Deep Reinforcement Learning-Based RAN Slicing with Efficient Inter-Slice Isolation in Tactical Wireless Networks

Abderrahime Filali; Diala Naboulsi; Georges Kaddoum

REVIEW 3 major objections 5 minor 38 references

Deep Reinforcement Learning-Based RAN Slicing with Efficient Inter-Slice Isolation in Tactical Wireless Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a two-stage, DRL-based slicing mechanism can keep tactical RAN slices isolated without wasting scarce bandwidth, with the TD3 implementation outperforming all baselines.

desk verdict A plausible two-stage DRL RAN slicing mechanism with a clear problem formulation, but the evaluation's variable user counts are incompatible with the fixed-dimensional DRL models, undermining the stated robustness results. read the letter →

arxiv 2506.09039 v1 pith:XYEQY5QY submitted 2025-06-10 cs.NI

Abderrahime Filali , Diala Naboulsi , Georges Kaddoum This is my paper

classification cs.NI

keywords RANslicingtacticalnetworksO-RANdeepreinforcementlearningsliceisolationbandwidthallocationTD3QoS

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Tactical networks have little bandwidth but many concurrent services, so they need to split the radio resource dynamically without letting one slice's demand degrade another slice's quality of service. This paper proposes the BS-IISI mechanism, a two-stage bandwidth allocation scheme: first a global agent divides the base station's total bandwidth among RAN slices, then each slice's agent partitions its own share among its users. Both stages are solved with deep reinforcement learning, and the mechanism adds constraints that only reconfigure a slice's bandwidth when it genuinely needs more or clearly has excess, which limits wasteful over-allocation and keeps slices isolated. The paper reports that the TD3-based implementation achieves the highest objective values across all tested numbers of users, outperforming DDPG, PPO, unconstrained variants, and the RSSI-IP baseline, while keeping reconfiguration cost and resource wastage low.

What carries the argument

The mechanism that carries the argument is the two-stage hierarchical allocation with gated reconfiguration. In the inter-slice stage, a global agent chooses the fraction $f_s^t$ of the total bandwidth $W$ for each slice, subject to a minimum-per-slice floor and to logical implication constraints: a slice flagged as needing resources must receive at least as much as before, and a slice flagged as having excess must not receive more. In the intra-slice stage, each slice's agent chooses user-level fractions $f_{u,s}^t$ of the slice bandwidth, subject to a sum constraint and per-user floors, which together enforce intra-slice isolation. The satisfaction utility $\Gamma^t_{u,s}$, a normalised function of the ratio of achieved rate to required rate $R_s^{\mathrm{req}}$, provides the reward signal for both agents and penalizes both under-provisioning and over-provisioning, so the same signal drives efficiency and isolation. The reconfiguration cost $\mathcal{C}^t_s$ records any drop in slice satisfaction after a bandwidth change, and the objective $\alpha\Gamma^t-(1-\alpha)\mathcal{C}^t$ is what the DRL agents maximize.

What would settle it

Run the trained TD3 intra-slice agents on a configuration where one slice's user count differs from its training value (for example, 19 eMBB users instead of 20) and observe whether the agent produces a valid allocation vector. If the fixed-size input layer rejects or misinterprets the different-sized observation, the claimed stability across total user counts from 108 to 300 cannot be reproduced, and the mechanism would need a variable-size input architecture to support the claim.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a RAN slicing mechanism can serve both goals at once: efficient use of scarce tactical bandwidth and inter- and intra-slice isolation defined through quality of service. The paper models user satisfaction with a utility function that peaks when the achieved data rate equals the required rate and decreases when the rate is either too low or too high, so giving a user more bandwidth than needed is penalized as waste. Slice-level reconfiguration constraints determine when a slice truly requires additional resources or has resources it can give up, and the global objective combines average slice satisfaction with average reconfiguration cost through a weight $\alpha$. Decomposing the problem into inter-slice and intra-slice DRL agents and training them with TD3, DDPG, or PPO yields, according to the paper, the best behavior with TD3: the TD3 implementation is claimed to achieve the highest objective function values in all tested configurations while maintaining stable reconfiguration cost and low resource wastage.

Load-bearing premise

The evaluation assumes that DRL models trained with fixed per-slice user counts (20 eMBB, 70 URLLC, 210 mMTC) can still produce valid bandwidth allocations when slices contain different numbers of users, even though the state and action vectors of the intra-slice agents have lengths equal to the number of users and the paper describes no padding or retraining to handle that change.

Editorial extensions

If this is right

A tactical RAN could run dynamic bandwidth sharing without sacrificing isolation: slices that need more resources get them only when demand genuinely requires it, and slices with excess give them up without being starved.
The TD3-based implementation, if correct, gives network operators a concrete algorithm choice: off-policy deterministic actor-critic with double critics outperforms DDPG, PPO, and the RSSI-IP baseline in this setting.
Deploying training in the non-real-time RIC and inference in the near-real-time RIC keeps DRL model updates off the field infrastructure, which matters for low-latency tactical operations.
Because over-allocation is penalized by the utility function, the mechanism is designed to avoid the resource wastage that the paper argues leads to frequent reconfiguration and weakened isolation.
The two-stage structure separates concerns: the global agent manages inter-slice isolation, and each slice's agent manages intra-slice isolation, so the two goals can be tuned independently through constraints and rewards.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the reported stability across total user counts from 108 to 300 depends on inference models trained on fixed per-slice counts (20 eMBB, 70 URLLC, 210 mMTC); a feedforward network has fixed input and output sizes, so a direct test would be to check whether the same trained agents can actually accept states with different numbers of users, or whether padding or retraining is silent
Beyond the paper: because the utility function already penalizes over-allocation, the isolation constraints mainly serve to stabilize reconfiguration; one could test whether a smaller reconfiguration penalty alone, without the logical implication constraints, reproduces the reported isolation.
Beyond the paper: the same two-stage hierarchical design could be lifted to multi-cell tactical deployments, with one global agent per cell and a higher-level agent arbitrating inter-cell bandwidth, if coordination overhead stays within tactical latency budgets.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the mechanism is thoughtfully put together and the paper is readable, but the empirical section has a flaw that takes down the central claim. The DRL models are trained on fixed per-slice user counts (20/70/210), then evaluated across total user counts from 108 to 300. Unless per-slice counts stay fixed for all those totals—which is impossible—the state and action vectors for the intra-slice agents change length. The paper describes no padding, masking, or retraining to handle that. So the robustness results in Figures 5–10 are not reproducible from the described method. This is not a minor omission; it is the evidence for the paper's main conclusion.

What is genuinely new: the two-stage bandwidth allocation with explicit reconfiguration constraints to avoid resource wastage, and the O-RAN deployment discussion. The problem formulation is coherent, the utility function in Eq. (9) is a reasonable way to penalize over-allocation, and the related-work classification table is helpful. Three DRL variants (TD3, DDPG, PPO) are implemented and compared against multiple baselines, including the RSSI-IP mechanism. The paper also describes model lifecycle management in the O-RAN context, which adds an architectural angle.

Soft spots beyond the main flaw: the reconfiguration cost used to measure inter-slice isolation is part of the training reward, so using it as an evaluation metric is partially circular, though the comparison with the WIC baselines partly mitigates this. There are no error bars on the averaged results, which makes it hard to judge the practical difference between TD3 and DDPG. The baseline set is not exhaustive; the WIC variants come from the same authors' implementations, and no code is released, which makes the variable-size issue harder to check independently.

Overall, the idea deserves attention and the flaw may be fixable with a clear description of how variable user counts are handled—for example, padding, masking, or retraining per configuration. But as submitted, the central empirical claim is unsupported. I would send this to peer review because the mechanism is interesting and the evaluation problem is likely correctable, but I would expect major revision before it could be accepted.

Referee Report

3 major / 5 minor

Summary. The paper proposes BS-IISI, a two-stage deep reinforcement learning (DRL) mechanism for RAN slicing in tactical wireless networks. In the first stage, a global agent allocates fractions of the total bandwidth to RAN slices subject to resource reconfiguration constraints aimed at inter-slice isolation; in the second stage, per-slice agents allocate the slice bandwidth among users subject to intra-slice isolation constraints. The objective combines a user satisfaction utility with a reconfiguration cost, and the paper describes deployment in an O-RAN architecture. Three DRL implementations (TD3, DDPG, PPO) are compared with unconstrained DRL variants and the RSSI-IP baseline, and the paper claims that TD3 consistently achieves the highest objective function values and robust performance as the number of users varies from 108 to 300.

Significance. If the empirical claims were valid, the mechanism would be a useful contribution to RAN slicing for tactical networks, particularly in formulating dynamic bandwidth sharing with isolation constraints and in detailing O-RAN integration and DRL lifecycle management. The problem formulation and the two-stage architecture are clearly presented. However, the evaluation contains a load-bearing flaw: the DRL policies are trained on fixed per-slice user counts and are then evaluated on configurations with different total user counts, although the state and action spaces of the intra-slice agents are defined as vectors of length equal to the number of users in the slice. The paper also evaluates algorithms on the same objective function that serves as their training reward, which weakens the comparative claims. As written, the central conclusion that TD3 outperforms all baselines across varying user counts is not supported.

major comments (3)

[Section VI-C with Section V-B (Eqs. (24)-(25)) and Table II] The claimed robustness across 108 to 300 users is not reproducible from the described method. The intra-slice state space in Eq. (24) is a vector of channel gains of length |U_s|, and the action space in Eq. (25) is [f_min^s, f_max^s]^{|U_s|}. Training is performed with fixed per-slice user counts (eMBB: 20, URLLC: 70, mMTC: 210), and Section VI-C states that the inference models were trained on these fixed counts. A feedforward TD3, DDPG, or PPO actor has fixed input and output dimensions, so it cannot process state vectors of different lengths at inference time. The paper describes no padding, masking, truncation, retraining, or variable-size architecture to handle configurations with fewer than 300 total users. Consequently, the objective function values in Fig. 5 and the downstream results in Figs. 6-10 cannot be generated by the trained models as described, and the robustness conclusion is unsupported.
[Section V-A (Eq. (23)) and Section VI-C] The primary evaluation metric is the same as the training objective, making the comparison partly circular. The reward R_g^t in Eq. (23) is exactly alpha*Gamma^t - (1-alpha)*C^t for valid actions, and the objective function plotted in Fig. 5 is Eq. (20a), with C^t defined in Eq. (15). Thus the DRL agents are being evaluated on the same function they were trained to maximize. Although the paper also reports user satisfaction, data rates, and resource wastage, the headline claim that TD3 'consistently achieves the highest objective function values across all configurations' is not independent evidence of superior trade-off performance; it largely reflects reward optimization. The authors should present the isolation and efficiency metrics separately and compare methods that were not trained on the same reward.
[Section VI-C, Fig. 5] The RSSI-IP baseline is evaluated with a different objective function, so the quantitative comparison is not apples-to-apples. The text explicitly states: 'For the RSSI-IP mechanism, the objective function value excludes the reconfiguration cost, as it is not considered in [21].' Since the reconfiguration cost C^t is nonnegative, subtracting it from the DRL objective changes the scale and meaning of the comparison. The claim that TD3 outperforms RSSI-IP on the objective function is therefore not established. A fair comparison would either compute RSSI-IP with the same full objective or report both components (satisfaction and reconfiguration cost) separately for all algorithms.

minor comments (5)

[Section VI-A] The text contains a typo: 'gNobeB' should be 'gNodeB'.
[Eq. (25)] The action vector notation repeats the index: a_s^t = [f_{1,s}^t, f_{1,s}^t, ..., f_{|U_s|,s}^t] should read [f_{1,s}^t, f_{2,s}^t, ..., f_{|U_s|,s}^t].
[Algorithm 1, line 6] The assignment 'f_s^{t-1} = f_s^t' is written backwards; the intended meaning is that the current bandwidth portion is kept from the previous slot, i.e., f_s^t <- f_s^{t-1}.
[Figures 3 and 5] The axis labels contain typos: 'Cummulative' should be 'Cumulative', and 'Number f users' should be 'Number of users'.
[Section VI-A, Table II] The table heading says 'Max. number of users per slice,' but these values are also described as the fixed training counts. Please clarify whether the evaluation at different total user counts uses proportional reductions of these per-slice counts, and if so, state explicitly how the model handles the changed dimensions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation in the BS-IISI mechanism; the fixed-dimensionality gap for varying user counts is a correctness/reproducibility issue, not circularity.

full rationale

The paper's derivation chain is self-contained with respect to its own optimization design. The utility function (Eq. 9) is taken from external prior work [32], the channel model follows 3GPP specifications, and the DRL algorithms (TD3, DDPG, PPO) and the RSSI-IP baseline are standard external references. The only self-citations ([33], [34]) support a background claim that DRL has been used for RAN slicing; they are not load-bearing for the proposed mechanism or its conclusions. The strongest candidate for circularity is that the reconfiguration cost C^t (Eq. 14) appears both in the training objective (Eq. 20a) and in the global reward (Eq. 23), and Section VI-C reports C^t and the objective-function value as performance metrics. This means Figs. 5 and 6 measure quantities the DRL agents were trained to optimize; however, this is a benchmarking/transparency limitation rather than a logical reduction of the paper's central claim, because the paper does not present these quantities as independent predictions from first principles, and the algorithm comparison is conducted across methods trained under the same objective. The more serious concern is the fixed state/action dimensionality of the intra-slice agents (Eqs. 24-25) versus evaluation configurations with total users from 108 to 300; that is a reproducibility and correctness gap, not circularity. No load-bearing step reduces by construction to its own inputs, and no self-citation chain forces the paper's conclusions.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a utility function imported from prior work, a self-defined reconfiguration cost used as isolation metric, hand-chosen parameters, and an unexplained assumption that DRL models can scale to variable user counts. No new physical or algorithmic entities are introduced beyond the mechanism itself.

free parameters (4)

alpha = 0.5
Weighting factor between slice satisfaction and reconfiguration cost in the objective function (Eq. 20a); chosen by hand, not optimized.
rho = 1.3
Elasticity parameter of the utility function in Eq. (9); set by hand and affects the shape of the satisfaction curve.
xi = 5
Elasticity parameter of the utility function in Eq. (9); set by hand and affects the normalization constant in Eq. (11).
Gamma_th = 0.8
Slice satisfaction threshold used in reconfiguration Constraints 3 and 4; chosen by hand and directly influences when slices are considered to have excess resources.

assumptions (5)

domain assumption 3GPP path loss model with shadow fading (Eq. 7) accurately represents tactical wireless channels
The simulation relies on a standard 5G channel model; the paper does not validate it against tactical propagation environments.
domain assumption Utility function in Eq. (9), imported from [32], correctly quantifies user satisfaction and resource wastage
The satisfaction metric underpins both the training objective and the evaluation; its fidelity is assumed from prior work.
ad hoc to paper Reconfiguration cost in Eq. (14) is a valid measure of inter-slice isolation
The paper defines isolation through this cost; it is not an externally validated or standardized metric.
domain assumption DRL agents can learn to satisfy logical constraints (20g)-(20i) solely through a -1 penalty for invalid actions
No formal guarantee is provided; the paper relies on RL exploration finding valid actions in a continuous space.
ad hoc to paper Fixed-size neural networks can generalize to variable numbers of users per slice
The evaluation varies total user count while training on fixed per-slice counts, requiring an unstated mechanism for variable input/output dimensions; this is likely false.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Reinforcement Learning-Based RAN Slicing with Efficient Inter-Slice Isolation in Tactical Wireless Networks." pith.science (2026). https://pith.science/paper/XYEQY5QY

@misc{pith2026250609039,
  author       = {Pith},
  title        = {Pith review of: Deep Reinforcement Learning-Based RAN Slicing with Efficient Inter-Slice Isolation in Tactical Wireless Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYEQY5QY}},
  note         = {Machine review of arXiv:2506.09039}
}

read the original abstract

The next generation of tactical networks (TNs) is poised to further leverage the key enablers of 5G and beyond 5G (B5G) technology, such as radio access network (RAN) slicing and the open RAN (O-RAN) paradigm, to unlock multiple architectural options and opportunities for a wide range of innovative applications. RAN slicing and the O-RAN paradigm are considered game changers in TNs, where the former makes it possible to tailor user services to users requirements, and the latter brings openness and intelligence to the management of the RAN. In TNs, bandwidth scarcity requires a dynamic bandwidth slicing strategy. Although this type of strategy ensures efficient bandwidth utilization, it compromises RAN slicing isolation in terms of quality of service (QoS) performance. To deal with this challenge, we propose a deep reinforcement learning (DRL)-based RAN slicing mechanism that achieves a trade-off between efficient RAN bandwidth sharing and appropriate inter- and intra-slice isolation. The proposed mechanism performs bandwidth allocation in two stages. In the first stage, the bandwidth is allocated to the RAN slices. In the second stage, each slice partitions its bandwidth among its associated users. In both stages, the slicing operation is constrained by several considerations related to improving the QoS of slices and users that in turn foster inter- and intra-slice isolation. The proposed RAN slicing mechanism is based on DRL algorithms to perform the bandwidth sharing operation in each stage. We propose to deploy the mechanism in an O-RAN architecture and describe the O-RAN functional blocks and the main DRL model lifecycle management phases involved. We also develop three different implementations of the proposed mechanism, each based on a different DRL algorithm, and evaluate their performance against multiple baselines across various parameters.

Figures

Figures reproduced from arXiv: 2506.09039 by the authors.

**Figure 2.** BS-IISI mechanism in a O-RAN framework. be updated. When an inference model suffers from a serious degradation in performance, the model management entity: (i) terminates the model, (ii) selects a new trained model from the catalog, and (iii) deploys it to the near-RT RIC. VI. SIMULATION RESULTS A. Simulation Settings The proposed BS-IISI mechanism is designed to meet the QoS requirements of any tactical RAN slice. … view at source ↗

**Figure 3.** Inter-slice training performance. tion and the reconfiguration cost defined in Eq. (20a) and (15), respectively. For the RSSI-IP mechanism, the objective function value excludes the reconfiguration cost, as it is not considered in [21]. Each algorithm is evaluated under various network configurations, particularly varying the number of users. The DRL algorithm’s inference model was trained on a fixed number of users… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Intra-slice training performance. 1 1 1 $ ! $"!" 1 1 # %$ # [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Objective function Eq. (20a) 1 1 1 # #! ! # " !" [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 7.** Figure 7: QoS achieved by eMBB users. bandwidth, ensuring that users’ data rates are consistently near the minimum required threshold (10 Mbps). The RSSIIP algorithm’s median degree of satisfaction is slightly lower than that of TD3. Its box plot is flat, as resources are unifo…

**Figure 8.** Figure 8: QoS achieved by URLLC users. forms TD3 in user satisfaction. Despite this, TD3 guarantees near-perfect user satisfaction, with an IQR between 0.9 and 1 (see Fig. 8b), demonstrating consistent satisfaction for most users. DDPG has several outliers below the required min…

**Figure 10.** Figure 10: Resource wastage by slice. a portion of the system’s total bandwidth. In the second stage, each slice partitions its allocated bandwidth among its users. The bandwidth allocation problem is formulated as an optimization task and solved using DRL algorithms. The BS-IIS…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

[21]

Multi-tenant resource sharing with equitable-priority- based performance isolation of slices for 5g cellular systems,

N. Yarkinaet al., “Multi-tenant resource sharing with equitable-priority- based performance isolation of slices for 5g cellular systems,”Computer Communications, vol. 188, pp. 39–51, 2022

work page 2022
[1]

New horizons in tactical communications: An overview of emerging technologies possibilities,

V. M. Baeza and L. C. Salor, “New horizons in tactical communications: An overview of emerging technologies possibilities,”IEEE Potentials, vol. 43, no. 1, pp. 12–19, 2024

work page 2024
[2]

5G RAN slicing for verticals: Enablers and challenges,

S. E. Elayoubi, S. B. Jemaa, Z. Altman, and A. Galindo-Serrano, “5G RAN slicing for verticals: Enablers and challenges,”IEEE Commun. Mag., vol. 57, no. 1, pp. 28–34, 2019

work page 2019
[3]

O-RAN: Towards an Open and Smart RAN,

O.-R. Alliance, “O-RAN: Towards an Open and Smart RAN,” , Tech. Rep., Oct. 2018, white Paper

work page 2018
[4]

5g o-ran potential for military communications,

D. C. Marqueset al., “5g o-ran potential for military communications,” in2023 International Conference on Military Communications and Information Systems (ICMCIS). IEEE, 2023, pp. 1–8

work page 2023
[5]

Applications of machine learning in resource manage- ment for RAN-slicing in 5G and beyond networks: A survey,

Y. Azimiet al., “Applications of machine learning in resource manage- ment for RAN-slicing in 5G and beyond networks: A survey,”IEEE Access, vol. 10, pp. 106581–106612, 2022

work page 2022
[6]

O-RAN Non-RT RIC Architecture,

O-RAN Working Group 2, “O-RAN Non-RT RIC Architecture,” , Tech. Rep., Feb. 2024, v05.00

work page 2024
[7]

O-RAN Near-RT RIC Architecture,

O-RAN Working Group 3, “O-RAN Near-RT RIC Architecture,” , Tech. Rep., Oct. 2023, v05.00

work page 2023

Show all 38 references

[8]

A Game Theoretic Framework for Distributed Mission Slice Allocation and Management for Tactical Networks,

F. Yanget al., “A Game Theoretic Framework for Distributed Mission Slice Allocation and Management for Tactical Networks,”Journal of Network and Systems Management, vol. 31, no. 1, p. 23, 2023

2023
[9]

Tactical Topology Optimization Methodology for Slice Aware and Reconfigurable Battlefield Networks,

A. Castañareset al., “Tactical Topology Optimization Methodology for Slice Aware and Reconfigurable Battlefield Networks,” inIEEE Military Commun. Conf. (MILCOM), Rockville, MD, USA, 2022, pp. 273–278

2022
[10]

Slice aware framework for intelligent and reconfigurable battlefield networks,

A. Castañares, D. K. Tosh, and C. A. Kamhoua, “Slice aware framework for intelligent and reconfigurable battlefield networks,” inIEEE Military Commun. Conf. (MILCOM), San Diego, CA, USA, 2021, pp. 489–494

2021
[11]

O-RAN-enabled Intelligent Network Slicing to Meet Service-Level Agreement (SLA),

J. Daiet al., “O-RAN-enabled Intelligent Network Slicing to Meet Service-Level Agreement (SLA),”IEEE Trans. Mobile Comput., 2024

2024
[12]

A Multi-Level Deep RL-Based Network Slicing and Resource Management for O-RAN-Based 6G Cell-Free Networks,

N. Ghafouriet al., “A Multi-Level Deep RL-Based Network Slicing and Resource Management for O-RAN-Based 6G Cell-Free Networks,” IEEE Trans. Veh. Technol., 2024

2024
[13]

Mobile Network Slicing under Demand Uncertainty: A Stochastic Programming Approach,

A. Gholami, N. Torkzaban, and J. S. Baras, “Mobile Network Slicing under Demand Uncertainty: A Stochastic Programming Approach,” arXiv preprint arXiv:2304.14556, 2023

2023 arXiv
[14]

Elastic O-RAN Slicing for Industrial Monitoring and Control: A Distributed Matching Game and DRL Approach,

S. F. Abedinet al., “Elastic O-RAN Slicing for Industrial Monitoring and Control: A Distributed Matching Game and DRL Approach,”IEEE Trans. Veh. Technol., vol. 71, no. 10, pp. 10808–10822, 2022

2022
[15]

Highly flexible RAN slicing approach to manage isolation, priority, efficiency,

D. Marabissi and R. Fantacci, “Highly flexible RAN slicing approach to manage isolation, priority, efficiency,”IEEE Access, vol. 7, pp. 97130– 97142, 2019

2019
[16]

Dynamic RAN Slicing with Effective Isolation under Imperfect CSI,

J. Zhang, Y. Zu, Y. Zhang, and B. Hou, “Dynamic RAN Slicing with Effective Isolation under Imperfect CSI,” inIEEE International Conf. on Comp. and Commun. Eng. Technol. (CCET). IEEE, 2022, pp. 238–242

2022
[17]

Wireless resource management in sliced networks based onisolationindexes,

L. Tianet al., “Wireless resource management in sliced networks based onisolationindexes,”inIEEEWirelessCommunicationsandNetworking Conference (WCNC). IEEE, 2021, pp. 1–6

2021
[18]

Energy-efficient deep reinforcement learning assisted resource allocation for 5G-RAN slicing,

Y. Azimi, S. Yousefi, H. Kalbkhani, and T. Kunz, “Energy-efficient deep reinforcement learning assisted resource allocation for 5G-RAN slicing,” IEEE Trans. Veh. Technol., vol. 71, no. 1, pp. 856–871, 2021

2021
[19]

Mobility aware and energy-efficient federated deep reinforcement learning assisted resource allocation for 5G-RAN slicing,

——, “Mobility aware and energy-efficient federated deep reinforcement learning assisted resource allocation for 5G-RAN slicing,”Computer Communications, vol. 217, pp. 166–182, 2024

2024
[20]

Resource allocation in an open ran system using network slicing,

M. K. Motalleb, V. Shah-Mansouri, S. Parsaeefard, and O. L. A. López, “Resource allocation in an open ran system using network slicing,”IEEE Trans. Netw. Service Manag., vol. 20, no. 1, pp. 471–485, 2022

2022
[22]

Data-Driven Estimation of Throughput Performance in Sliced Radio Access Networks via Supervised Learning,

C. Gijón, M. Toril, and S. Luna-Ramírez, “Data-Driven Estimation of Throughput Performance in Sliced Radio Access Networks via Supervised Learning,”IEEE Trans. Netw. Service Manag., 2022

2022
[23]

Latency Equalization Policy of End-to-End Network Slicing Based on Reinforcement Learning,

H. Bai, Y. Zhang, Z. Zhang, and S. Yuan, “Latency Equalization Policy of End-to-End Network Slicing Based on Reinforcement Learning,” IEEE Trans. Netw. Service Manag., vol. 20, no. 1, pp. 88–103, 2022

2022
[24]

An inter- slice rb leasing and association adjustment scheme in o-ran,

Y. Hou, K. Zhang, X. Liu, G. Chuai, W. Gao, and X. Chen, “An inter- slice rb leasing and association adjustment scheme in o-ran,”IEEE Transactions on Network and Service Management, 2023

2023
[25]

Analyticalmodelingandimprovement of interference-coupled ran slicing,

S.A.HashemianandF.Ashtiani,“Analyticalmodelingandimprovement of interference-coupled ran slicing,”IEEE Trans. Mobile Comput., 2024

2024
[26]

Hierar- chical reinforcement learning based resource allocation for ran slicing,

H. A. Akyıldız, Ö. F. Gemici, I. Hökelek, and H. A. Çırpan, “Hierar- chical reinforcement learning based resource allocation for ran slicing,” IEEE Access, 2024

2024
[27]

Random waypoint mobility model in cellular networks,

E. Hyytiä and J. Virtamo, “Random waypoint mobility model in cellular networks,”Wireless Networks, vol. 13, no. 2, pp. 177–188, 2007

2007
[28]

Study on new radio access technology: Radio access architecture and interfaces (Release 14),

3GPP, “Study on new radio access technology: Radio access architecture and interfaces (Release 14),” , Tech. Rep., March 2017, TR 38.801 v14.0.0

2017
[29]

Control, User and Synchronization Plane Specification,

O-RAN Working Group 4, “Control, User and Synchronization Plane Specification,” O-RAN.WG4.CUS.0-R004-v16.01, Tech. Rep., october 2024

2024
[30]

O-RAN Use Cases and Deployment Scenarios White Paper,

O. Alliance, “O-RAN Use Cases and Deployment Scenarios White Paper,”O-RAN ALLIANCE, Tech. Rep, 2020

2020
[31]

5G Study on channel model for frequencies from 0.5 to 100 GHz,

3GPP TR 38.901, “5G Study on channel model for frequencies from 0.5 to 100 GHz,” , Tech. Rep. V15.0.0, Release 15, July 2018

2018
[32]

Network slice selection in softwarization-based mobile networks,

G. Zhao, S. Qin, G. Feng, and Y. Sun, “Network slice selection in softwarization-based mobile networks,”Trans. on Emerging Telecom. Technol., vol. 31, no. 1, p. e3617, 2020

2020
[33]

Dynamic SDN-based radio access network slicing with deep reinforcement learning for URLLC and eMBB services,

A. Filaliet al., “Dynamic SDN-based radio access network slicing with deep reinforcement learning for URLLC and eMBB services,”IEEE Trans. Netw. Sci. Eng., vol. 9, no. 4, pp. 2174–2187, 2022

2022
[34]

Federated deep reinforcement learning for open ran slicing in 6g networks,

A. Abouaomar, A. Taik, A. Filali, and S. Cherkaoui, “Federated deep reinforcement learning for open ran slicing in 6g networks,”IEEE Commun. Mag., vol. 61, no. 2, pp. 126–132, 2022

2022
[35]

Addressing function approxi- mation error in actor-critic methods,

S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approxi- mation error in actor-critic methods,” inInternational conference on machine learning. PMLR, 2018, pp. 1587–1596

2018
[36]

Continuous control with deep reinforcement learning,

T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,”arXiv preprint arXiv:1509.02971, 2015

2015 arXiv
[37]

Proximal policy optimization algorithms,

J. Schulmanet al., “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

2017 arXiv
[38]

AI/ML Workflow Description and Require- ments,

O-RAN Working Group 2, “AI/ML Workflow Description and Require- ments,” O-RAN.WG2.AIML-v01.03, Tech. Rep., October 2021

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [21]

Multi-tenant resource sharing with equitable-priority- based performance isolation of slices for 5g cellular systems,

N. Yarkinaet al., “Multi-tenant resource sharing with equitable-priority- based performance isolation of slices for 5g cellular systems,”Computer Communications, vol. 188, pp. 39–51, 2022

work page 2022

[2] [1]

New horizons in tactical communications: An overview of emerging technologies possibilities,

V. M. Baeza and L. C. Salor, “New horizons in tactical communications: An overview of emerging technologies possibilities,”IEEE Potentials, vol. 43, no. 1, pp. 12–19, 2024

work page 2024

[3] [2]

5G RAN slicing for verticals: Enablers and challenges,

S. E. Elayoubi, S. B. Jemaa, Z. Altman, and A. Galindo-Serrano, “5G RAN slicing for verticals: Enablers and challenges,”IEEE Commun. Mag., vol. 57, no. 1, pp. 28–34, 2019

work page 2019

[4] [3]

O-RAN: Towards an Open and Smart RAN,

O.-R. Alliance, “O-RAN: Towards an Open and Smart RAN,” , Tech. Rep., Oct. 2018, white Paper

work page 2018

[5] [4]

5g o-ran potential for military communications,

D. C. Marqueset al., “5g o-ran potential for military communications,” in2023 International Conference on Military Communications and Information Systems (ICMCIS). IEEE, 2023, pp. 1–8

work page 2023

[6] [5]

Applications of machine learning in resource manage- ment for RAN-slicing in 5G and beyond networks: A survey,

Y. Azimiet al., “Applications of machine learning in resource manage- ment for RAN-slicing in 5G and beyond networks: A survey,”IEEE Access, vol. 10, pp. 106581–106612, 2022

work page 2022

[7] [6]

O-RAN Non-RT RIC Architecture,

O-RAN Working Group 2, “O-RAN Non-RT RIC Architecture,” , Tech. Rep., Feb. 2024, v05.00

work page 2024

[8] [7]

O-RAN Near-RT RIC Architecture,

O-RAN Working Group 3, “O-RAN Near-RT RIC Architecture,” , Tech. Rep., Oct. 2023, v05.00

work page 2023