PaperlayerView exact revision on arXiv

Source-linked AI summary

A DRL-Driven Optimization of RAN Slice Resource Partitioning for V2X SLA Compliance in 5G Networks

M. Martínez, I. de-la-Bandera, D. E. García, P. Vera, S. Fortes, M. L. Luque, A. Mendo, J. Ramiro, R. Barco

arXiv:2609.27659v1cs.NI

TL;DR

V2X services must meet strict latency and reliability requirements while sharing scarce radio resources with heterogeneous services. The paper trains a PPO-based agent to partition downlink PRBs among slices, and simulations report SLA compliance across considered loads alongside improved resource efficiency.

  • Problem

    V2X services must meet stringent latency and reliability targets while competing with heterogeneous services for limited radio resources.

    Source: §I · p. 1, §VI · p. 12

  • Method

    A PPO-based DRL strategy trains on precomputed simulations to allocate PRBs among RAN slices while targeting V2X SLA compliance, resource efficiency, and reduced eMBB impact.

    Source: §I · p. 2, §VI · p. 12

  • Results

    For every considered load level, the learned policy adapts PRB allocation to network demand, ensures V2X SLA compliance, and improves resource efficiency in each partition.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The methodology is presented as a flexible dynamic PRB-allocation strategy that can extend to other SLA, throughput, and load conditions through appropriate training.

    Source: §VI · p. 13

Abstract

from arXiv · show

Vehicle-to-Everything (V2X) communications impose very demanding requirements in terms of latency and reliability, which must be met in scenarios where multiple services with diverse performance targets coexist. In such scenarios, traffic-intensive services compete for limited radio resources, complicating the fulfillment of V2X service demands. Within this context, Network Slicing (NS) emerges as a key factor that enables the creation of multiple slices and the allocation of resources among them to satisfy heterogeneous service requirements. In particular, this work addresses the Radio Access Network (RAN) slicing problem from the perspective of Physical Resource Block (PRB) partitioning under high traffic demand conditions. To this end, a reinforcement learning approach based on Proximal Policy Optimization (PPO) is proposed to determine PRB allocations that satisfy the strict latency and reliability requirements of V2X services, while improving resource utilization efficiency and minimizing performance degradation of enhanced Mobile BroadBand (eMBB) services. The proposed solution is evaluated through simulation-based experiments under various traffic loads and different V2X service requirements, demonstrating its ability to adapt resource partitioning to network conditions and service demands.

I. INTRODUCTION

V2X services must meet stringent latency and reliability targets while sharing limited radio resources with services such as eMBB. The paper motivates adaptive RAN slicing and a PPO-based approach, noting that prior work often does not jointly address V2X delay and reliability in heterogeneous multi-service settings.

I. INTRODUCTION

  • V2X SLA compliance becomes harder under high traffic and bandwidth-intensive competition, while vehicle mobility and changing demand make radio allocation time-varying.Inadequate allocation risks over- or under-provisioning, SLA noncompliance, and degraded network performance.Source: §I · p. 1
  • The work uses PPO-based reinforcement learning to partition downlink PRBs among slices, targeting V2X latency and reliability while limiting eMBB impact.Its stated contributions include jointly optimizing V2X SLA compliance, resource utilization, and eMBB throughput, with V2X compliance defined using both delay and reliability.Source: §I · p. 2, §II · p. 3

II. RELATED WORK

  • Prior approaches often use generic slice profiles or omit explicit V2X latency and reliability requirements; other studies do not capture resource competition across diverse V2X services.The cited work includes studies limited to two V2X slices or using other representations of reliability and latency.Source: §II · p. 2
  • The paper evaluates its PPO-based RAN-slicing approach across different loads, performance targets, and 3GPP-adapted V2X traffic patterns.The stated setting includes multiple V2X and eMBB slices with different traffic characteristics and performance targets.Source: §II · p. 3

III. SYSTEM MODEL AND PROBLEM FORMULATION

The model considers downlink PRB allocation among service-specific slices and characterizes V2X traffic and SLA compliance using packet-delay statistics. V2X packets exceeding their delay budget are discarded and counted in the packet error rate.

A. System Model

  • The downlink cell has a finite PRB capacity, and the scheduler allocates resources per slot with a 1 ms TTI.The network deploys service-associated slices, with the model distinguishing services that have different traffic characteristics and performance requirements.Source: A. System Model · p. 3

1) V2X Service Model:

  • V2X services generate periodic packets, with packet size and generation period allowing the model to represent heterogeneous traffic patterns.The service-level arrival rate is the average number of packets arriving at the downlink buffer per millisecond over the observation interval.Source: 1) V2X Service Model · p. 3
  • The V2X traffic model measures service-level packet arrivals over the observation interval, whose TTIs have a duration of 1 ms.The passage defines the rate from arrivals across the set of TTIs in that interval.Source: 1) V2X Service Model · p. 3, A. System Model · p. 3
  • V2X SLA compliance is assessed by comparing a service's measured packet-delay percentile with its maximum tolerable delay.The SLA also specifies the required reliability percentage, and delay percentiles jointly assess how many packets satisfy the latency constraint.Source: 1) V2X Service Model · p. 3
  • Packets that exceed the Packet Delay Budget are treated as lost and discarded because late V2X information may be outdated.The stated rationale is that discarding stale packets can help reduce network congestion.Source: 1) V2X Service Model · p. 3
  • The model defines packet error rate using the number of packets discarded for exceeding the delay budget and the total number generated during the observation period.The two counts are specified for each slice.Source: 1) V2X Service Model · p. 3

2) eMBB Service Model:

The eMBB model treats traffic as a continuous data flow whose users remain active and request resources at each TTI. It evaluates eMBB service performance through user throughput.

2) eMBB Service Model:

  • eMBB traffic is modeled as a continuous data flow, with active users persistently requesting radio resources at each TTI.The service packet arrival rate is defined with reference to the packet-arrival model introduced earlier.Source: 2) eMBB Service Model · pp. 3–4
  • The model adopts throughput as the main eMBB performance metric because it depends on allocated radio resources and channel conditions.Performance is characterized using average throughput experienced by users over the observation interval.Source: 2) eMBB Service Model · p. 4
  • The eMBB service-level throughput measure is based on the users associated with that service and their average throughput over the observation set.The passage identifies the associated-user set and its cardinality in the metric definition.Source: 2) eMBB Service Model · p. 4

B. Scheduler Description

The scheduler first partitions PRBs among slices, then schedules users within each slice and can share unused resources. The formulation seeks a PRB distribution that satisfies V2X SLAs while improving slice resource efficiency and eMBB throughput under high load.

B. Scheduler Description

  • Inter-slice allocation assigns each slice a subset of PRBs subject to the cell's total-capacity constraint, before intra-slice scheduling begins.The stated scheduler then uses Round Robin to allocate each slice's assigned PRBs among its active users.Source: B. Scheduler Description · p. 4
  • PRBs left unassigned or unused after intra-slice scheduling may be shared among users still requiring resources, using Round Robin.This sharing is performed independently of the user's associated service or slice.Source: B. Scheduler Description · p. 4

C. Problem Formulation For RAN Slicing

  • The problem allocates PRBs across four dedicated slices: two V2X services and two eMBB services.The allocation vector contains one PRB count for each slice; the service models permit heterogeneous V2X traffic and eMBB throughput targets.Source: C. Problem Formulation For RAN Slicing · p. 4
  • Under high-load conditions, every slice must receive at least one PRB and all available PRBs are allocated while pursuing SLA compliance, resource utilization, and eMBB throughput needs.Remaining PRBs are assigned to eMBB slices according to their throughput needs.Source: C. Problem Formulation For RAN Slicing · p. 4
  • The formulation treats V2X SLA compliance as a hard constraint and optimizes slice resource efficiency and eMBB throughput within the feasible set.V2X SLA assessment uses the relevant delay percentile.Source: C. Problem Formulation For RAN Slicing · p. 4

IV. LATENCY AWARE RAN SLICING SOLUTION

The section frames RAN slicing as an adaptive cell-level PRB allocation problem: V2X requirements must be met while limiting effects on overall performance. It uses an RL agent to choose allocations from observed slice KPIs and learn from their outcomes.

IV. LATENCY AWARE RAN SLICING SOLUTION

  • RAN slicing seeks PRB allocations that satisfy V2X requirements while minimizing impacts on overall performance as traffic demand and channel quality change.The section presents this as an adaptive decision-making problem suited to reinforcement learning.Source: §IV · p. 4

A. DRL-Based Slicing Strategy

  • During training, the agent selects actions, observes network behavior, and uses the resulting reward to learn which actions yield higher rewards.The strategy periodically allocates PRBs based on service performance in the cell.Source: A. DRL-Based Slicing Strategy · pp. 4–5

1) State space:

  • The agent observes a slice-level KPI state vector, including allocated and used PRBs, SINR, packet arrival rate, and active connections.The state is computed over an observation interval.Source: 1) State space · p. 5

2) Action space:

  • With four slices, each action specifies a PRB allocation vector; the implemented discrete actions index vectors that satisfy the PRB budget constraint.The action components specify the number of PRBs allocated to each slice.Source: 2) Action space · p. 5

3) Reward Computation:

The reward function uses separate branches for V2X SLA-compliant and noncompliant allocations. It penalizes delay and packet loss when requirements are missed, and otherwise favors resource efficiency and eMBB throughput.

3) Reward Computation:

  • When both V2X slices meet their SLAs, the reward pursues resource efficiency and eMBB user throughput.The compliant-allocation branch applies after the percentile constraints are guaranteed.Source: 3) Reward Computation · p. 5
  • The resource-efficiency term penalizes over-provisioning by comparing allocated PRBs with average PRB use, with arrival-rate weighting increasing penalties for inefficient allocations at lower arrival rates.The allocation-use difference is normalized by the allocated PRBs.Source: 3) Reward Computation · p. 5
  • The eMBB reward uses throughput relative to expected throughput, with ρ_i weighting intended to avoid systematically favoring services that are easier to satisfy.The weighting promotes a more balanced optimization when services have heterogeneous performance expectations.Source: 3) Reward Computation · p. 5
  • When V2X SLA requirements are not met, the reward penalizes normalized delay degradation and packet loss, with stronger penalties as violations become more severe.The delay term is normalized by the maximum tolerable delay, while the packet-loss indicator is normalized to a 0–1 scale.Source: 3) Reward Computation · p. 5

4) PPO Agent:

The solution uses PPO with actor and critic networks to select allocations and estimate policy value. Training uses pre-collected observations, varying episode loads, and transitions retrieved from a dataset rather than direct interaction with a real deployment.

4) PPO Agent:

  • PPO uses an actor to produce action probabilities from the current state and a critic to estimate the policy value.Both networks are MLPs with two hidden layers and tanh activation.Source: 4) PPO Agent · p. 6
  • The PPO loss uses clipping to prevent abrupt policy updates, while Generalized Advantage Estimation estimates the advantage and supports convergence.The passage also states that clipping reduces policy-update variance.Source: 4) PPO Agent · p. 6

5) Environment and Training Methodology:

  • Training uses a dataset of pre-collected network observations to learn action consequences without interacting directly with a real deployment.The dataset covers different PRB allocations and load conditions.Source: 5) Environment and Training Methodology · p. 6
  • Each episode randomly selects a load level that remains fixed, and starts from a PRB allocation chosen for the highest aggregated V2X delay percentiles.The initial allocation is selected from the configurations associated with that load level.Source: 5) Environment and Training Methodology · p. 6
  • At each step, the agent samples a PRB allocation, retrieves a dataset state for the next state, and computes reward metrics while preserving the observed traffic conditions.The transition uses the selected allocation and the load level in the observed state.Source: 5) Environment and Training Methodology · p. 6

A. Simulation Setup

The simulation setup evaluates cell-level PRB partitioning using multiple traffic loads, service configurations, and allocation samples. It also specifies the observation window and how experienced packet delay treats discarded packets.

A. Simulation Setup

  • The setup uses a 5G system-level simulator with RAN slicing and a cell-level scenario with 50 available PRBs.The scenario is a cell from a trisectorized site.Source: A. Simulation Setup · p. 7
  • The setup defines V2X traffic patterns and two eMBB services with expected throughputs of 2.24Mbps and 11.2Mbps, corresponding to 1 and 5 PRBs under favorable channels.V2X-1 has a 10ms inter-packet arrival time and 6400-bit packets; V2X-2 has a 50ms inter-packet arrival time.Source: A. Simulation Setup · p. 7
  • Three operating scenarios represent low, medium, and high demand, characterized by packet-generation-rate statistics and service user counts.The packet-generation rates are expressed in packets per millisecond.Source: A. Simulation Setup · p. 7
  • The dataset uses 47 PRB configurations spanning SLA violations, compliant allocations, and V2X over-provisioning; 564 simulation samples comprise the training dataset.The configurations are simulated under three load levels.Source: A. Simulation Setup · pp. 7–8
  • Slice-level KPIs are computed over 30,000 TTIs, or 30 seconds, selected as a trade-off between simulation time and statistical representativeness.The passage says this interval provides enough packet transmissions for stable KPIs.Source: A. Simulation Setup · p. 8
  • Experienced V2X delay runs from scheduler arrival until transmission completion or discard; discarded packets are included in the delay distribution, and PDB overruns are recorded at PDB+1 ms.The simulator's 1ms resolution identifies an overrun at the following TTI boundary.Source: A. Simulation Setup · p. 8

B. Training performance Analysis

Across two independently trained SLA settings and varying load levels, the agent learns allocations that meet V2X requirements while improving resource efficiency. Training also reshapes eMBB allocations, including a pronounced throughput shift at Load level 2 under SLA 2.

B. Training performance Analysis

  • The 90th-percentile delay stabilizes below 15 ms and the 99.99th-percentile around or below 6 ms; V2X SLA violations become isolated across traffic loads.The analysis evaluates both V2X slices under each SLA scenario.Source: B. Training performance Analysis · p. 8, Figure 2 · p. 9
  • At Load level 2 under SLA 2, eMBB-1 throughput falls from approximately 2.65 Mbps to 1.6 Mbps, while eMBB-2 rises from 5.08 to 6.3 Mbps.The prose attributes the shift to a change in PRB allocation policy.Source: Figure 3 · p. 9, B. Training performance Analysis · p. 10
  • With a 99.99% SLA versus 90%, eMBB-1 achieves higher throughput and eMBB-2 lower throughput, as periodic V2X traffic leaves reusable PRBs that benefit eMBB-1.The text says eMBB-1 requires fewer resources to achieve relatively high performance.Source: B. Training performance Analysis · p. 9
  • As training progresses, allocated-to-utilized PRB gaps shrink and stabilize; around episode 7000 under Load level 2 and SLA 2, fewer PRBs are dedicated to V2X.The adjustment leaves fewer opportunities for eMBB-1 to reuse unused V2X PRBs and allows a larger share to be allocated explicitly to eMBB-2.Source: B. Training performance Analysis · p. 10
  • The eMBB-2 slice fully utilizes its allocated PRBs throughout training, while eMBB-1's allocation efficiency improves from negative values toward zero.The reported pattern is linked to eMBB-2's high throughput expectations and eMBB-1's exploration of allocations that sometimes over-provision resources.Source: B. Training performance Analysis · p. 10
  • A dominant selected action emerges for each load level, and under SLA 2 at Load level 2 the policy shifts around episode 7000 from 25-9-3-13 to 13-5-5-27.The new action replaces an allocation that had over-provisioned the V2X segments at Load level 2.Source: B. Training performance Analysis · p. 10

C. Model deployment analysis

The trained model is integrated into the simulation tool to periodically use network KPIs to recommend inter-slice PRB allocations. In two experiments under SLA 2 and the second load level, it responds to SLA violations and V2X over-provisioning while maintaining the V2X delay constraint.

C. Model deployment analysis

  • The trained model periodically receives network state and proposes an inter-slice allocation, which is applied when it differs from the configured allocation.The deployment evaluation uses SLA 2 at load level 2 and presents one experiment starting from a suboptimal distribution and another addressing V2X over-provisioning.Source: C. Model deployment analysis · p. 10, Figure 6 · p. 12
  • Initially, 99.84% of V2X-1 delays exceed 6 ms and its 99.99th-percentile delay is 31 ms; the model then recommends C = {13, 5, 5, 27}.In the same initial period, V2X-2 has 0.29% of delays above 6 ms and a 9 ms 99.99th-percentile delay.Source: C. Model deployment analysis · p. 11
  • In the first experiment’s third period, both V2X services reach a 99.99th-percentile delay of 5 ms; only 0.005% of V2X-1 delays exceed 6 ms, and all V2X-2 delays meet that constraint.The experiment begins with a 31 ms 99.99th-percentile delay for V2X-1, prompting a new allocation; the reported third-period results follow the transition.Source: C. Model deployment analysis · p. 11
  • In the second experiment’s first period, V2X-1 and V2X-2 use only 41% and 45.5% of allocated resources, respectively, while their 99.99th-percentile delay stays below 6 ms.The eMBB slices achieve average utilization above their assigned share by competing for PRBs unused by V2X slices.Source: C. Model deployment analysis · p. 11
  • 72.3% V2X-1 and 78.96% V2X-2 average resource utilization are achieved after reallocation, while the V2X delay constraint remains below 6 ms.The second experiment reduces V2X PRBs that were over-provisioned in its first period.Source: C. Model deployment analysis · pp. 11–12
  • After reallocation, eMBB-2 throughput rises while eMBB-1 throughput falls, producing more similar relative performance while V2X SLAs remain met.The eMBB-1 slice receives more allocated PRBs but has fewer unused V2X PRBs available to reuse.Source: C. Model deployment analysis · p. 12

VI. CONCLUSIONS AND FUTURE WORK

The paper addresses V2X SLA support amid heterogeneous services by using a DRL-based RAN slicing strategy to allocate PRBs among slices. In simulations, the learned policy adapts to network demand and meets V2X SLAs across considered loads; simulation deployment tests also cover SLA violations and over-provisioning.

VI. CONCLUSIONS AND FUTURE WORK

  • The proposed DRL-based RAN slicing strategy determines PRB distributions among slices while targeting SLA compliance, resource efficiency, and limited eMBB performance impact.Training uses precomputed simulation data rather than direct agent interaction with the network.Source: §VI · p. 12
  • Across the considered load levels, the algorithm converges to an adaptive PRB allocation policy that ensures V2X SLA compliance and improves resource efficiency for each partition.The results are reported for a multi-slice scenario.Source: §VI · p. 12
  • In simulation deployment under the most restrictive SLA, the learned policy responds effectively to both SLA-violation and over-provisioning scenarios.The deployment evaluation is shown through results from the first and second experiments.Source: §VI · p. 13, Figure 7 · p. 13, Figure 8 · p. 13
  • Real-network validation remains future work, including assessment of feasibility, performance under realistic operating conditions, and practical challenges beyond simulation.The methodology is presented as extendable to other SLA and throughput targets and network loads through appropriate training.Source: §VI · p. 13
Loading 2609.27659v1…