Pith. sign in

REVIEW 3 major objections 4 minor 15 references

Deep Reinforcement Learning for Backhaul Link Selection for Network Slices in IAB Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A Double Deep Q-Network with one hidden layer of 32 ReLU neurons can select which neighboring base station should wirelessly backhaul each network slice of a congested 5G base station, converging in about 21 episodes and earning the…

desk verdict A small DDQN can learn a capacity-feasibility lookup in a simplified IAB scenario, but the reported perfect test score is probably an artifact of a chronological split that puts the test phase entirely outside congestion windows. read the letter →

arxiv 2502.05707 v1 pith:S4NOTYCZ submitted 2025-02-08 cs.NI

António J. Morgado , Firooz B. Saghezchi , Pablo Fondo-Ferreiro , Felipe Gil-Castiñeira , Maria Papaioannou , Kostas Ramantas , Jonathan Rodriguez This is my paper · ORCID
classification cs.NI
keywords integratedaccessandbackhauldeepreinforcementlearningnetworkslicinglinkselectionDDQN5GresourceallocationIABdonor
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

The paper proposes using Integrated Access and Backhaul (IAB), the 5G feature that lets base stations with wired backhaul lend wireless capacity to neighboring base stations, to relieve a congested base station's network slices. Because traffic demands and slice combinations change over time, the authors model the choice of which neighboring base station should backhaul each slice as a deep reinforcement learning problem. They report that a Double Deep Q-Network with one hidden layer of 32 ReLU neurons selects feasible backhaul links successfully, converges in about 21 training episodes, and earns the maximum possible reward throughout the test phase in their simulator. The point is that a very small neural-network critic may be enough for slice-level backhaul selection, making the approach practical for deployment and retraining.

What carries the argument

The central object is a Double Deep Q-Network (DDQN) agent, a value-based reinforcement-learning method that estimates the optimal state-action value $Q^*(s,a)$ with a fully connected neural network. The agent's input is a 23-dimensional vector that concatenates a 16-parameter observation (the slice's required DL/UL throughput plus the currently free DL/UL bandwidth of BS1 and BS2–BS7) with a 7-parameter one-hot action vector selecting one backhaul link; the output is the predicted reward for that slice-link pairing. The reward is +1 exactly when the selected directional link can carry the slice's required throughput to the core network, which makes the maximum achievable episode reward easy to compute and lets the authors measure convergence against a known ceiling.

What would settle it

Run the trained DDQN agent on a scenario where IAB links share the same spectrum and interfere with each other's access traffic; if the test-phase reward drops below the maximum attainable 60, the orthogonal-resources assumption is load-bearing and the reported no-failure performance does not extend to interference-limited operation.

Watch

Extended reading notes

Core claim

The central claim is that a DDQN agent, whose critic is a fully connected neural network with a single ReLU hidden layer of 32 neurons, can decide every 15 minutes which of the six neighboring base stations should wirelessly backhaul each of three slices of a congested base station, and can do so without a single failed allocation during testing. In the simulation, the agent receives a reward of +1 whenever the selected link provides the slice's required DL and UL throughput and 0 otherwise, so the maximum attainable per-episode reward is 288 and the maximum test-phase reward is 60. The authors report that the chosen network reaches 97.5 percent of the maximum training reward after 21 episodes and then collects the full 60-point test reward, meaning every tested slice allocation was feasible. They interpret the fast convergence and small model as evidence that the approach can be implemented with minimal computation on IAB base stations and can be retrained when traffic profiles change.

Load-bearing premise

The model assumes that all wireless IAB links use orthogonal resources and never interfere with one another or with any incumbent network, which turns the selection problem into a pure capacity-feasibility check on the observed free bandwidth.

Editorial extensions

If this is right

  • In the simulated scenario, a single hidden layer with 32 neurons is enough; deeper and wider critics do not systematically improve the test reward, so the learning task is simple enough for a small model.
  • Because training stops after about 21 episodes, the agent can be retrained quickly when surrounding traffic profiles change, which the authors identify as important for practical deployment.
  • The reward model gives an exact upper bound (288 per episode, 60 in the test split), so the reported convergence and no-failure test results are checkable quantities rather than relative improvements.
  • The approach extends the prior use of DRL in IAB from radio-resource allocation to network-slice-level backhaul selection, which is the paper's stated contribution.

Reading between the lines

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

  • If the orthogonal-resources assumption is relaxed, the problem becomes one of managing interference, and the same state representation would need to include channel conditions or interference measurements; a natural test is to retrain the agent on a shared-spectrum scenario and compare the test reward.
  • A simpler non-learning policy could probably match this performance in the current setup, because the decision is a capacity-feasibility lookup; comparing the DDQN agent against such a greedy baseline would clarify whether learning adds value beyond the simulator's assumptions.
  • The action model decides each slice sequentially with one action at a time, so the paper leaves open the question of jointly optimizing all slices; extending to multi-agent or joint-action formulations would be needed for larger topologies, as the authors note for future work.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a Double Deep Q-Network (DDQN) agent to select, every 15 minutes, a backhaul link (the wired BS1 or one of six IAB donor base stations) for each of three network slices served by a congested base station in an Integrated Access and Backhaul (IAB) network. The agent observes slice throughput requirements and the available bandwidth of candidate links, and receives a reward of +1 if the selected link provides the required QoS and 0 otherwise. Simulations with a custom SDN-based environment show that a single-hidden-layer network with 32 neurons reaches near-optimal episode reward after about 21 episodes and attains a perfect test reward of 60/60. The paper claims this demonstrates that the DRL agent performs the BS selection task successfully and can be implemented with minimal computational impact.

Significance. If the evaluation were sound, the paper would provide useful evidence that a very compact DDQN can solve a capacity-feasibility backhaul selection task in a deterministic simulated IAB scenario, which could have implications for lightweight and fast-retraining IAB control. The problem formulation is clear, the action/observation/reward scheme is straightforward, and the hyperparameter sweep across network sizes is a useful engineering exploration. However, the reported test result is undermined by a likely artifact of the chronological train/validation/test split, and the lack of baselines, multiple seeds, and out-of-distribution evaluation means the central claim is not yet supported. The paper is a plausible candidate after substantial revision and re-evaluation.

major comments (3)
  1. [Section V-F, Table IV and Section V-C] The reported perfect test-phase reward (60/60) does not substantiate the claim that the agent learned a valid IAB donor-selection policy. If the split is chronological, as the terms 'training phase', 'validation', and 'test phase' suggest, the 60 test samples correspond to the last 20% of the 288-sample day (roughly 19:00-24:00), which lies entirely outside the congestion windows identified in Section V-C (05:30-12:30 for UL and 14:30-18:30 for DL). A constant policy that always selects the wired backhaul (BS1) would receive +1 on every test sample because BS1's wired link has sufficient capacity in those hours. The paper must specify whether the split is chronological or random; if it is chronological, the test-phase metric is trivial and must be replaced by a test regime that includes congested periods, for example a held-out day with shifted traffic profiles or randomly sampled time slots across the whole day. The performance of the trivial always-BS1 policy should also be reported as a reference.
  2. [Table IV and Section V-A] The results are reported for a single run per architecture with no multiple seeds, confidence intervals, or significance testing. DDQN training is stochastic, so the claimed fast convergence in 21 episodes and the perfect test reward need variance estimates. Moreover, no baseline is compared, such as a greedy 'any feasible donor' policy, a least-loaded donor policy, or the trivial always-wired policy. Without such baselines, the paper cannot support the conclusion that DRL is the right tool for this task or that the learned policy is superior to simple heuristics. The authors should at least add the always-BS1 baseline and a greedy feasibility policy to the comparison in Table IV and report mean and standard deviation over multiple seeds.
  3. [Section V-F, early stopping] The early-stopping criterion is based on the moving average of the episode reward reaching 97.5% of the maximum achievable value over a training episode (0.975 * 201 = 195.975). Since the training portion is the first 70% of the day, which contains the congestion windows, this threshold could be achieved by a policy that simply selects BS1 during non-congested hours and any feasible donor during congested hours, without learning a sophisticated allocation. The paper should report the reward over the full episode (all 288 samples) for the trained agent, not only over the training/validation/test fractions, to demonstrate that the agent indeed handles the congested periods correctly.
minor comments (4)
  1. [Section V-F, Fig. 8] The value '1,479 Gbps' appears to be a typo; given the 1 Gbps link capacities and the scale of the traffic profiles, the intended value is likely 1.479 Gbps (or 1479 Mbps).
  2. [Section III.A.1] The assumption that wireless IAB links operate in orthogonal resources and therefore do not interfere with each other or with incumbent networks removes the central radio-resource coupling of in-band IAB. This limitation should be discussed more prominently, since in real deployments interference and shared spectrum are core to IAB operation and would affect both the state representation and the learned policy.
  3. [Sections IV.A and IV.B] The action vector has seven options and the observation vector has 16 parameters, but the paper does not specify the exact ordering of the observation vector entries or the normalization method. A precise definition would improve reproducibility.
  4. [Section IV.C] The reward model gives +1 for any feasible link and 0 otherwise; this binary immediate reward does not account for future capacity consumption or the impact of sequential decisions on other slices. At least a brief discussion of why this reward is adequate is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the DDQN result is a measured training outcome, and the reward/performance identity is a definition, not a derivation reducible to its inputs.

full rationale

The paper's central claim is empirical: a DDQN agent with one hidden layer of 32 ReLU neurons converges in 21 episodes and collects the maximum attainable test reward (Section V-F). This is a reported simulation outcome, not a first-principles derivation, so the usual circularity patterns (definitional equivalence, fitted input renamed as prediction, load-bearing self-citation, or uniqueness imported from the authors) do not apply. The reward model (+1 when the selected link meets the slice QoS requirements, 0 otherwise) defines what the RL agent optimizes; reporting the collected reward as the performance metric is standard for control tasks and is not circular. The episode maximum 288 = 3 slices x 24 hours x 60/15 is an arithmetic identity, and the test-phase maximum 60 equals the number of test samples; these are counting definitions, not reductions of the result to an input. The paper lacks external-benchmark or held-out-scenario validation, and the chronological train/validation/test split may weaken the generalization claim if the test phase falls outside congestion windows, but that is a soundness/robustness limitation, not circularity. No quoted step exhibits a derivation whose conclusion is equivalent to its premise. Hence no significant circularity is present.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small set of hand-picked hyperparameters and, more importantly, on domain assumptions that make the problem a deterministic capacity-feasibility lookup: no interference, fixed 1 Gbps links, throughput-only QoS, and repeated daily traffic profiles. These assumptions drastically simplify the IAB selection task relative to real deployments.

free parameters (9)
  • Learning rate alpha = 0.0001
    Hand-picked hyperparameter; affects convergence speed but not the feasibility condition. Reported in Table III.
  • Discount factor gamma = 0.99
    Hand-picked; the task has no meaningful temporal credit assignment beyond sequential slice decisions, so gamma is not physically motivated.
  • Epsilon decay rate = 0.01
    Exploration schedule chosen by hand; single-run behavior is sensitive to this schedule.
  • Minibatch size = 64 samples
    Hand-picked DDQN hyperparameter (Table III).
  • Replay buffer size = 10000 samples
    Hand-picked DDQN hyperparameter (Table III).
  • Target network update period = 4 timesteps
    Hand-picked DDQN hyperparameter (Table III).
  • Early-stopping threshold = 97.5% of max episode reward
    Ad hoc stopping rule; determines the reported convergence episodes, including the 21 episodes for the preferred configuration.
  • Train/validation/test split = 70/10/20 within each 288-sample episode
    Chosen split; test samples come from the same deterministic daily profiles as training samples.
  • Traffic profile assignment = p=((bs-1) mod 3)+1 for BS2-7
    Hand-assigned synthetic load profiles; this assignment shapes which donors are feasible at each time slot.
assumptions (5)
  • domain assumption Wireless IAB links operate in orthogonal resources (space domain) and do not interfere with each other or with incumbent networks (Section III.A.1).
    This removes interference and makes available capacity additive, converting the selection problem into a deterministic capacity-feasibility check.
  • domain assumption Each BS has a wired backhaul of 1 Gbps in DL and UL, and IAB links share a 1 Gbps DL and UL with access traffic (Section V.C).
    Scenario capacity numbers are not derived; they define the only resource constraint tested.
  • domain assumption Slice QoS is fully described by required DL and UL throughput; latency and other QoS dimensions are ignored (Sections III.B and IV.C).
    Reward +1 is granted iff requested throughput fits; no other SLA constraint is modeled.
  • domain assumption Traffic profiles are deterministic and identical every day (Figs. 4 and 5).
    The environment has no stochasticity; train and test data come from the same repeated daily cycle, limiting generalization claims.
  • standard math DDQN with function approximation converges to a good Q-function under the given hyperparameters (Section IV.B).
    No convergence guarantee is proven; the paper relies on standard DRL practice and empirical reward plots.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Reinforcement Learning for Backhaul Link Selection for Network Slices in IAB Networks." pith.science (2026). https://pith.science/paper/S4NOTYCZ

@misc{pith2026250205707,
  author       = {Pith},
  title        = {Pith review of: Deep Reinforcement Learning for Backhaul Link Selection for Network Slices in IAB Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4NOTYCZ}},
  note         = {Machine review of arXiv:2502.05707}
}
read the original abstract

Integrated Access and Backhaul (IAB) has been recently proposed by 3GPP to enable network operators to deploy fifth generation (5G) mobile networks with reduced costs. In this paper, we propose to use IAB to build a dynamic wireless backhaul network capable to provide additional capacity to those Base Stations (BS) experiencing congestion momentarily. As the mobile traffic demand varies across time and space, and the number of slice combinations deployed in a BS can be prohibitively high, we propose to use Deep Reinforcement Learning (DRL) to select, from a set of candidate BSs, the one that can provide backhaul capacity for each of the slices deployed in a congested BS. Our results show that a Double Deep Q-Network (DDQN) agent using a fully connected neural network and the Rectified Linear Unit (ReLU) activation function with only one hidden layer is capable to perform the BS selection task successfully, without any failure during the test phase, after being trained for around 20 episodes.

Figures

Figures reproduced from arXiv: 2502.05707 by the authors.

Figure 1
Figure 1. illustrates our addressed scenario, where all the seven 5G BSs (gNodeBs) have a wired (fiber optic) backhaul connecting them to the core network, and may also use IAB to borrow additional backhaul capacity from the neighboring BSs. We assume that the congested BS, i.e., the BS needing to borrow capacity from the neighboring BSs is BS1. Different aspects that need to be considered in this scenario include: the networ… view at source ↗
Figure 2
Figure 2. Simulator architecture the flow tables of the SDN switches according to the desired forwarding path for each network slice. The network topology ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Daily traffic of the network slices in BS1 D. Traffic Profiles of the Surrounding BS The surrounding IAB BSs (BS2 to BS7) have to primarily carry the access network traffic for their associated UEs, and for this reason, they can only lend to BS1 the remaining capacity for wireless backhauling [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Simulator achieves the optimum episode reward (288) steadily overfitting. Hence, we stop the training when the moving average of the episode reward, for a window size of 10, reaches 97.5% of the maximum achievable value over a training episode, i.e., 0.975 * 201 = 195.…
Figure 7
Figure 7. Figure 7: Training of DDQN agent using a critic with 1 layer and 32 neurons terms of the total throughput of BS1 (aggregated 3 slices, DL+UL). We see that the 3 slices requested an average traffic demand of 1,479 Gbps (the dashed constant line in the figure), which was first rea…
Figure 8
Figure 8. Figure 8: Evolution of average throughput achieved during training The fast training of the proposed DDQN model makes it interesting for practical implementation in real wireless environments where the channel conditions and traffic profiles may dynamically change. However, for …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Integrated Access Backhauled Networks,

    O. Teyeb, A. Muhammad, G. Mildh, E. Dahlman, F. Barac and B. Makki, "Integrated Access Backhauled Networks," 2019 IEEE 90th Vehicular Technology Conference (VTC2019-Fall), Honolulu, HI, USA, 2019, pp. 1-5

    work page 2019

  2. [2]

    A Novel Machine Learning-Based Scheme for Spectrum Sharing in Virtualized 5G Networks,

    A. J. Morgado, F. B. Saghezchi, S. Mumtaz, V. Frascolla, J. Rodriguez and I. Otung, "A Novel Machine Learning-Based Scheme for Spectrum Sharing in Virtualized 5G Networks," in IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 10, pp. 19691-19703, Oct. 2022,

    work page 2022

  3. [3]

    Integrated access and backhaul in 5G mmWave networks: Potential and challenges,

    M. Polese, M. Giordani, T. Zugno, A. Roy, S. Goyal, D. Castor, and M. Zorzi, “Integrated access and backhaul in 5G mmWave networks: Potential and challenges,”, in IEEE Communications Magazine, vol. 58, no. 3, pp. 62-68, March 2020

    work page 2020

  4. [4]

    Access and radio resource management for IAB networks using deep reinforcement learning,

    M. M. Sande, M. C. Hlophe, and B. T. Maharaj, “Access and radio resource management for IAB networks using deep reinforcement learning,” in IEEE Access, vol. 9, pp. 114218-114234, August 2021

    work page 2021

  5. [5]

    Deep reinforcement learning-based spectrum allocation in integrated access and backhaul networks,

    W. Lei, Y. Ye, and M. Xiao, “Deep reinforcement learning-based spectrum allocation in integrated access and backhaul networks,” in IEEE Transactions on Cognitive Communications and Networking, vol. 6, no. 3, pp. 970-979, September 2020

    work page 2020

  6. [6]

    Deep reinforcement learning-based spectrum allocation and power management for IAB networks,

    Q. Cheng, Z. Wei, and J. Yuan, “Deep reinforcement learning-based spectrum allocation and power management for IAB networks,” in 2021 IEEE International Conference on Communications Workshops (ICC Workshops), pp. 1-6, June 2021

    work page 2021

  7. [7]

    Routing and Resource Allocation for IAB Multi-Hop Network in 5G Advanced,

    H. Yin, S. Roy, and L. Cao, “Routing and Resource Allocation for IAB Multi-Hop Network in 5G Advanced,” in IEEE Transactions on Communications, vol. 70, no. 10, pp. 6704-6717, October 2022

    work page 2022

  8. [8]

    Instantaneous Load-Based User Association in Multi-Hop IAB Networks using Reinforcement Learning,

    M. M. Sande, M. C. Hlophe and B. T. Maharaj, "Instantaneous Load-Based User Association in Multi-Hop IAB Networks using Reinforcement Learning," GLOBECOM 2020 - 2020 IEEE Global Communications Conference, Taipei, Taiwan, 2020, pp. 1-6, doi: 10.1109/GLOBECOM42002.2020.9322230

    work page arXiv 2020

Show all 15 references
  1. [9]

    Intelligent Spectrum Allocation for mmWave Integrated Backhaul and Access Network

    X. Zhou and X. Dong, "Intelligent Spectrum Allocation for mmWave Integrated Backhaul and Access Network", 2021 9th International Conference on Intelligent Computing and Wireless Optical Communications (ICWOC), pp. 69-74, 2021

    2021

  2. [10]

    A Backhaul Adaptation Scheme for IAB Networks Using Deep Reinforcement Learning With Recursive Discrete Choice Model,

    M. M. Sande, M. C. Hlophe and B. T. S. Maharaj, "A Backhaul Adaptation Scheme for IAB Networks Using Deep Reinforcement Learning With Recursive Discrete Choice Model," in IEEE Access, vol. 11, pp. 14181- 14201, 2023, doi: 10.1109/ACCESS.2023.3243519

    2023

  3. [11]

    Resilient Topology Design for Wireless Backhaul: A Deep Reinforcement Learning Approach,

    A. Abdelmoaty, D. Naboulsi, G. Dahman, G. Poitau and F. Gagnon, "Resilient Topology Design for Wireless Backhaul: A Deep Reinforcement Learning Approach," in IEEE Wireless Communications Letters, vol. 11, no. 12, pp. 2532-2536, Dec. 2022, doi: 10.1109/LWC.2022.3207358

    2022

  4. [12]

    Deep reinforcement learning for resource management in network slicing

    R. Li, Z. Zhao, Q. Sun, I. Chih-Lin, C. Yang, X. Chen, et al., "Deep reinforcement learning for resource management in network slicing", IEEE Access, vol. 6, pp. 74429-74441, 2018

    2018

  5. [13]

    Network Slice Reconfiguration by Exploiting Deep Reinforcement Learning With Large Action Space,

    F. Wei, G. Feng, Y. Sun, Y. Wang, S. Qin and Y. -C. Liang, "Network Slice Reconfiguration by Exploiting Deep Reinforcement Learning With Large Action Space," in IEEE Transactions on Network and Service Management, vol. 17, no. 4, pp. 2197-2211, Dec. 2020

    2020

  6. [14]

    Dynamic reservation and deep reinforcement learning based autonomous resource slicing for virtualized radio access networks

    G. Sun, Z. T. Gebrekidan, G. O. Boateng, D. Ayepah-Mensah and W. Jiang, "Dynamic reservation and deep reinforcement learning based autonomous resource slicing for virtualized radio access networks", IEEE Access, vol. 7, pp. 45758-45772, 2019

    2019

  7. [15]

    Deep reinforcement learning with double Q-learning,

    H. van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double Q-learning,” in Proc. of the Thirtieth AAAI Conference on Artificial Intelligence, vol. 30, no. 1, pp. 2094–2100, March 2016

    2016

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.