PaperlayerView exact revision on arXiv

Source-linked AI summary

RRM-GPT: A Framework and Vision for Radio Resource Management Foundation Models

Ahmed Aboulfotouh, Akram Bin Sediq, Koosha Pourtahmasi Roshandeh, Omar Mashaal, Ahmad M. Nagib, Jale Sadreddini, Hatem Abou-Zeid

arXiv:2610.04296v1eess.SPcs.AI

TL;DR

Learning-based RRM models are commonly specialized to a function and deployment, so new settings require repeating development. RRM-GPT generates structured decisions autoregressively, with network-log pretraining and deployment-specific post-training. A 5G NR case study demonstrates complete scheduling-grant generation and transfer to an unseen scenario.

  • Problem

    Learning-based RRM models are typically specialized to one function and deployment, requiring development to be repeated for new settings.

    Source: INTRODUCTION · p. 1

  • Method

    RRM-GPT autoregressively generates RRM decisions and uses network-log pretraining followed by post-training for deployment-specific objectives.

    Source: INTRODUCTION · pp. 1–2

  • Results

    The 5G NR case study demonstrates complete scheduling-grant generation; initial-transmission MCS accuracy rises from 40% to 68% for downlink and from 47% to 91% for uplink.

    Source: INTRODUCTION · p. 2, NUMERICAL RESULTS · p. 8

  • Takeaways & Limitations

    The reported transfer to an unseen scenario supports the framework’s intended reuse across deployments, though transfer is not uniform across fields.

    Source: INTRODUCTION · p. 1, NUMERICAL RESULTS · p. 8

    Imitation-based policy learning is limited by the behavior represented in the logs; objective-driven post-training may refine that policy.

    Source: LEARNING FROM NETWORK LOGS · p. 4

Abstract

from arXiv · show

Learning-based models for radio resource management (RRM) are typically built for a single function and deployment, so each new setting repeats the development pipeline. RRM decisions, however, share a common structure: each is assembled from interdependent fields, defined by the standard, whose values are selected in view of the network state. We propose RRM-GPT, an autoregressive framework for RRM foundation models that generate these decisions as a language model generates text. An encoder maps heterogeneous network observations into a common token representation, and a decoder emits the decision one field at a time, each conditioned on the network state and the fields already committed. Pretraining on unannotated network logs teaches the model what makes a decision valid and how controllers choose among valid decisions; post-training then adapts it to deployment-specific operator objectives through imitation or reinforcement learning. The framework targets two forms of reuse: a function-specific model reused across deployments, and a model shared across RRM functions that generates their interdependent decisions as one sequence. In a 5G New Radio (NR) case study, we demonstrate that a single model generates complete scheduling grants spanning user selection, timing, link adaptation, resource allocation, and control signaling. The model captures dependencies among grant fields and transfers learned behavior to an unseen scenario without adaptation.

INTRODUCTION

RRM-GPT frames radio resource management decisions as interdependent fields generated autoregressively from network observations, with pretraining on network logs and post-training for deployment objectives. It targets reuse across deployments and RRM functions, and demonstrates complete scheduling-grant generation and evaluation on an unseen scenario.

INTRODUCTION

  • Learning-based RRM models are typically specialized to one function and deployment, requiring the development pipeline to be repeated for new settings.Independent models also make it harder to capture dependencies among RRM decisions in a shared context.Source: INTRODUCTION · p. 1
  • The framework targets function-specific model reuse across deployments and, more broadly, a shared model across RRM functions.A scheduling model could adapt to different cell configurations, environments, and traffic conditions.Source: INTRODUCTION · p. 1, Figure 1 · p. 2
  • RRM decisions share a field-based structure, but their fields are interdependent within decisions and across functions.For example, channel conditions constrain rate, while handover and interference coordination affect later scheduling choices.Source: INTRODUCTION · p. 1
  • RRM-GPT generates feasible RRM decisions field by field, using network-log pretraining and post-training for deployment-specific objectives.Post-training can imitate a reference controller or use reinforcement learning to optimize an operator-defined objective.Source: INTRODUCTION · pp. 1–2
  • The representation accommodates heterogeneous network observations and function-specific decision fields and constraints.The framework is also outlined for mobility, interference coordination, and load management.Source: INTRODUCTION · p. 2
  • The 5G NR case study demonstrates complete scheduling-grant generation and evaluates generalization to an unseen scenario.The grant spans user selection, timing, link adaptation, resource allocation, and control signaling.Source: INTRODUCTION · p. 2

SCHEDULING AS SEQUENCE GENERATION

A 5G NR grant combines interdependent data and control-resource choices, so RRM-GPT generates fields conditionally rather than choosing from the enormous set of complete grants.

SCHEDULING AS SEQUENCE GENERATION

  • A DL grant jointly assigns data and control resources, with queue demand, CQI, and recent transmission outcomes shaping user, MCS, PRB, timing, and CCE choices.The example allocates twelve PRBs starting at PRB 24 and uses four CCEs starting at CCE 8.Source: SCHEDULING AS SEQUENCE GENERATION · p. 2

FROM SENTENCES TO GRANTS

  • Fig. 2 contrasts word-by-word text with grant generation: RRM emits fields sequentially, conditioned on network state and prior choices, rather than selecting among millions of complete grants.The slot may contain several grants, and each grant changes which choices remain available.Source: FROM SENTENCES TO GRANTS · pp. 2–3, Figure 2 · p. 3

TOKENS AND VOCABULARIES

  • Output tokens choose field values from standards- and configuration-defined vocabularies, while heterogeneous observations are structured and projected into common input tokens.The same generation principle covers scheduling grants, muting patterns, and handovers, with zero or more actions in a sequence.Source: TOKENS AND VOCABULARIES · p. 3, Table I · p. 4

ENCODER–DECODER ARCHITECTURE

  • The encoder jointly contextualizes state tokens, then the decoder reuses that encoded state while conditioning each next field on choices already generated.The state is encoded once per decision step.Source: ENCODER–DECODER ARCHITECTURE · p. 4

LEARNING FROM NETWORK LOGS

  • Next-field prediction on logged state-action records teaches both validity rules and controller behavior without explicitly implementing controller algorithms.Operational-network logs and simulator records provide state and decision pairs for this training objective.Source: LEARNING FROM NETWORK LOGS · p. 4, Figure 3 · p. 5

GUARANTEEING FEASIBILITY

  • Because rules learned from logs provide no hard guarantees, inference masks values conflicting with committed slot resources and updates a tracker after each generated field.A model trained on sufficient logs largely respects these rules, so masking rarely changes the outcome.Source: GUARANTEEING FEASIBILITY · p. 5

BEYOND SCHEDULING: AUTOREGRESSIVE MODELING OF JOINT RRM

RRM-GPT extends autoregressive generation to interdependent RRM functions while preserving their distinct decision structures and timescales. Post-training specializes the model to deployment objectives, and a 5G NR scheduling case study instantiates the framework.

BEYOND SCHEDULING: AUTOREGRESSIVE MODELING OF JOINT RRM

  • A shared model can generate load-management, handover, interference-coordination, and scheduling decisions as an interdependent sequence, allowing decisions across functions to condition on one another.The scheduling formulation takes serving cell and coordination restrictions as given, although other RRM functions shape those conditions.Source: BEYOND SCHEDULING: AUTOREGRESSIVE… · p. 5

A UNIFIED GENERATIVE ARCHITECTURE

  • Functions retain distinct decision structures and control timescales, with generation ordered so resource constraints precede affected grants and slower decisions persist in the state.Handover, load management, and interference coordination can complement or substitute for one another when addressing an overloaded cell.Source: A UNIFIED GENERATIVE ARCHITECTURE · pp. 5–6

POST-TRAINING AND SPECIALIZATION

  • Post-training preserves the shared state representation and output structure while adapting model weights through controller imitation or reward feedback for deployment objectives.Behavioral cloning uses logged controller decisions without an explicit reward; reinforcement learning adjusts decisions against long-term return, and pretraining may reduce deployment-specific data or interaction needed for adaptation.Source: POST-TRAINING AND SPECIALIZATION · p. 6

CASE STUDY: 5G NR SCHEDULING

A simulator-trained 5G NR model generates scheduling grants field by field and is evaluated on held-out mobile data and a static scenario excluded from training. It reproduces dependencies among grant fields and transfers some scheduler behavior without adaptation, though transfer is not uniform across fields.

EXPERIMENTAL SETUP

  • The simulator study trains on approximately 6.5 million grants from mobile users, tests held-out runs and a wholly excluded static scenario, and emits seven ordered grant fields.The encoder and decoder each use two Transformer layers, 256-dimensional embeddings, and four attention heads; the model has approximately 4.3 million parameters and trains for 50 epochs.Source: EXPERIMENTAL SETUP · p. 6

NUMERICAL RESULTS

  • Uplink starting-PRB accuracy rises from 48% to 99% under teacher forcing; MAE falls from 2.8 to 0.1 PRB indices.This suggests error propagation accounts for much of the rollout mismatch in resource placement.Source: NUMERICAL RESULTS · p. 6, Figure 4 · p. 7
  • 40%→68% (downlink) and 47%→91% (uplink) initial-transmission MCS accuracy, but 48%→37% uplink starting-PRB accuracy, marked transfer to static testing.The static scenario uses static users in a smaller cell and is tested without adaptation.Source: Figure 5 · p. 7, NUMERICAL RESULTS · p. 8
  • Rollout MAE is 1.0 MCS indices for downlink and 1.2 for uplink initial transmissions, with low MAEs in both modes indicating scheduler choices beyond exact-match accuracy alone.Source: NUMERICAL RESULTS · p. 7, Figure 4 · p. 7
  • The results support function-specific reuse, while broader generalization, post-training benefits, and network-performance impact remain open for evaluation.Source: NUMERICAL RESULTS · p. 8

CONCLUSION AND OPEN PROBLEMS

The 5G NR case study demonstrates complete scheduling-grant generation and partial generalization to an unseen scenario; broader adaptation and cross-function reuse remain future directions with open representation, data, deployment, and operator-intent challenges.

CONCLUSION AND OPEN PROBLEMS

  • The 5G NR case study demonstrates complete scheduling-grant generation and partial generalization to an unseen scenario.Reliable adaptation across a wider range of deployments and objectives remains a next step.Source: CONCLUSION AND OPEN PROBLEMS · p. 8
  • A shared decision representation must balance transfer across standards and systems with the specificity needed to generate valid decisions.The appropriate fields, granularity, and order remain open design questions.Source: CONCLUSION AND OPEN PROBLEMS · p. 8
  • Pretraining for reuse depends on logs covering diverse configurations, channel conditions, traffic demands, and controller policies.Cross-function learning also requires traces preserving decision timing and dependencies alongside relevant network observations.Source: CONCLUSION AND OPEN PROBLEMS · p. 8
  • Architectures must account for tight inference budgets on slot-timescale decisions and avoid repeating expensive computation at fast timescales.Decisions and inputs processed every few minutes allow greater computational complexity.Source: CONCLUSION AND OPEN PROBLEMS · p. 8
  • Deploying model instances at multiple network points leaves the challenge of keeping decisions consistent across their different local observations.Each instance may operate on information available to its local control loop.Source: CONCLUSION AND OPEN PROBLEMS · p. 8
  • Post-training must address operator goals that may be implicit, preference-based, or variable with time, deployment conditions, and commercial priorities.The paper also identifies safe adaptation without live-network exploration and lightweight specialization for diverse objectives as open challenges.Source: CONCLUSION AND OPEN PROBLEMS · p. 8
Loading 2610.04296v1…