PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding Conflict and Compatibility Conditions for Agents in AI-RAN

Arshia Zolghadr, Joao F. Santos, Imtiaz Nasim, Deniz Aytemiz, Nicholas J. Kaminski, Jacek Kibilda

arXiv:2610.07415v1cs.NIcs.SYeess.SY

TL;DR

Existing conflict mitigation can arbitrate between agents that control the same parameter without checking whether their objectives can be met simultaneously. This paper models agents as sampled deadband controllers, derives pairwise compatibility tests, and uses compatibility-aware mitigation. In simulation, outcomes distinguish compatibility from whether the shared acceptable range is attainable, and a reported comparison gives lower combined SLA drift for the proposed policy than Priority Mitigation.

  • Problem

    Existing methods treat agents controlling the same parameter as conflicts requiring arbitration without assessing whether their objectives can be satisfied simultaneously.

    Source: §I · p. 1

  • Method

    The paper models agents as sampled deadband feedback controllers, derives compatibility conditions that can be tested pairwise, and arbitrates only when objectives are incompatible.

    Source: §I · p. 1, B. Modeling the Operation of RAN Agents · p. 2, §IV · p. 3

  • Results

    SLA drift stays below 0.2% for ∆r ≤17 Mbps, rises to ≈5.8% for compatible ∆r from 18 to 20 Mbps when no discrete PRB allocation attains the shared range, and one comparison reports ≈0.2% versus ≈4% combined drift for the proposed policy and Priority Mitigation.

    Source: C. Evaluating Compatibility Conditions · p. 5, D. From Compatibility Test to… · p. 5

  • Takeaways & Limitations

    Compatibility is distinct from practical coexistence: even a non-empty shared acceptable range may be unattainable with available PRB allocations.

    Source: C. Evaluating Compatibility Conditions · p. 5

Abstract

from arXiv · show

Disaggregated mobile networks expose open interfaces for independent agents to control the Radio Access Network (RAN). The interactions between agents with contrasting objectives can result in conflicts. However, existing conflict mitigation methods treat agents that control the same parameter as being in a conflict that requires arbitration of their actions, without assessing whether their objectives can be satisfied simultaneously. In this paper, we introduce a control-theoretic approach to determine when agents can coexist and share control of the same parameters. We model threshold-based agents often encountered in the literature as sampled deadband feedback controllers and represent their objectives through Key Performance Indicator (KPI) targets and tolerances that define acceptable deadband bounds. We derive compatibility conditions from these bounds and show that compatibility for any number of agents and monitored KPIs can be established through pairwise tests. We propose a conflict mitigation policy that allows compatible agents to share control of the same network parameters while restricting arbitration to cases where their objectives cannot be satisfied simultaneously. We evaluate our policy through simulations using NIST's ns3-oran extension of ns-3, and our results show that it enables agents to share control with near-zero Service Level Agreement (SLA) drift when their objectives are compatible and the shared acceptable range is attainable, while matching Priority Mitigation when their objectives are incompatible.

I. INTRODUCTION

AI-RAN agents can control the same RAN parameter while pursuing different KPI objectives, but existing approaches may treat every such interaction as requiring arbitration. This paper models agents as sampled deadband controllers, derives compatibility tests, and proposes mitigation that arbitrates only when objectives are incompatible.

I. INTRODUCTION

  • Existing methods may arbitrate direct conflicts even when agents’ objectives can both be satisfied, as when throughput remains within both agents’ targets.The example concerns throughput-maximization and energy-saving agents controlling PRB allocation,; O-RAN defines direct conflict by shared control of a parameter.Source: §I · p. 1
  • The paper models RAN agents as sampled deadband feedback controllers and defines compatibility by whether a network state can satisfy all agents’ objectives simultaneously.Incompatible agents have no such state, so at least one objective must be violated.Source: §I · p. 1
  • Compatibility for arbitrary numbers of agents and monitored KPIs can be established through pairwise tests.The paper presents this as a contribution of its control-theoretic formulation of AI-RAN agents.Source: §I · p. 1
  • The proposed policy lets compatible agents share control and prioritizes one agent only when their objectives are incompatible.This policy focuses mitigation on cases where simultaneous satisfaction is impossible.Source: §I · p. 1
  • The evaluation compares the proposed method with Priority Mitigation in a representative AI-RAN scenario using NIST’s ns3-oran extension of ns-3.The supplied introduction passages describe the simulator and comparison, but do not report their outcomes.Source: §I · pp. 1–2

II. RELATED WORKS

Prior work detects agent interactions and mitigates detected conflicts, but does not assess whether agents’ objectives are compatible. The paper addresses this gap and situates its model in the Near-RT RIC.

II. RELATED WORKS

  • Conflict-detection work identifies interactions among agents, RAN parameters, and KPIs, while typically treating shared-parameter agents as requiring mitigation.Examples include statistical profiles and data-driven correlation and similarity [6] methods; these approaches do not establish whether mitigation is necessary.Source: §II · p. 2
  • Mitigation methods either merge DRL agents, sacrificing modularity and requiring retraining, or prioritize one agent and suppress others’ actions.Prioritization methods –, suppress lower-ranked agents regardless of whether their actions conflict with or support the prioritized agent’s objective.Source: §II · p. 2
  • Without objective-compatibility analysis, mitigation can unnecessarily restrict simultaneous control and potentially curb achievable performance.The paper therefore derives conditions for agents to coexist and share control of the same parameters.Source: §II · p. 2

III. SYSTEM MODEL

  • The model focuses on the Near-RT RIC, where xApps use E2 measurements and actions for near-real-time control and optimization; the framework may extend to other AI-RAN control domains.The stated focus reflects the multi-xApp conflict-mitigation setting; extension beyond it is presented as potential, not established.Source: §III · p. 2

A. RAN Model

The model specifies which RAN parameters and KPIs agents control or monitor, then identifies shared control of a parameter as a conflict. It also collects the relevant monitored KPI values for each parameter.

A. RAN Model

  • Each agent controls a subset of exposed RAN parameters and monitors a subset of available KPIs.Examples of parameters include PRB allocation, transmit power, and handover threshold.Source: A. RAN Model · p. 2
  • A parameter is in conflict when at least two agents control it; the model identifies the agents controlling each parameter.Source: A. RAN Model · p. 2
  • For each parameter, the model collects the KPI values monitored by its controlling agents into a vector.The KPI set includes those monitored by at least one agent controlling that parameter.Source: A. RAN Model · p. 2

B. Modeling the Operation of RAN Agents

The paper models threshold-based agents as sampled deadband feedback controllers, with KPI targets and tolerances defining acceptable ranges. Agents remain inactive within a satisfactory range and act when a monitored KPI leaves it.

B. Modeling the Operation of RAN Agents

  • Threshold-based agents are modeled as sampled deadband feedback controllers that act when a monitored KPI leaves its satisfactory range.The paper relates this behavior to throughput maximization agents requesting additional resources when throughput falls below a threshold.Source: B. Modeling the Operation of RAN Agents · p. 2
  • For each monitored KPI, an agent has a target and positive tolerance, which define lower and upper deadband bounds.Source: B. Modeling the Operation of RAN Agents · p. 2

IV. COMPATIBILITY ANALYSIS

Compatibility means that agents’ objectives can be satisfied simultaneously, represented by a nonempty shared acceptable range for each monitored KPI. Pairwise tests establish compatibility for the full agent set, and the policy uses that test to decide whether agents may share control.

IV. COMPATIBILITY ANALYSIS

  • Agents are compatible when their deadbands share an acceptable range for every monitored KPI.For each KPI, the shared range is bounded by the maximum lower bound and minimum upper bound across agents.Source: §IV · p. 3
  • Pairwise compatibility tests are sufficient to establish compatibility for any number of agents and monitored KPIs.The result applies the one-dimensional case of Helly’s theorem independently to each KPI [17].Source: §IV · p. 3
  • With equal tolerances, the compatibility boundary is Δr*=2δ, where the shared acceptable range collapses to one KPI value.The equal-tolerance range width depends on target separation and tolerance.Source: §IV · p. 3
  • The policy allows agents to share a parameter when every pairwise test passes; otherwise, it prioritizes the higher-priority agent.Tests apply to each pair and each KPI they share; prioritization follows.Source: §IV · p. 3

V. NUMERICAL EVALUATION

The evaluation simulates a direct conflict over downlink PRB allocation to examine compatibility-aware mitigation. A canonical two-agent, single-KPI case isolates compatibility’s effects on arbitration and runtime.

V. NUMERICAL EVALUATION

  • The evaluation simulates a direct conflict over downlink PRB allocation.The policy is assessed through simulations of agents controlling the same parameter.Source: §V · p. 3
  • A canonical two-agent, single-KPI case isolates the effect of compatibility on arbitration and runtime behavior.Source: §V · p. 3
  • The section presents the simulation and conflict scenario, metric and baselines, and numerical results.Source: §V · p. 3

A. Simulation Setup and Conflict Scenario

The simulations use NIST’s ns3-oran extension of ns-3 to model O-RAN-like behavior in an LTE network. A controlled single-cell, two-agent throughput conflict over PRB allocation isolates the effects of shared control.

A. Simulation Setup and Conflict Scenario

  • NIST’s ns3-oran extension of ns-3 models O-RAN-like behavior and agent control of LTE network parameters.Source: A. Simulation Setup and Conflict… · p. 3
  • A single-cell topology isolates direct-conflict effects from inter-cell interference, handovers, and inter-cell coordination.Source: A. Simulation Setup and Conflict… · p. 3
  • Runs last 15 seconds, with the first 2 seconds for initialization; reported medians aggregate 100 independent runs.Runs use distinct random seeds, and shaded bands show the interquartile range.Source: A. Simulation Setup and Conflict… · p. 3
  • The scenario pairs a Throughput Maximization agent with an Energy Saving agent, both monitoring downlink throughput and controlling PRB allocation.The agents act every 100 ms; their targets are calibrated to feasible throughput under evaluated PRB allocations.Source: A. Simulation Setup and Conflict… · pp. 3–4

B. Evaluation Metric and Baselines

The evaluation measures agents’ objective violations using SLA drift for two-sided deadbands and compares compatibility-aware mitigation with unarbitrated shared control and priority-based arbitration.

B. Evaluation Metric and Baselines

  • SLA drift quantifies each agent’s violation of its two-sided deadband objective.It adapts the normalized SLA-drift metric in; drift is zero inside the deadband and measures relative distance to the violated boundary outside it.Source: B. Evaluation Metric and Baselines · p. 4
  • Combined SLA drift is the average of TM and ES drift, averaged over evaluation samples and expressed as a percentage.The evaluation reports combined drift alongside monitored downlink throughput.Source: B. Evaluation Metric and Baselines · p. 4, Figure 1 · p. 4
  • Ad Hoc Shared Control applies each request upon arrival, while Priority Mitigation keeps the higher-priority agent’s decisions effective.The proposed policy shares control only after establishing compatibility and otherwise invokes Priority Mitigation; ES has higher priority in the simulations.Source: B. Evaluation Metric and Baselines · p. 4

C. Evaluating Compatibility Conditions

The analysis varies target separation while centering equal-width objectives on an achievable 28 Mbps throughput, distinguishing compatibility from whether discrete PRB allocations attain the shared range. The proposed policy avoids arbitration while objectives overlap, but drift rises when that range is unattainable or empty.

C. Evaluating Compatibility Conditions

  • At ∆r*=2δ=20 Mbps, the shared acceptable throughput range [18 + ∆r/2, 38 − ∆r/2] Mbps collapses, with targets centered on 28 Mbps and tolerances of 10 Mbps.The targets are set symmetrically around the calibration throughput.Source: C. Evaluating Compatibility Conditions · pp. 4–5
  • Priority Mitigation suppresses one agent’s actions, violating its objective even when the agents are compatible.The proposed policy instead permits shared control when the shared range is attainable.Source: C. Evaluating Compatibility Conditions · p. 5
  • SLA drift stays below 0.2% for ∆r ≤17 Mbps, rises to ≈5.8% at 18–20 Mbps despite compatibility, and reaches ≈26% at ∆r=25 Mbps after incompatibility triggers ES priority.The results distinguish a non-empty shared range from a range that available PRB allocations can actually realize.Source: C. Evaluating Compatibility Conditions · p. 5
  • For 18–20 Mbps target separation, no discrete PRB allocation reaches the narrowing shared range, so SLA drift rises despite compatibility and without substantial throughput degradation.This separates objective compatibility from practical attainability.Source: C. Evaluating Compatibility Conditions · p. 5

D. From Compatibility Test to Practical Coexistence

The section tests coexistence under changing traffic demand and network conditions. Shared control reduces SLA drift when the shared range is attainable, while compatibility alone does not guarantee that the network can reach it.

D. From Compatibility Test to Practical Coexistence

  • For up to 30 UEs, traffic demand is below the shared range’s lower bound, so all evaluated policies show high SLA drift despite compatible objectives.At 50 UEs, throughput can enter the shared range and the mitigation policies begin to diverge.Source: D. From Compatibility Test to… · p. 5
  • ≈0.2% combined SLA drift under shared control versus ≈4% under Priority Mitigation, a 20× increase, once throughput enters the shared acceptable range.The comparison is reported as traffic demand increases with active UEs.Source: Figure 2 · p. 5, D. From Compatibility Test to… · p. 5
  • Across 30, 35, and 40 Mbps traffic demands, network conditions can make the shared acceptable range unreachable even when target separation satisfies ∆r ≤2δ.The analysis therefore distinguishes whether a shared range exists from whether network conditions and available control actions can attain it.Source: D. From Compatibility Test to… · p. 5

VI. CONCLUSION

The conclusion presents a control-theoretic formulation of objective compatibility and derives a pairwise test for agents sharing parameters. Simulations indicate that compatibility-aware mitigation can reduce SLA drift when compatible objectives are attainable.

VI. CONCLUSION

  • The paper presents a control-theoretic formulation of objective compatibility for AI-RAN agents and derives conditions for sharing control while satisfying their objectives simultaneously.The authors describe this as the first such formulation to their knowledge.Source: §VI · pp. 5–6
  • Figure 3 plots SLA drift across target separations, tolerances, and traffic demands, marking ∆r*=2δ as the compatibility boundary.Source: Figure 3 · p. 6
  • Under the componentwise deadband model, compatibility for any number of agents and monitored KPIs can be established through pairwise tests.Source: §VI · p. 6
  • Compatibility-aware mitigation can substantially reduce SLA drift relative to priority mitigation when compatible objectives are attainable.The evaluation uses simulations with NIST’s ns3-oran module.Source: §VI · p. 6
  • Future work will add an explicit RAN plant to assess attainability and incompatible-agent closed-loop behavior, and examine heterogeneous periods, coupled KPIs, multiple shared parameters, and indirect conflicts.Source: §VI · p. 6
Loading 2610.07415v1…