Understanding Conflict and Compatibility Conditions for Agents in AI-RAN

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

Understanding Conflict and Compatibility Conditions for Agents in AI-RAN Arshia Zolghadr† , Joao F. Santos† , Imtiaz Nasim‡ , Deniz Aytemiz‡ , Nicholas J. Kaminski‡ , and Jacek Kibiłda† † Commonwealth Cyber Initiative, Virginia Tech, USA, e-mail: {arshiaz,joaosantos,jkibilda}@vt.edu ‡ Idaho National Laboratory, USA, e-mail: {imtiaz.nasim,deniz.aytemiz,nicholas.kaminski}@inl.edu

conflicts. While effective, this approach fails to recognize that not all detectable conflicts necessitate mitigation and that considering agents’ objectives can improve performance. For instance, in the case of a direct conflict, defined by O-RAN as two or more agents controlling the same RAN parameter [10], existing works typically treat any detected conflict as warranting arbitration. However, consider a throughput-maximization agent and an energy-saving agent that both monitor downlink throughput and control the Physical Resource Block (PRB) allocation [3], [9]. The throughput-maximization agent requests more PRBs when the measured throughput falls below its target, while the energy-saving agent requests fewer PRBs when the throughput is above its target. Although both control the same parameter and are in direct conflict over it, if the throughput performance stays within both of their targets, this can satisfy both objectives, making arbitration unnecessary. Motivated by this observation, we focus on designing a mitigation policy that accounts for agents’ objectives, enabling the agents to share control over parameters. To achieve this goal, we model a RAN agent as a sampled deadband feedback controller model [11], [12], reflecting the behavior of a typical RAN agent that monitors a given KPI and only takes action when its value falls outside its target objective [3]. Then, we show that a set of agents are compatible if there exists at least one network state, as characterized by their monitored KPIs, in which the objectives of all agents are simultaneously satisfied. On the contrary, a collection of agents is incompatible when no such network state exists, so at least one objective must be violated. This distinction allows us to propose a conflict mitigation policy that arbitrates only between incompatible RAN agents, thereby improving network performance and target performance compliance. The main contributions of this paper are as follows: • We provide the first control-theoretic formulation of AIRAN agents and define their compatibility conditions. • We show that the compatibility of an arbitrary number of agents over an arbitrary number of KPIs can be established through pairwise tests. • We propose a compatibility-aware conflict mitigation policy that enables compatible agents to share control of a parameter, while resorting to prioritizing a single agent only when their objectives are incompatible. • We quantify the impact of conflicts and compare state-ofthe-art priority conflict mitigation policy [7] against our proposed method in a representative AI-RAN scenario using

Abstract—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. Index Terms—AI-RAN, Conflict Mitigation, Control Theory

I. I NTRODUCTION Mobile networks are increasingly adopting AI-driven optimization of Radio Access Network (RAN) to achieve different objectives. This evolution is prominently reflected in the AI-RAN vision, which has gained substantial momentum through the establishment of a global consortium with broad participation from industry stakeholders [1]. Such AI-driven optimization can be realized through a collection of thirdparty agents deployed on the network infrastructure, e.g., dApps, xApps, and rApps in the context of Open RAN (ORAN) [2], that monitor Key Performance Indicators (KPIs) and make decisions to modify RAN parameters. However, agents can make conflicting decisions with the potential to degrade performance, cause instability, or disrupt network operation [3]. The existing literature on conflicts in mobile networks focuses on conflict detection and conflict mitigation as distinct problems. Conflict detection identifies interactions among agents, RAN parameters, and KPIs either based on agent subscription information or RAN data analysis [3]–[6]. Mitigation methods determine which actions take effect through different resolution mechanisms [7]–[9], assuming known or detected

1

NIST’s ns3-oran extension of the ns-3 simulator [13]. II. R ELATED W ORKS There is a growing literature exploring methods to detect conflicts based on the interactions between agents, RAN parameters, and KPIs. For example, the work of [3] adopted statistical profiles of agents in a sandbox environment to infer interactions between agents, parameters, and KPIs. In our prior works, we proposed data-driven methods to identify interactions based on the correlation [5] and similarity [6] between parameters and KPIs. While differing in their methods to identify interactions, existing works classify agents that control or influence the same parameter as being in conflict and require mitigation. This stringent approach overlooks cases in which multiple agents have compatible objectives and can share control or influence over a parameter without negatively affecting each other. In other words, existing conflict detection methods identify which agents interact on the same parameters, but do not determine whether mitigation is actually necessary. Another group of works explores methods to mitigate conflicts after they are detected. For example, distillationbased methods explore knowledge distillation to merge multiple Deep Reinforcement Learning (DRL)-based agents into a single aggregate agent trained to prevent conflicts [14]. However, this approach forfeits the modularity of independent agents, requires retraining, and can only be applied to DRL-based agents. Alternatively, prioritization-based methods (also referred to as rule-based in some papers) rank agents according to an arbitrary priority and only enact the actions of the highest-ranked agent [7]–[9], [15]. While these works adopt different prioritization strategies, they rely on the same mechanism: the actions of one agent take precedence, and others’ are suppressed, either permanently or at each decision epoch. This suppression is applied regardless of whether the other agents’ actions would conflict with the prioritized agent or contribute to achieving the same (or a similar) objective. The existing literature focuses on preventing potential effects arising from multiple agents interacting over the same parameters, without assessing the compatibility of their objectives. As a result, existing conflict mitigation methods can unnecessarily restrict the number of agents that can simultaneously control the network infrastructure, potentially curbing the achievable performance. To address this limitation, we examine the behavior of agents operating on the network infrastructure and leverage control-theoretic concepts to derive the conditions under which multiple agents can coexist and share control of the same parameters. III. S YSTEM M ODEL In the following, we present the RAN model and then define the agents as sampled deadband feedback controllers. We focus on the Near Real-Time RAN Intelligent Controller (NearRT RIC) because it is the RAN control domain in which xApps use E2 measurements and actions for near-real-time control and optimization of RAN elements and resources, and RAN

conflict mitigation is defined for multi-xApp operation [10], [16]. However, the framework can potentially be extended to other control domains of AI-RAN. A. RAN Model We consider a set of agents A = {1, . . . , A} operating in the Near-RT RIC. Let P denote the set of RAN parameters exposed to the agents, e.g., PRB allocation, transmit power, and handover threshold, and let K denote the set of KPIs available from the RAN. Each agent a ∈ A controls a subset Pa ⊆ P and monitors a subset Ka ⊆ K. Following the O-RAN definition of a direct conflict [10], for each p ∈ P we define the set of agents that control it as Ap = {a ∈ A : p ∈ Pa }.

(1)

Thus, there is a conflict over parameter p when |Ap |≥ 2. For such a parameter, let Kp =

⋃︂

Ka

(2)

a∈Ap

denote the KPIs monitored by at least one of the agents controlling p. For each k ∈ Kp , let yk (t) denote the reported value of KPI k at time t. We collect the KPIs relevant to parameter p in the vector y(t) = (yk (t))k∈Kp .

(3)

B. Modeling the Operation of RAN Agents Agents monitor KPIs and act only when a monitored KPI leaves a range of values they consider satisfactory. For example, a throughput maximization agent only requests additional resources when the throughput falls below its intended threshold [3]. This behavior corresponds to a sampled deadband feedback controller: the agent remains inactive while its monitored KPI lies within a range and acts to steer it back when it leaves that range [11], [12]. Leveraging the language of control theory, we define each RAN agent through the following characteristics. For every KPI k ∈ Ka monitored by agent a, let ra,k denote the agent’s target for k and δa,k > 0 the agent’s tolerance. At a time instant t, the control error is then ea,k (t) = yk (t) − ra,k .

(4)

We say that the KPI k is within the agent’s deadband when |ea,k (t)|≤ δa,k ,

(5)

with lower and upper bounds la,k = ra,k − δa,k ,

ha,k = ra,k + δa,k .

(6)

IV. C OMPATIBILITY A NALYSIS In this section, we define objective compatibility for the agents controlling a common RAN parameter and establish how this compatibility extends from pairs of agents to the full set. Objective compatibility concerns whether the agents’ deadbands can be satisfied simultaneously. Let Ap,k = {a ∈ Ap : k ∈ Ka }

(7)

denote the set of agents with a direct conflict over parameter p that monitor KPI k. For each k ∈ Kp , define Lk = max la,k , a∈Ap,k

Hk = min ha,k , a∈Ap,k

(8)

as the maximum lower bound and minimum upper bound on KPI k across the conflicting agents’ deadbands. Definition 1 (Objective Compatibility). The agents in Ap are said to be compatible if Lk ≤ Hk ,

∀k ∈ Kp .

(9)

For each k ∈ Kp , the interval [Lk , Hk ] is called the shared acceptable range for KPI k. If (9) fails for any k ∈ Kp , the agents are incompatible. The shared acceptable range contains the KPI k values that fall within the deadband of every agent in Ap,k . We next show that for a set of agents, it suffices to perform pairwise compatibility tests, which will be useful for implementing an efficient conflict mitigation mechanism. Proposition 1 (Pairwise Sufficiency). The agents in Ap are compatible if and only if, for every pair of distinct agents i, j ∈ Ap and every KPI k ∈ Ki ∩ Kj , max(li,k , lj,k ) ≤ min(hi,k , hj,k ). (10) (︁|Ap |)︁ There are 2 agent pairs to evaluate, with each pair i, j requiring |Ki ∩ Kj | KPI checks. Proof: Suppose first that the agents in Ap are compatible. For any distinct i, j ∈ Ap and any k ∈ Ki ∩ Kj , Definition 1 gives Lk ≤ Hk . Since i, j ∈ Ap,k , max(li,k , lj,k ) ≤ Lk ≤ Hk ≤ min(hi,k , hj,k ),

(11)

and therefore (10) holds. Then, suppose (10) holds for every pair and fix any k ∈ Kp . Since Ap,k is finite and non-empty, there exist i, j ∈ Ap,k such that Lk = li,k , Hk = hj,k . (12) If i = j, then Lk ≤ Hk because li,k ≤ hi,k . If i ̸= j, (10) implies li,k ≤ hj,k , and therefore Lk ≤ Hk . Since this holds for every k ∈ Kp , the agents in Ap are compatible by Definition 1. The above result is the one-dimensional case of Helly’s theorem [17]. Applying this independently to each k ∈ Kp gives the multi-KPI result.

Corollary 1 (Two-Agent Compatibility). Consider two agents monitoring the same KPI, with targets r1 and r2 and tolerances δ1 and δ2 . Let ∆r = |r1 − r2 |

(13)

denote their target separation. The agents are compatible if and only if ∆r ≤ δ1 + δ2 . (14) For equal tolerances δ1 = δ2 = δ, the width of the shared acceptable range is W = max(0, 2δ − ∆r).

(15)

Hence, the compatibility boundary is ∆r∗ = 2δ, at which the shared acceptable range collapses to a single KPI value. Based on Definition 1, we propose a compatibility-aware conflict mitigation policy. For each set Ap of agents in a direct conflict, the policy evaluates (10) for every pair i, j ∈ Ap and every KPI k ∈ Ki ∩ Kj . If the condition holds for every such pair and KPI, the agents are compatible and are allowed to share control of p without arbitration. If it fails for any pair on any shared KPI, the agents are incompatible, and the policy prioritizes the higher-priority agent following [7]. V. N UMERICAL E VALUATION In this section, we evaluate our compatibility-aware conflict mitigation policy through simulations of a direct conflict over the downlink PRB allocation. We use the canonical twoagent, single-KPI case to isolate the effect of compatibility on arbitration and runtime behavior. We first describe the simulation and conflict scenario, then the metric and baselines, and finally the numerical results. A. Simulation Setup and Conflict Scenario To assess the compatibility of agents in AI-RAN and evaluate our conflict mitigation policy, we use NIST’s ns3-oran module [13], an extension of ns-3 that models O-RANlike behavior and enables agents to monitor KPIs and control parameters of an LTE network. Our simulations consider a single-cell topology to isolate the effects of direct conflicts from those of inter-cell interference, handovers, and intercell coordination. We place stationary User Equipments (UEs) according to the 3GPP TR 38.901 [18] UMa deployment model, so changes in the reported throughput result from changes to the PRB allocation rather than mobility. We use constant-rate UDP downlink traffic so throughput changes reflect PRB control and cell capacity rather than transportlayer rate adaptation. Each run lasts 15 s with the 2 s used for initialization, and KPIs are evaluated over the remaining 13 s. We report median results across 100 independent runs with distinct random seeds, with shaded bands showing the interquartile range. We summarize our simulation parameters in Table I. We consider a network infrastructure hosting two agents with distinct objectives commonly adopted in the conflict detection literature [3], [9]: a Throughput Maximization (TM)

Component

Configuration

Simulator Topology

ns-3 with NIST’s ns3-oran module [13] Single-cell LTE/EPC; 1 eNB; stationary UEs uniformly placed in [10, 289] m 3GPP TR 38.901 UMa path loss, probabilistic LOS/NLOS, log-normal shadowing [18] 2110 MHz (LTE FDD band 1) eNB 46 dBm [19]; UE 23 dBm [20] Round-robin Ideal RRC model Configured DL PRB allocation available to the scheduler, selected from {6, 15, 25, 50, 75, 100} [19]; initial allocation of 50 PRBs Downlink constant-rate UDP; 1400-byte packets TM and ES; TTM = TES = 100 ms; both monitor DL throughput and control the DL PRB allocation 15 s per run; 2 s initialization; 100 runs per configuration

Traffic Agents Simulation duration

TABLE I: Simulation parameters.

B. Evaluation Metric and Baselines To quantify the performance impact of multiple agents sharing control of a parameter, we adopt the following metric: Service Level Agreement (SLA) Drift: We quantify the violation of each agent’s objective using SLA drift, adapting the normalized SLA-drift metric in [21] to the two-sided deadband objectives considered in this work. For agent a, let y(ti ) denote the reported throughput at sampling instant ti , and let la = ra − δa and ha = ra + δa denote its lower and upper deadband bounds. We define the instantaneous SLA drift as (︃ )︃ la − y(ti ) y(ti ) − ha da (ti ) = max 0, , . (16) la ha Thus, da (ti ) = 0 when the reported throughput lies within the agent’s deadband. When the lower or upper bound is violated, da (ti ) measures the relative distance to the corresponding deadband boundary. This extends the one-sided normalized SLA violation in [21] to objectives with both lower and upper acceptable bounds. We average the drift over the sampling instants in the evaluation interval: Neval 100 ∑︂ da (ti ), Neval i=1

(17)

Δr *

40 30 20 10 0

40

Δr *

30

20

10

0

agent and an Energy Saving (ES) agent, both monitoring the aggregate downlink throughput KPI, and controlling the downlink PRB allocation parameter. The TM increases the PRB allocation when the throughput falls below its acceptable range, whereas the ES reduces the allocation when its objective can be met with fewer radio resources [9]. Both agents monitor throughput and take actions every 100 ms, i.e., TT M = TES = 100 ms, consistent with the near-real-time control loop from O-RAN [16]. Since the throughput delivered by our simulated cell depends on both the PRB allocation and the traffic demand, we calibrate the agents’ target based on the achievable throughput under each evaluated PRB allocation to ensure they correspond to feasible network states.

Drifta =

Throughput (Mbps)

Carrier frequency Transmit power MAC scheduler RRC PRB control

SLA Drift (%)

Channel model

50

0

5 Ad Hoc Shared Control

10

Δr (Mbps)

15

Priority Mitigation

20

25

Proposed

Fig. 1: Impact of the coexistence between agents with different target separations, under our proposed policy and two baselines, in terms of combined SLA drift and their monitored downlink throughput. where Neval is the number of samples collected during the evaluation interval. The factor 100 expresses the result as a percentage. All evaluated throughput bounds satisfy la > 0, so the normalization in (16) is well-defined. To assess the impact of sharing control of a parameter on both the TM and ES agents, we report our results in terms of the combined SLA drift, as an average of DriftTM and DriftES . We compare our policy against two baselines representing its two operating choices: (i) Ad Hoc Shared Control, an unarbitrated baseline in which every request is applied upon arrival, so the most recent request determines the PRB allocation; and (ii) Priority Mitigation, which assigns priorities to the agents. Following [7], decisions from the prioritized agent remain effective when conflicting requests target the same parameter, limiting lower-priority actions. Our policy permits shared control only after establishing compatibility; otherwise, it invokes Priority Mitigation. Without loss of generality, the ES has a higher priority than the TM in our simulations. C. Evaluating Compatibility Conditions In this analysis, we evaluate how differences between the agents’ objectives impact their ability to coexist while sharing the control of a parameter. Since their shared acceptable range is determined by their target values and tolerances, we set equal tolerances, δTM = δES = 10 Mbps, so that both agents have acceptable ranges of the same width, and vary the separation between their targets, ∆r. We consider 10 active UEs generating 4.5 Mbps of traffic each, and use 28 Mbps (an achievable throughput under this traffic demand) as the calibration point for the agents’ targets. We place the targets symmetrically around this reference value, with rTM = 28 + ∆r/2 and rES = 28 − ∆r/2, yielding the shared

D. From Compatibility Test to Practical Coexistence In this analysis, we examine how network conditions affect the coexistence between agents. First, we consider an exemplar case where the TM and ES agents have equal tolerances of δTM = δES = 11 Mbps and a target separation of 7 Mbps, resulting in the shared acceptable throughput range [37, 52] Mbps. To assess how the traffic demand affects the achievable throughput, we vary the number of active UEs from 10 to 150, generating 1 Mbps of traffic each as shown in Figure 2. Here, the proposed policy permits shared control and is operationally equivalent to the Shared Control baseline, whose identical curve is omitted for readability. For up to 30 UEs, the traffic demand remains below the lower bound of the shared acceptable range. Consequently, the RAN cannot produce throughput within the shared acceptable range that satisfies both agents, despite their compatible objectives, and we can observe a high SLA drift across all evaluated policies. Once the number of active UEs reaches 50, the cell can deliver a throughput within the shared acceptable range, and the mitigation policies begin to diverge. By permitting shared

Throughput (Mbps)

60 50 40 30 20 10 0 70 60

SLA Drift (%)

50 40 30 20

Number of UEs Priority Mitigation ES deadband

TM deadband

150

100

50

0

30

10 10

acceptable throughput range [ 18 + ∆r/2, 38 − ∆r/2 ] Mbps. Varying ∆r allows us to reduce the alignment between the agents’ objectives while keeping their targets centered around a feasible network state. By (15), its width W decreases linearly with ∆r and collapses at ∆r∗ = 2δ = 20 Mbps. Figure 1 shows the results of our simulations. The Ad Hoc Shared Control maintains near-zero SLA drift while the agents are compatible and the shared acceptable range is attainable with the evaluated PRB allocations. As ∆r approaches ∆r∗ , the shared acceptable range narrows until the available PRB allocations can no longer induce a throughput inside it, and the SLA drift increases even though the agents remain compatible. In contrast, the Priority Mitigation suppresses one agent’s actions at the cost of violating that agent’s objective, even when the agents are compatible. Our proposed policy displays three separate regimes according to ∆r. For ∆r ≤ 17 Mbps, the SLA drift remains below 0.2%, indicating that the agents’ objectives are jointly satisfied. The throughput changes with ∆r, but it remains within the shared acceptable range. For 18 Mbps ≤ ∆r ≤ 20 Mbps, the agents remain compatible, but none of the discrete PRB allocations induces a throughput within the narrowing shared acceptable range. This separates objective compatibility from practical attainability: the shared acceptable range remains non-empty, but no available PRB configuration realizes it. Consequently, SLA drift increases to ≈5.8% despite compatibility, with no substantial throughput degradation. For ∆r > ∆r∗ , the shared acceptable range is empty and agents are incompatible. Only then does our policy perform arbitration, prioritizing the ES. The combined SLA drift grows with ∆r, reaching ≈26% at ∆r = 25 Mbps, as TM’s objective is no longer satisfied after its conflicting requests are suppressed. These results show that our compatibility test can identify conditions where agents are compatible and can share control of parameters without unnecessary arbitration.

Proposed Shared acceptable range

Fig. 2: Effect of network conditions on the coexistence between agents, as traffic demand increases with the number of active UEs, measured in terms of SLA drift and throughput. control, the proposed policy reduces the combined SLA drift to approximately 0.2%. In contrast, Priority Mitigation suppresses one agent and incurs a combined SLA drift of ≈4%, a 20× increase. These results show that, while our compatibility test can identify whether the agents’ objectives can be satisfied simultaneously, the network conditions and their impact on the monitored KPIs influence whether the agents can coexist in practice. Finally, we extend the analysis across a wide range of target separations and tolerances under three traffic demands, and measure their effects on SLA drift, as shown in Figure 3. The agents are designed around the 35 Mbps operating point used in our reference configuration and the 30 Mbps and 40 Mbps cases evaluate how their behavior changes as the network operates below or above this reference. The black line marks the theoretical compatibility boundary ∆r = 2δ from (15). When ∆r > 2δ, the agents are incompatible because their acceptable ranges do not overlap. However, even when ∆r ≤ 2δ, we can observe some SLA drift as the traffic demand moves away from the reference operating point. This occurs because network conditions impose additional constraints on the network state, potentially making the shared acceptable range unreachable despite the agents being compatible. This shows that compatibility determines whether a shared acceptable range exists, while available control actions and network conditions determine whether it is attainable in practice. VI. C ONCLUSION In this paper, to the best of our knowledge, we provided the first control-theoretic formulation of objective compatibility

Traffic demand = 30 Mbps

ACKNOWLEDGEMENTS

35

Δr *

14

30 25

δ (Mbps)

10

Compatible

20

8 15 6 10 4

Incompatible

SLA Drift (%)

12

5

2

Traffic demand = 35 Mbps Δr *

14

0

R EFERENCES

35

[1] AI-RAN Alliance, “AI-RAN Alliance Vision and Mission White Paper,” AI-RAN Alliance, White Paper, 2024. [Online]. Available: https://ai-ran. org/wp-content/uploads/2024/12/AI-RAN Alliance Whitepaper.pdf [2] J. F. Santos et al., “Managing O-RAN Networks: xApp Development from Zero to Hero,” IEEE Communications Surveys & Tutorials (COMST), vol. 28, pp. 800–840, 2026. [3] P. Brach del Prever et al., “PACIFISTA: Conflict Evaluation and Management in Open RAN,” IEEE Transactions on Mobile Computing (TMC), vol. 24, no. 10, pp. 10 590–10 605, 2025. [4] P. B. del Prever et al., “Predicting Conflict Impact on Performance in O-RAN,” arXiv preprint arXiv:2603.08685, 2026. [5] A. Zolghadr et al., “Learning and Reconstructing Conflicts in O-RAN: A Graph Neural Network Approach,” in IEEE Wireless Communications and Networking Conference (WCNC), 2025, pp. 01–06. [6] J. F. Santos et al., “Conflict Detection in AI-RAN: Efficient Interaction Learning and Autonomous Graph Reconstruction,” arXiv preprint arXiv:2601.13213, 2026. [7] C. Adamczyk and A. Kliks, “Conflict Mitigation Framework and Conflict Detection in O-RAN Near-RT RIC,” IEEE Communications Magazine (ComMag), vol. 61, no. 12, pp. 199–205, 2023. [8] A. Wadud et al., “QACM: Qos-Aware xApp Conflict Mitigation in Open RAN,” IEEE Transactions on Green Communications and Networking (TGCN), 2024. [9] A. Giannopoulos et al., “COMIX: Generalized Conflict Management in O-RAN xApps—Architecture, Workflow, and a Power Control Case,” IEEE Access, 2025. [10] O-RAN Alliance, “O-RAN Conflict Mitigation,” O-RAN Alliance, Technical Report O-RAN.WG3.TR.ConMit-R004-v01.00, Oct. 2024. [11] H. K. Khalil and J. W. Grizzle, Nonlinear systems, 2002, vol. 3. [12] A. Gelb and W. E. Vander Velde, Multiple-Input Describing Functions and Nonlinear System Design. McGraw-Hill, 1968. [13] W. Garey et al., “O-RAN with Machine Learning in ns-3,” in 2023 Workshop on ns-3 (WNS3), 2023. [14] H. Erdol et al., “xApp Distillation: AI-based Conflict Mitigation in B5G O-RAN,” Computer Networks, vol. 274, 2026. [15] A. Wadud et al., “xApp-level Conflict Mitigation in O-RAN, A Mobility Driven Energy Saving Case,” in IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), 2025, pp. 1–6. [16] O-RAN Alliance, “Near-RT RIC Architecture,” O-RAN Alliance, Technical Specification O-RAN.WG3.RICARCH-R003-v04.00, 2024. [17] E. Helly, “Über mengen konvexer körper mit gemeinschaftlichen punkten,” Jahresbericht der Deutschen Mathematiker-Vereinigung, vol. 32, pp. 175–176, 1923. [18] 3GPP, “Study on Channel Model for Frequencies from 0.5 to 100 GHz,” 3rd Generation Partnership Project (3GPP), Technical Report TR 38.901, 2022. [19] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station (BS) Radio Transmission and Reception,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 36.104, 2021. [20] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); User Equipment (UE) Radio Transmission and Reception,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 36.101, 2021. [21] D. Campos et al., “DREAMIN: Channel-Aware Inter-Slices Radio Resource Scheduling for Efficient SLA Assurance,” in IEEE International Conference on Communications (ICC), 2025.

30 25

δ (Mbps)

Compatible

20

8 15 6 10 4

Incompatible

SLA Drift (%)

12 10

5

2 0

Traffic demand = 40 Mbps

35

Δr *

14

30 25

δ (Mbps)

Compatible

20

8 15 6 10 4

Incompatible

SLA Drift (%)

12 10

5

2 0

8

16

Δr (Mbps)

24

Work supported through the INL Laboratory Directed Research & Development (LDRD) Program under DOE Idaho Operations Office Contract DE-AC07-05ID14517. This research was also partially funded by the NSF IUCRC WISPER, under NSF Award No. 2412872 and by the NSF US-Ireland R&D Partnership program under NSF Award No. 2421362.

0

Fig. 3: SLA drift for different target separations and tolerances, for traffic demands of 30, 35, and 40 Mbps. The line ∆r∗ = 2δ delimits the compatible and incompatible regions.

among AI-RAN agents. Through this model, we derived compatibility conditions for determining whether agents can share control of the same parameters while allowing their objectives to be satisfied simultaneously, and proved that, under the componentwise deadband model, compatibility for any number of agents and monitored KPIs can be established through pairwise tests. We evaluated our model through simulations using NIST’s ns3-oran module, and our results demonstrated that compatibility-aware mitigation can substantially reduce SLA drift relative to priority mitigation when compatible objectives are attainable. Future work will relax the assumptions of the current formulation by incorporating an explicit RAN plant model to assess attainability alongside compatibility and to analyze the closed-loop behavior of incompatible agents under the deadband nonlinearity. We will also consider heterogeneous sampling periods, coupled multi-KPI objectives, multiple shared control parameters, and indirect and implicit conflicts. Together, these directions would extend objective compatibility beyond the direct-conflict setting considered here toward more general coordination of heterogeneous agents in AI-RAN.

Related documents

• Understanding Conflict and Compatibility Conditions for Agents in AI-RAN#1216790
• Intermediate System to Intermediate System (IS-IS) Extensions in Support of Generalized Multi-Protocol Label Switching (GMPLS)#213276
• Multiuser Diversity Regarding System Performance, Design and Evaluation of Radio Resource Allocation#411830
• Value-Based Massive Access through Goal-Oriented Irregular Repetition Slotted ALOHA#1034382
• Value-Based Massive Access through Goal-Oriented Irregular Repetition Slotted ALOHA#978381