLLM)-guided constrained optimization for RAN intelligent control (MLLM-coRIC), a requirement-adaptive hierarchical Open RAN (O-RAN) framework for joint resource and power allocation. MLLM-coRIC jointly exploits the operator's natural-language specification and radio-frequency (RF)-derived network context as multimodal inputs, enabling a single deployed framework to address different optimization problems without redesigning task-specific algorithms or retraining control policies. At the Non-Real-Time RAN Intelligent Controller (Non-RT RIC), the MLLM performs holistic, longer-timescale reasoning over the operator requirement, predicted network evolution, and measured optimization outcomes to design the numerical control loss. The relative constraint penalties are iteratively refined through a closed-loop process, allowing the optimization criterion to reflect both the intended network behavior and the expected operating context. At the Near-Real-Time RIC (Near-RT RIC), a loss-conditioned hybrid executor translates the synthesized criterion into fast radio actions by combining learned ramp-constrained resource allocation with model-based interference-aware power control. A multi-cell evaluation environment integrating the CARLA urban mobility simulator and the Sionna RT ray-tracing-based wireless propagation simulator is implemented for validation.

I. INTRODUCTION

Existing RAN controllers are often tied to predefined scenarios and optimization criteria, while O-RAN separates slower policy generation from fast radio control. MLLM-coRIC uses that hierarchy to adapt the optimization problem and its execution to changing requirements and network conditions.

I. INTRODUCTION

  • O-RAN separates slower, network-wide policy generation from latency-sensitive radio control based on live measurements.The Non-RT RIC can use historical and network-wide observations, while the Near-RT RIC handles fast tasks such as resource allocation and power control.Source: §I · p. 1

A. Related Works

  • Prior work spans O-RAN control, environment-adaptive optimization, and LLM-based control, but dynamic construction of numerical loss from changing operator requirements and network conditions remains unexplored.The cited discussion distinguishes environment generalization from adapting the optimization criterion itself.Source: A. Related Works · pp. 1–2

B. Challenges

  • Existing controllers often specialize in a predefined environment or optimization criterion, motivating a unified RIC architecture that can adapt both after deployment.Related approaches may generalize across network conditions while retaining a fixed optimization problem.Source: B. Challenges · p. 2, A. Related Works · p. 2

C. Contributions

  • MLLM-coRIC treats the operator-defined optimization problem as a runtime input and uses language and RF-derived context to support resource and power allocation without per-request redesign or retraining.The Non-RT MLLM interprets requirements and predicted context to construct and refine the loss; the Near-RT executor applies it.Source: C. Contributions · pp. 2–3
  • The Near-RT executor combines learned ramp-constrained resource allocation with model-based, interference-aware power control.This assigns sequential adaptation to learning and instantaneous power coordination to analytical optimization.Source: C. Contributions · p. 3
  • Evaluations across user loads and operator-defined objectives, constraints, and thresholds report adaptation to different conditions and problems while satisfying service requirements without retraining.The evaluation environment integrates CARLA urban mobility simulation with Sionna RT wireless propagation simulation.Source: C. Contributions · p. 3

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system model divides control between slow network-wide orchestration and faster cell-level resource and power decisions. It models co-channel cells, traffic and radio conditions, power states, and a ramp limit on PRB provisioning.

A. Preliminary: O-RAN Control Hierarchy

  • The Non-RT RIC sends policy or enrichment information over A1, while the Near-RT RIC uses live E2 measurements to set cell-level PRB budgets and per-PRB power.Instantaneous UE-level PRB scheduling remains with the O-DU.Source: A. Preliminary: O-RAN Control Hierarchy · p. 3

B. Communication Model

  • The modeled network has co-channel cells serving time-varying UEs, with channel gain reflecting path loss and shadowing and UE demand set by its attached service.UE association is based on strongest channel.Source: B. Communication Model · p. 3
  • The Near-RT RIC selects each cell’s provisioned PRB budget and per-PRB transmit power, which affect capacity, provisioning cost, received signal, and inter-cell interference.UE-level allocations are bounded by the associated cell’s provisioned budget.Source: B. Communication Model · pp. 3–4
  • The communication model derives PRB utilization, interference-dependent SINR, capped spectral efficiency, and delivered UE rate from allocations and radio conditions.The utilization factor captures the share of the PRB pool over which another cell contributes interference.Source: B. Communication Model · p. 4
  • A fixed demand-aware round-robin O-DU scheduler is treated as underlying operation, while cell power accounts for state, provisioning, and PRBs actually carrying traffic.The power model follows the EARTH short-sleep model and 3GPP cell discontinuous transmission.Source: B. Communication Model · p. 4
  • Cells may be active, light, or deep; increasing provisioned PRBs can prewarm a cell at added provisioning power, while RF cost applies to utilized PRBs.Active cells serve traffic or remain sufficiently provisioned; light and deep cells serve no traffic, with deep cells not maintained provisioned.Source: B. Communication Model · p. 4
  • Provisioned PRB budgets are ramp-limited between epochs because bandwidth increases incur wake-up delay, whereas transmit power can change within an epoch.This actuation difference motivates forecast-aided resource allocation and power allocation responsive to current channel and interference.Source: B. Communication Model · p. 4

C. Problem Formulation

The formulation makes the operator’s objective and service constraints variable across control episodes. The RIC chooses per-cell PRB budgets and transmit powers, subject to service and actuation constraints, in a stochastic, coupled problem.

C. Problem Formulation

  • The operator specifies a potentially changing network objective and service constraints, and the RIC chooses per-cell PRB budgets and transmit powers.The problem is stated as minimizing over the resource and power allocation variables.Source: C. Problem Formulation · p. 4
  • The service constraints are time-averaged, with the averaging population set to one for per-cell metrics or the active UE count for per-UE metrics.The formulation also bounds PRB budgets and transmit power and limits per-epoch PRB-budget variation.Source: C. Problem Formulation · p. 4
  • The resulting control problem is mixed-integer, non-convex, temporally coupled, and stochastic because of discrete PRB budgets, inter-cell interference, scheduling effects, ramp limits, and evolving conditions.These factors make network performance depend non-smoothly on the control variables.Source: C. Problem Formulation · p. 5

III. PROPOSED LANGUAGE-GUIDED HIERARCHICAL RESOURCE AND POWER ALLOCATION

The Non-RT RIC turns operator requirements and RF context into a requirement-dependent control loss and load forecast for Near-RT control. It refines loss penalties using measured optimization outcomes.

A. Overview: Hierarchical Decomposition of P1

  • The Non-RT RIC sends a synthesized loss specification and load forecast to the Near-RT RIC, which combines them with live E2 measurements to set per-cell PRB budgets and transmit powers.The A1 interface carries the enrichment from Non-RT to Near-RT control.Source: A. Overview: Hierarchical… · p. 5

B. Non-RT rApp: Requirement Compilation, Forecasting, and Measured Loss Synthesis

  • Recent per-cell RF measurements summarize offered load, UE count, serving-link SINR, mobility, and cross-cell channel gains for forecasting future cell loads.A GRU forecaster predicts expected loads and their uncertainty from RF history.Source: B. Non-RT rApp: Requirement… · p. 5
  • The MLLM compiles the operator’s natural-language request into a typed specification of the objective, optimization direction, active constraints, and thresholds.A positive signed constraint deviation indicates a violation.Source: B. Non-RT rApp: Requirement… · p. 5
  • A context-dependent MLP initializes loss coefficients, while the MLLM refines them using measured objective and constraint outcomes from evaluated candidates.The MLLM also adjusts a forecast-guided resource floor for delivery to the Near-RT RIC.Source: B. Non-RT rApp: Requirement… · pp. 5–6
  • The refinement loop samples candidate coefficients, evaluates them with the network toolkit, and uses MLLM-provided relative multipliers to recenter the distribution for another round.Only the finalized coefficients and resource floor leave the loop.Source: Figure 2 · p. 6, B. Non-RT rApp: Requirement… · p. 6

1) Tokenization and embedding:

The requirement, network context, and measured outcomes are serialized as a token sequence and embedded into a representation for transformer processing.

1) Tokenization and embedding:

  • The requirement, network context, and current-round measured outcomes are serialized into a token sequence using a prompt template.Source: 1) Tokenization and embedding · p. 6
  • Each token is embedded with E, and the resulting token embeddings are stacked into a latent representation matrix.Source: 1) Tokenization and embedding · p. 6
  • The representation is processed by sequential transformer blocks, with multiple attention heads operating in parallel within each block.Source: 2) Backbone · p. 6

2) Backbone:

The backbone applies multi-head attention and feed-forward processing in repeated transformer blocks to produce a final hidden representation.

2) Backbone:

  • Each attention head forms query, key, and value projections using trainable matrices, with projection dimension d/H.Source: 2) Backbone · p. 6
  • Attention uses row-wise softmax with a causal mask that excludes future token positions.Source: 2) Backbone · p. 6
  • The head outputs are combined through an output projection, and repeated block operations yield the final hidden representation.Source: 2) Backbone · p. 6

3) Autoregressive decoding:

The language-model head autoregressively generates structured updates to candidate loss coefficients and the provisioning floor. Repeated evaluation and refinement produce the finalized values delivered to the Near-RT RIC.

3) Autoregressive decoding:

  • The pretrained language-model head decodes the final transformer representation autoregressively, selecting each next token from the current hidden representation.Source: 3) Autoregressive decoding · p. 7
  • The fixed-format generated sequence specifies multiplicative factors for constraint coefficients and the forecast-guided provisioning floor, rather than directly generating coefficients or radio actions.Source: 3) Autoregressive decoding · p. 7
  • The MLLM’s factors update the best-performing candidate coefficients and the provisioning floor; candidate sampling and refinement repeat for Nref rounds before delivery to the Near-RT RIC.Source: 3) Autoregressive decoding · p. 7

C. Near-RT xApp: Loss-Conditioned Hybrid Executor and Training

The Near-RT RIC uses current measurements and Non-RT guidance to select resource and power allocations, then trains the resource allocator across requirements using a loss surrogate and budget-variation regularization.

C. Near-RT xApp: Loss-Conditioned Hybrid Executor and Training

  • The Near-RT RIC sets a PRB budget from load forecasts and the requirement-dependent loss, then optimizes transmit power using current channel and interference conditions.It uses current E2 measurements to translate Non-RT guidance into these decisions.Source: C. Near-RT xApp: Loss-Conditioned… · p. 7
  • Candidate PRB budgets are formed from prior-epoch measured demand within a forecast-guided provisioning range, then evaluated jointly with their power allocations under the requirement-dependent loss.The final resource–power pair is selected after evaluating the candidate budgets.Source: C. Near-RT xApp: Loss-Conditioned… · p. 7
  • The power-allocation problem is approximately solved by block-coordinate descent, updating one cell’s transmit power at a time until the allocation converges.Each update minimizes the loss while holding the other cell powers fixed.Source: C. Near-RT xApp: Loss-Conditioned… · p. 7
  • Offline training minimizes a differentiable surrogate of the requirement-dependent loss plus a regularizer on inter-epoch PRB-budget variation, while treating the power solution as fixed during resource-allocator updates.The requirements are sampled over their admissible range, and rollouts follow ramp-constrained budget dynamics.Source: C. Near-RT xApp: Loss-Conditioned… · p. 8

A. Evaluation Setup

In evaluation, MLLM-coRIC addresses different constrained optimization requests and user loads within one controller. It achieves low-power feasible operation under QoS constraints and low QoS violations near power limits.

A. Evaluation Setup

  • The evaluation uses a three-cell CARLA Town05 urban-road deployment with a moving vehicle convoy, stationary pedestrian traffic, and heterogeneous service demands.Convoy size varies from 5 to 25 vehicles; the radio setup uses a 3.5 GHz 3GPP urban-micro channel.Source: A. Evaluation Setup · p. 8
  • The comparisons include Lyapunov drift-plus-penalty control, demand-based greedy provisioning, and end-to-end deep reinforcement learning for joint PRB and power control.The three baselines use distinct control approaches.Source: A. Evaluation Setup · p. 8

B. Problem-Solving Capability

  • At 35 users, relaxing V from 1% to 5% raises MLLM-coRIC’s QoS violation rate from 0.79% to 4.99%, while it still achieves the lowest network power.Under P2, it is the only method satisfying the stringent 1% constraint; Fig. 4 compares network power and violation rate at 25 and 35 users.Source: Figure 4 · p. 8, B. Problem-Solving Capability · p. 9
  • Under P3, MLLM-coRIC achieves the lowest QoS violation rate near the prescribed power limit, staying within 1.5 W and 1.2 W of the limits at 35 and 25 users, respectively.Fig. 5 presents the QoS violation rate and network power under the power constraint.Source: B. Problem-Solving Capability · p. 9, Figure 5 · p. 9
  • Across the tested user loads under P2 with V = 1%, MLLM-coRIC maintains feasibility and the lowest network power, while all three baselines violate the QoS constraint at 35 users.Fig. 6 plots network power and QoS violation rate against user count.Source: B. Problem-Solving Capability · p. 9, Figure 6 · p. 9

C. Adaptation to Dynamic Environment and Operator Requirements

MLLM-coRIC handles changing network loads and operator requirements across four phases without retraining or manual retuning. Under comparable power, forecast-aided provisioning tracks requested traffic more closely as users move between cells.

C. Adaptation to Dynamic Environment and Operator Requirements

  • The four-phase sequence holds the trajectory, channel model, interference coupling, A1 cadence, and power model fixed, isolating changes to network load and operator requirements.Phase II increases load; Phase III exchanges the objective and constraint; Phase IV restores power minimization with a relaxed QoS requirement.Source: C. Adaptation to Dynamic Environment… · p. 9
  • Across four phases, MLLM-coRIC adapts to changed loads and operator requests without retraining or manual retuning, while maintaining a better stated-objective operating point than the retrained DRL baseline.The phases include increased load, a change from power minimization to QoS minimization, and a return to power minimization with a relaxed QoS requirement.Source: C. Adaptation to Dynamic Environment… · pp. 9–10, Figure 7 · p. 10

D. Forecast-Aided Proactive Provisioning

  • At comparable network power, MLLM-coRIC tracks requested traffic more closely as users move across cells, producing substantially fewer QoS violations than the baselines.Forecasting enables advance provisioning; reactive allocation may lag rapid demand increases because of the PRB ramp constraint.Source: D. Forecast-Aided Proactive Provisioning · p. 10

E. Ablation Study

The ablations indicate that adaptive loss synthesis and Near-RT execution play complementary roles, while the hybrid executor performs favorably across the constraint sweep. A small number of MLLM refinements suffices for a feasible, efficient criterion, and larger-scale and real O-RAN deployments remain future work.

E. Ablation Study

  • Fixed loss coefficients are overly conservative under lighter loads but insufficient under heavy loads, while removing the Near-RT executor degrades QoS despite comparable power.The results attribute gains to coordinating context-aware loss construction, feedback-driven refinement, and state-dependent radio control.Source: E. Ablation Study · pp. 10–11, Figure 9 · p. 11
  • Across the constraint sweep, MLLM-coRIC operates near the requested QoS boundary with the lowest network power, outperforming learning-based and non-learning alternatives.The hybrid design learns ramp-constrained resource allocation and uses model-based optimization for transmit power based on current channel and interference state.Source: E. Ablation Study · pp. 11–12, Figure 10 · p. 12

F. Computational Complexity

  • A single MLLM refinement brings QoS violations below the stated constraint while lowering power, and execution time remains below 5 s after seven calls.Additional calls improve the operating point with diminishing marginal gains; cumulative execution time increases approximately linearly.Source: Figure 11 · p. 12, F. Computational Complexity · p. 12

V. CONCLUSION

  • The framework treats operator-defined optimization as a runtime input and separates Non-RT loss construction from Near-RT hybrid radio execution.Future work targets larger-scale and real O-RAN deployments, where imperfect measurements, forecasting errors, and interface delays may affect control.Source: §V · p. 12
Loading 2609.06122v1…