FedRIC: Heterogeneity-Aware Federated Learning with Hard Deadline Constraints in Open RAN

Adhwaa Al-Chaab; Ayman Younis; Dario Pompili

Source-linked AI summary

FedRIC: Heterogeneity-Aware Federated Learning with Hard Deadline Constraints in Open RAN

Adhwaa Al-Chaab, Ayman Younis, Dario Pompili

arXiv:2610.20530v1Oct 14, 2026eess.SYcs.SY

TL;DR

RIC control-loop deadlines and heterogeneous UE resources make real-time O-RAN learning challenging, while centralized learning has communication, latency, and data-locality drawbacks. The paper formulates and decomposes a resource-optimization problem and proposes FedRIC and FedRICE. Across two datasets, the frameworks improve accuracy and loss-reduction rates over the baselines, with FedRIC reporting 96.76% accuracy and minimal training latency.

Problem
Centralized ML transfers locally collected data, strains latency and scalability during aggregation, and exposes raw data, while deadline-constrained FL must handle heterogeneous UE compute, channels, and traffic.
Source: §I · p. 1
Method
The paper formulates O-RAN FL as a deadline-constrained MINLP, decomposes it into UE Assignment and Computation and Bandwidth Allocation, and proposes FedRIC and FedRICE with compressed gradients, latency-driven allocation, and hierarchical aggregation.
Source: §I · p. 2, §III · p. 3, B. Local Training with Compressed AGD · p. 4
Results
Across ColO-RAN and Telecom Italia, FedRIC and FedRICE achieve higher accuracy and faster loss reduction than baselines; FedRIC reports 96.76% accuracy with minimal training latency.
Source: B. Experiment Results · pp. 5–6
Takeaways & Limitations
The reported experiments indicate that the proposed frameworks satisfy hard RIC timing constraints while achieving faster convergence and reduced overhead.
Source: §I · p. 2

Abstract

from arXiv · show

Open Radio Access Networks (O-RANs) bring intelligence to disaggregated RANs through virtualized platforms and open interfaces, but the stringent latency budgets of the RAN Intelligent Controller (RIC) and heterogeneous traffic conditions make real-time deployment costly. Centralized learning is ill-suited to the RIC due to its computational burden, timing constraints, and communication overhead, motivating Federated Learning (FL) as a distributed alternative. We formulate FL in O-RAN as a Mixed-Integer Nonlinear Program (MINLP) that minimizes communication cost and learning time under heterogeneity in User Equipment (UE) compute, traffic, and channel conditions, and decompose it into a UE Assignment (UA) subproblem and a Computation and Bandwidth Allocation (CBA) subproblem. We then propose FedRIC, a heterogeneity-aware FL framework combining compressed gradient descent, slice-aware resource allocation, and multilevel hierarchical aggregation. Experiments show that FedRIC outperforms FedCOM, FedAvg, and QFedAvg in convergence rate, communication cost, and latency.

I. INTRODUCTION

O-RAN’s real-time control loops make centralized learning and deadline-unaware FL poorly suited to RIC deployment. The paper motivates deadline-aware FL, formulates its system and learning model, and proposes an optimization framework for heterogeneous conditions.

I. INTRODUCTION

Centralized ML transfers locally collected data, strains latency and scalability during aggregation, and exposes raw data, motivating distributed learning that preserves data locality.O-RAN’s open interfaces and RIC support use cases including QoE optimization, SLA assurance, uRLLC, eMBB, and real-time analytics.Source: §I · p. 1
Above roughly 40 KiB, E2AP processing exceeds the 1 s Near-RT RIC deadline, causing the xApp to miss its deadline and rendering the control loop ineffective.The passage identifies UE compute, channel conditions, and traffic loads as sources of heterogeneity for deadline-constrained FL.Source: §I · p. 1
Prior FL approaches address data heterogeneity, communication efficiency, or wireless settings, but the cited work does not account for hard RIC control-loop timing budgets.The paper frames deadline-constrained learning amid heterogeneous UE compute, channels, and traffic as an open gap.Source: §I · pp. 1–2
FL in O-RAN is formulated as a MINLP that jointly minimizes communication cost and training time under RIC deadlines, then decomposed into UE Assignment and Computation and Bandwidth Allocation subproblems.The proposed FedRIC and FedRICE frameworks combine compressed gradient descent, slice-aware resource allocation, and hierarchical aggregation.Source: §I · p. 2

II. SYSTEM MODEL

The system model considers one logical E2 node serving stationary UEs over orthogonal OFDM channels, with channel effects including fading, path loss, and shadowing.The E2 node comprises the CU, DU, and RU; the RIC node combines Non-RT and Near-RT RIC functions.Source: §II · p. 2

A. Learning Model

Each UE holds a disjoint, non-i.i.d. local dataset, and the learning model distinguishes global parameters from each UE’s local parameters and loss.The stated loss examples include MSE and cross-entropy.Source: A. Learning Model · p. 2

B. FL Model Accuracy

The model defines local and global accuracy targets and relates them to the number of global training rounds under stated regularity assumptions.

B. FL Model Accuracy

The local target ξ_l requires the gradient norm to shrink by at least a factor of ξ_l between successive global iterations, while the global target ξ_g is defined through global loss change.The targets are specified at the UE and RIC node, respectively.Source: B. FL Model Accuracy · p. 2
Under convexity, ρ-Lipschitz, and β-smoothness assumptions on φ(v), the model relates the global target ξ_g to the number of global iterations needed.The excerpt identifies I as a sufficient number of global iterations.Source: B. FL Model Accuracy · p. 2
The bound Υ(ξ_g, ξ_l) links the accuracy targets to training rounds, which feed into the cost and latency models.The displayed bound’s form is not fully given in the supplied passage.Source: B. FL Model Accuracy · p. 2

C. Resource Model

The resource model represents UE participation with binary variables and accounts for communication and computation costs.

D. Resource Model

A binary variable µ_u[i] indicates whether UE u participates in global iteration i.The participation vector collects these UE-level indicators.Source: C. Resource Model · p. 2
Active UEs receive at least w_min of total bandwidth W, while inactive UEs receive none.The passage states w_min ≤ 1/U.Source: C. Resource Model · p. 2
The total communication and computation costs are modeled using per-unit bandwidth cost ε and per-unit processing-time cost τ, respectively.The computation model also names UE CPU rate κ_u and cycles per bit f_u.Source: C. Resource Model · p. 2

D. Latency Model

With synchronous rounds, completion time is governed by the slowest selected UE; total training time and the combined cost therefore depend on per-round duration.

D. Latency Model

Each synchronous global round waits for every selected UE to finish local training and upload, so the slowest UE paces the round.The per-UE round duration includes computation and update-upload time.Source: D. Latency Model · p. 3
Downlink delay is omitted because its rate is much higher than the uplink and the global model is much smaller than aggregate local data.This omission follows the cited modeling assumption [18].Source: D. Latency Model · p. 3
Total training time is the number of global rounds multiplied by the maximum UE round duration, and the per-iteration objective combines resource cost with latency using δ.The maximum round duration is taken across UEs.Source: D. Latency Model · p. 3

E. Problem Formulation

The formulation casts O-RAN federated learning as a deadline-constrained MINLP, then motivates practical solution variants because exact optimization is computationally intractable. FedRIC uses two aggregation layers, while FedRICE adds an O-Cloud layer.

E. Problem Formulation

The objective is to minimize Ω[i] under system constraints, including normalized bandwidth allocation, minimum bandwidth for active UEs, zero allocation for inactive UEs, and total bandwidth and compute limits.The formulation also imposes local and global accuracy targets.Source: E. Problem Formulation · p. 3
The formulation caps aggregate compute demand at Fmax and requires local and global accuracy targets.Bandwidth constraints also limit total usage to W.Source: E. Problem Formulation · p. 3

III. PROPOSED FEDRIC SOLUTION

The MINLP is NP-hard because UE-selection combinations grow exponentially, selection and resource-allocation variables are non-convexly coupled, and exact solvers can take minutes to hours against millisecond deadlines.The framework therefore considers two deployment variants: FedRIC's two-layer hierarchy and FedRICE's additional O-Cloud aggregation layer.Source: §III · p. 3
The MINLP's binary UE-selection variables create 2^|U| possible UE-selection combinations per round and exponential complexity across global iterations.Its objective also couples discrete selection variables with continuous allocation variables.Source: §III · p. 3
FedRIC aggregates selected UEs' compressed gradients at their serving DU with momentum smoothing, then forwards the result to the CU for global aggregation every ΔCU rounds.FedRICE adds a third aggregation layer at the O-Cloud level.Source: §III · p. 3
The UE-assignment algorithm requires a UE set and deadline and returns the selected UE set.Its procedure stops when adding another UE would exceed the deadline and enforces minimum slice representation when feasible.Source: §III · p. 3

A. Problem Decomposition

The UE Assignment subproblem selects participating UEs to maximize participation while keeping each round within its deadline. A slice-aware heuristic approximates the NP-hard time-budgeted selection problem.

A. Problem Decomposition

The UE Assignment stage determines binary participation variables indicating which UEs take part in each round.Its resulting subproblem is formulated as a 0–1 integer program.Source: A. Problem Decomposition · p. 3
UE Assignment aims to maximize the number of participating UEs in each global iteration while keeping round duration within the deadline.Including more UEs is intended to incorporate more data and accelerate convergence.Source: A. Problem Decomposition · p. 3
The per-round objective weights incoming-slice time contribution by ζ, while the incremental time constraint must fit within the remaining deadline.The time cost combines transmission and computation terms for selected UEs.Source: A. Problem Decomposition · p. 3
The slice-aware heuristic admits UEs with the smallest incremental time cost, boosts underrepresented eMBB, uRLLC, and mMTC slices, and stops before a deadline violation.The underlying time-budgeted selection problem is NP-hard.Source: A. Problem Decomposition · p. 3

2) Computation and Bandwidth Allocation (CBA, P2 subproblem):

After UE selection, Computation and Bandwidth Allocation assigns resources to minimize the round's duration. Because its coupled min-max optimization is NP-hard, the method uses a fast proportional allocation heuristic.

2) Computation and Bandwidth Allocation (CBA, P2 subproblem):

For the selected UEs, CBA allocates bandwidth and computational resources for the current round, choosing computation allocations to minimize round duration.The selected UE variables are fixed from the UA stage before the reduced optimization is formed.Source: 2) Computation and Bandwidth… · pp. 3–4
The reduced CBA optimization follows substitution of fixed UA selection variables into the original problem.Its resource allocations are sent to each selected UE before local training and model upload.Source: 2) Computation and Bandwidth… · p. 4
The CBA subproblem is NP-hard because its min-max objective couples bandwidth and computation time, so a fast proportional heuristic balances UE completion times.The heuristic operates in O(|Mu|) time and avoids exact-optimization overhead.Source: 2) Computation and Bandwidth… · p. 4
Computational resources are allocated proportionally to processing requirements, while bandwidth is distributed proportionally to upload sizes with a minimum guarantee.Bandwidth shares are normalized to sum to one.Source: 2) Computation and Bandwidth… · p. 4

B. Local Training with Compressed AGD

FedRIC combines randomized gradient compression and momentum-based accelerated local training within allocated UE resources, with hierarchical aggregation across the RAN. Its convergence guarantee is stated under smoothness, strong-convexity, and compression assumptions.

B. Local Training with Compressed AGD

Selected UEs apply randomized compression to local gradients before transmission, with the compressor required to be unbiased and to have bounded relative variance.The stochastic quantizer's per-round bit budget decreases from 16 to 2 bits over successive global rounds.Source: B. Local Training with Compressed AGD · p. 4
Momentum incorporates compressed gradients and accumulated historical gradient information to accelerate convergence under communication constraints.The paper states that combining compression and momentum maintains convergence guarantees under deadline constraints.Source: B. Local Training with Compressed AGD · p. 4
Each UE performs its local iterations within its assigned computation budget, with the number of iterations constrained by dataset size, per-sample complexity, processing power, and allocated time.The training process respects CBA resource allocations.Source: B. Local Training with Compressed AGD · p. 4
FedRIC's compressed AGD has an expected optimization-error bound under an unbiased compressor with bounded variance and standard smoothness and strong-convexity assumptions.The proof adapts AGD-with-compression analysis to the momentum update; compression variance and momentum appear in the bound.Source: B. Local Training with Compressed AGD · p. 4
FedRIC aggregates compressed updates by data-size-weighted averaging at each DU with DU-level momentum, then forwards them to the CU every ΔCU rounds.FedRICE inserts an O-Cloud edge-cluster aggregation step before the DU stage, adding one hop per round.Source: B. Local Training with Compressed AGD · p. 4

IV. PERFORMANCE EVALUATION

The evaluation setup varies uplink bandwidth and compresses updates over successive rounds; the supplied visuals identify model and dataset comparison dimensions.

Updates are quantized from 16 bits down to 2 bits over successive rounds, with uplink bandwidth constrained between 10 and 1000 units per round.Source: §IV · p. 4
Table I is identified as a comparison of models on key metrics.Source: Table I · p. 4
Figure 3 compares four sequence models on ColO-RAN using ROC curves, accuracy versus training time and rounds, and loss versus communicated bits.Source: Figure 3 · p. 5
Figure 4 compares Telecom Italia results using accuracy and loss versus rounds, transmission volume versus UE count, and the compression operator's effect.Source: Figure 4 · p. 5

A. Experimental Setting

The evaluation uses a KPI-violation detection task on ColO-RAN and compares two hierarchical deployment schemes against three established FL baselines.

The ColO-RAN task uses 18,329 one-second snapshots from seven gNBs, labeled for missed 1 s Near-RT deadlines and uniformly partitioned across 50 simulated RIC nodes.Source: A. Experimental Setting · p. 5
FedRIC uses two-layer hierarchical aggregation, while FedRICE adds an O-Cloud aggregation layer before final aggregation at the E2 node.Source: A. Experimental Setting · p. 5
The benchmarks are FedAvg, QFedAvg, and FedCom, representing fixed-subset uncompressed learning, fairness-weighted UE contributions, and compressed updates, respectively.Source: A. Experimental Setting · p. 5

B. Experiment Results

FedRIC and FedRICE outperform the compared FL methods in convergence, communication cost, and latency. Results span ColO-RAN and Telecom Italia, with transmission volume also measured as UE counts increase.

B. Experiment Results

The simulated O-RAN cell uses distance-dependent path loss and models small-scale Rayleigh fading with Jake’s first-order complex Gauss–Markov model.Source: B. Experiment Results · p. 5
FedRIC reaches 96.76% accuracy with minimal training latency, while FedRICE attains 94.82% accuracy with latency consistently below 0.9s.Source: B. Experiment Results · p. 5
On both datasets, FedRIC and FedRICE achieve higher accuracy and faster loss reduction than the baselines, with high accuracy reached in fewer ColO-RAN rounds.Source: B. Experiment Results · p. 6
At 100 UEs, transmission volume is 5,200 Mbits for FedAvg, 1,000 Mbits for FedRIC, 500 Mbits for FedRICE, and 1,800 Mbits for FedCOM.Source: B. Experiment Results · p. 6
FedRIC and FedRICE reach lower loss in fewer rounds while transmitting fewer bits than FedCOM, QFedAvg, and FedAvg.Source: B. Experiment Results · p. 6

V. CONCLUSION

FedRIC and FedRICE outperform FedCOM, FedAvg, and QFedAvg in convergence rate, communication cost, and latency.Source: §V · p. 6