Compositional Learning for Modular Multi-Agent Self-Organizing Networks

Parijat Bhattacharjee; Qi Liao

REVIEW 4 major objections 5 minor 17 references

Compositional Learning for Modular Multi-Agent Self-Organizing Networks

T0 review · 4 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 This paper claims that a modular two-tier agent architecture with compositional reward prediction—and especially a predictive decision-making variant that searches over small parameter steps—achieves a 37.2% reduction in handover…

desk verdict Two-tier cell-pair agent design is a real contribution, but CPDM's cell-level state omits the current TTT, so the headline result rests on an ill-posed action representation. read the letter →

arxiv 2506.02616 v1 pith:EPAETFCQ submitted 2025-06-03 cs.LG

Qi Liao , Parijat Bhattacharjee This is my paper

classification cs.LG

keywords self-organizingnetworksmulti-agentdeepreinforcementlearningcompositionalpredictivedecision-makingtwo-tieragentsmobilityrobustnessoptimizationhandoverfailuresampleefficiency

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

This paper argues that the control of a self-organizing network can be decomposed into a two-tier multi-agent system in which cell-level agents tune time-to-trigger (TTT) and cell-pair-level agents tune directional cell individual offsets (CIO), with the two tiers' rewards aligned so that a cell's reward equals the average of its cell-pair rewards. On top of this architecture the authors build two compositional learning methods: CDRL, which decomposes the critic into sub-predictors for handover cost, throughput/latency class, and radio link failure (RLF) anomaly, and CPDM, which uses those predictors to exhaustively select small step-size parameter changes instead of exploring actions. In a self-developed system-level simulation of a 32-cell Helsinki scenario with 12 days of training and 2 days of evaluation, CPDM achieves a 37.2% reduction in total handover failures compared to a conventional per-cell MADRL baseline, improves throughput and latency, cuts throughput and RLF anomalies, and is the only method that converges within the constrained training budget. The paper's central practical claim is that prediction-based decision-making with a modular, reusable agent design is more stable, safer, and more sample-efficient than end-to-end deep reinforcement learning for SON optimization.

What carries the argument

The load-bearing mechanism is the decomposition of the global SON optimization problem into two aligned tiers of small, fixed-dimension agents accompanied by a compositional objective. Cell-level agents observe aggregated cell-pair KPIs (the mean of neighboring-pair KPIs) and act on TTT; cell-pair-level agents observe both cells' parameters and KPIs plus the directional pair KPIs, and act on both directional CIOs. Reward alignment (the cell reward equals the average of its cell-pair rewards, per Remark 1) ensures that improving a CPA's reward also improves the associated CA's reward, so the two tiers can be trained alternately without conflicting gradients. The compositional value function replaces a single critic with three sub-critics that predict HO cost, throughput/latency class, and RLF anomaly, and the CPDM variant shrinks the action set to small step sizes so that the best action can be chosen by exhaustive inference rather than by exploration.

What would settle it

Run the same CPDM and MADRL training protocols on a public network simulator (for instance an LTE/NR system-level simulator distinct from the authors' own) or on the same simulator with five or more random seeds and a different map/traffic profile, and compare the 37.2% handover-failure reduction and the 0.02% RLF anomaly rate; if the advantage over MADRL shrinks or reverses, the finding is scenario-specific rather than a general property of the method.

Watch

Extended reading notes

Core claim

The central discovery is that compositionality—splitting the network into a two-tier agent hierarchy and splitting the value function into separately predictable network metrics—turns an intractable global MDP into a set of small, fixed-dimension learning problems, and that replacing exploratory action selection with an exhaustive search over small parameter steps makes the learning fast enough for practical training windows. The authors show that cell-level agents (CAs) controlling TTT and cell-pair-level agents (CPAs) controlling directional CIOs can be trained jointly under a centralized-training-decentralized-execution scheme, with order-agnostic sample augmentation that doubles the usable CPA experiences. Their compositional critics predict handover cost, throughput/latency class, and RLF anomaly with reported accuracies of 0.006 mean absolute error, 86.8% precision/81.1% recall, and 94.6% precision/92.0% recall respectively. CPDM uses these predictions to search over a five-step TTT change and 25-step CIO change space at each decision, selecting the maximum predicted value; this yields 33.3% fewer too-late handovers and 30.2% fewer too-early handovers relative to a fixed default configuration, with average throughput rising from 29.62 to 30.14 Mbps and downlink latency falling from 6.13 to 5.90 ms.

Load-bearing premise

The load-bearing premise is that the authors' self-developed system-level simulator faithfully represents real-world radio propagation, mobility, and traffic, and that the single scenario and single training/evaluation run are enough to establish the claimed improvements; if the simulator is not faithful or the run is unrepresentative, the 37.2% reduction may not generalize.

Editorial extensions

If this is right

CPDM reduces total handover failures by 37.2% relative to the per-cell MADRL baseline in the simulated scenario, and cuts too-late handovers by 33.3% and too-early handovers by 30.2% relative to the default configuration.
The two-tier heterogeneous agent design lowers state and action dimensions (fixed at 2(L(p)+L(ρ)+L(ψ)) for CPAs instead of growing with neighbor count), improving model reusability and scalability.
Compositional reward decomposition accelerates convergence over monolithic critics: CDRL converges faster than H-MADRL, and both improve on MADRL, which fails to converge within 12 days of training.
CPDM is more sample-efficient and stable than DRL-based methods: it achieves throughput anomalies of 6.26% and RLF anomalies of 0.02%, versus 28.84% and 0.63% for MADRL.
Prediction-based decision-making provides safer training because it avoids exploratory actions, searching only over small, bounded parameter changes.

Reading between the lines

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

I infer that the predictive decomposition, not the RL loop, is what drives the sample-efficiency gains: the three sub-predictors are trained by supervised losses on observed KPIs, so a purely supervised policy that selects the argmax action at each step would likely match CPDM's performance. A direct ablation replacing the TD3 actor with a supervised classifier on the same features would test this
The order-agnostic sample augmentation and aligned rewards could transfer directly to other pairwise coordination problems, such as inter-cell interference coordination or slice resource negotiation, where each pair of entities must agree on directional parameters.
Because CPDM searches a discrete step-size space, its advantage depends on the existence of small, safe adjustments between consecutive parameter settings; for parameters with unbounded or coarsely quantized action spaces, the exhaustive search would need a different search strategy.
The 37.2% figure comes from a single scenario and a single training/evaluation run; until multiple radio maps, traffic profiles, and random seeds are tested, the number should be read as indicative rather than as a guaranteed improvement.

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

Dear colleague,

The two-tier agent split is the real contribution here. Existing MADRL SON work uses per-cell agents whose state/action dimensions grow with neighbor count; the CA/CPA decomposition with fixed dimensions and the order-agnostic augmentation is a clean, reusable way around that, and Remark 1's reward alignment is a neat property worth stealing. The compositional sub-critic setup with symlog normalization is also sensible engineering.

The soft spots are not all equal. The missing seeds, single unreleased simulator, and no error bars are standard weaknesses, and the authors are at least explicit that the DRL baselines do not converge. That alone warrants a conditional verdict. But the stress-test note is right, and it is worse than the reader's report suggests. Section III-D defines CPDM actions as deltas, a_n(t)=p_n(t+1)-p_n(t), then Eq. (10) picks the delta by maximizing f^(C)(s_n,a_n). The CA state from Section III-A-1 and Table I is s_n=[rho_n,g0(psi_n)]; p_n(t) is not in it. Two different current TTTs with identical KPI features and the same delta produce different absolute TTTs and different rewards, so no f^(C) can correctly evaluate the deltas. The CPA version is fine because its state includes p_n and p_m. The cell-level CPDM results in Table II therefore rest on an unspecified or misspecified state/action representation. This is fixable by adding p_n(t) to the CA state for CPDM and retraining the predictor on the joint input, but as written the method is not defined. The CDRL branch does not have this issue because it uses absolute actions.

What else: the KPI table against DFLT is the most informative result, and the 33%/30% HO reductions are plausible. But the comparison to MADRL is weak because the baselines did not converge. The 'first to introduce cell-pair-level agents' claim needs a literature check; it may be true, but a couple of citations to earlier cell-pair formulations would make it credible.

My take: this deserves a serious referee. The core architecture and the cell-pair agent idea have value, and the main flaw is concrete and fixable rather than fatal to the whole program. I would send it out with a request that the CPDM formulation be corrected, the state/action semantics clarified, and the experiments rerun with multiple seeds and at least one more scenario.

Referee Report

4 major / 5 minor

Summary. The paper addresses self-organizing network (SON) optimization by proposing two compositional approaches: CDRL, which decomposes the reward/value function into sub-critics that predict intermediate KPI-related metrics, and CPDM, which uses prediction functions to exhaustively select step-size updates of control parameters (TTT and CIOs). Both operate within a two-tier multi-agent structure of cell-level agents (CAs) and cell-pair-level agents (CPAs) under centralized training with decentralized execution. The experiments in a self-developed Helsinki-like system-level simulator compare CDRL/CPDM against a conventional MADRL baseline, an ablation (H-MADRL), and a default configuration, reporting reduced handover failures, improved throughput/latency, and faster/safer training. The core claims are the 37.2% handover-failure reduction, CPDM's superiority to DRL-based methods, and better sample efficiency and safety.

Significance. If the results hold, the two-tier decomposition and the replacement of risky exploration by supervised prediction of KPI metrics are practically attractive for industrial SON, where online training must respect safety and limited time. The paper's formalization is clean in places: the agent definitions and dimension reduction in Table I are explicit, Remark 1 gives an aligned reward relation, and the CDRL loss in Eq. (9) is concrete. However, the significance is currently conditional: the empirical evidence is a single proprietary scenario with no uncertainty quantification, and the CPDM mechanism has a state/action identifiability gap that needs to be closed before the central results can be interpreted.

major comments (4)

[§III-D, Eq. (10), with §III-A-1 and Table I] The CPDM cell-level action in Eq. (10) is a step size a_n(t)=p_n(t+1)-p_n(t) in ΔA(C), but the CA state defined in Section III-A-1 and Table I is s_n(t)=[ρ_n(t), g0(ψ_n(t))], which does not contain the current TTT p_n(t). The predicted value and the realized reward depend on the absolute parameter p_n(t+1)=p_n(t)+a_n through the handover condition in Eq. (12); two states that coincide in ρ_n and ψ_n but arise from different p_n(t) yield different transitions and rewards for the same a_n. Thus no function f^(C)(s_n,a_n) can be a well-defined predictor unless p_n(t) is included in the state or the action is redefined as an absolute parameter choice. Moreover, if f^(C) is taken from the CDRL critic, it was trained on absolute actions p_n and feeding it deltas changes the action semantics; if it is retrained on deltas, the state is non-identifiable. The paper does not state how f^(C) is trained for CPDM. Since Table II attributes the main gains to CPDM, this issue must be resolved before the reported HO reductions can be accepted.
[§IV-B, Table II] The empirical support rests on a single evaluation scenario with a single training/evaluation run: 12 days of training and 2 days of evaluation in one self-developed simulator, with no multiple seeds, no confidence intervals, and no error bars in Table II or Figures 4-5. With DRL baselines non-converged, the statement in Section IV-B-3 that 'CPDM significantly outperforms all DRL-based methods' is not statistically supported. To support the central claim, report at least 5-10 independent seeds with mean±std (or confidence intervals) for all KPIs, and either extend the baseline training budget until convergence or include a converged reference policy so that the comparison is not between converged and non-converged methods.
[Abstract and §I vs §IV-B-4, Table II] The headline '37.2% reduction in total handover failures' is not reconciled with Table II. From HOLR and HOER, the default-to-CPDM reductions are 33.3% and 30.2%; if 'total handover failures' is defined as HOLR+HOER or any combination of HOLR, HOER, HOWR, and HOPPR in Table II, the implied reduction differs from 37.2%. Please define the metric exactly and show the computation; otherwise the central numerical claim in the abstract and introduction is unverifiable.
[§IV-B-5] The accuracy of the compositional predictive functions is reported only as aggregate numbers (0.006 MAE on HO cost, 86.8% precision, 81.1% recall, 94.6% precision, 92.0% recall), with the plots omitted 'due to the limited space.' Because these prediction functions are the mechanism by which CPDM selects actions, the paper should include confusion matrices or per-class performance tables so the reader can check that the prediction quality does not hide failure modes on rare but critical events such as RLF anomalies.

minor comments (5)

[Table II] The entries '0,63' and '0,35' should use decimal points ('0.63', '0.35'), and the RLF anomaly column should be formatted consistently with the other percentage columns.
[§IV-A, Eq. (12)] The notation in Eq. (12) mixes a condition 'RSRP_m > RSRP_n + q_{n,m} holds for p_n ms' with a mathematical inequality; consider writing the full event with a dwell time condition, since the dwell time is essential to the TTT semantics.
[§III-B-2] The text invokes 'alternating direction methods' and cites ADMM, but the convergence claim is only empirical; please make explicit that the alternating update is a heuristic and state its update schedule (e.g., how many CPA updates per CA update) in the algorithm description.
[§II and Table I] The definition of the state space dimension for the conventional neighbor-aware agent in Table I omits the cell-pair parameters in the state; clarify whether q_n is part of the action only, or also part of the observation, to make the dimension comparison unambiguous.
[§IV-A, rewards] The sentence defining the auxiliary function g2 contains a malformed formula ('/BD{...} + 0.5 · /BD{...} + /BD{...}'); please rewrite it in standard math notation and specify the indicator domains explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the main performance claims are measured simulator outcomes, not fitted predictions; self-citations are contextual rather than load-bearing.

full rationale

The paper's central quantitative claims, including the 37.2% handover-failure reduction and the KPI improvements in Table II, are measured in a system-level simulator rather than produced by the trained value functions or prediction networks. The CDRL/CPDM methods are compared against default and MADRL baselines on the same evaluation protocol, so the claimed improvements are empirical outcomes of an architectural comparison, not analytic consequences of the model's own definitions. The self-citations to prior work by the authors ([5], [9], [15]) are used for context, for a standard mean-aggregation mapping, and for the general motivation that predictive decision-making can be safer than DRL; none is invoked as a uniqueness theorem or as the sole justification for a derived result. There is some overlap between the reward construction and the reported anomaly metrics: the reward uses classifiers h2 and h3 whose labels are derived from historical/default-configuration data, and Table II reports throughput and RLF anomalies defined by deviations from the default configuration. However, the table reports actual measured anomalies, and Section IV-B-5 separately validates the prediction functions, so the evaluation is not the same as reporting a fitted value. The paper also explicitly states a limitation, namely that ADMM-style alternating training is supported only by empirical studies and not by a convergence proof, which lowers confidence but does not constitute circularity. Overall, no load-bearing step in the claimed derivation reduces to its own inputs by construction.

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

The main free parameters are reward shaping weights and classification thresholds; the central claims rest on the fidelity of the unreleased simulator and on an unproven decomposition of the value function.

free parameters (5)

Reward weights w1, w2 = not specified
Balance cell-level and cell-pair-level KPIs in rewards (3) and (4); values are chosen by hand and not reported, but they directly shape the optimization objective.
HO cost weights alpha0..alpha3 = 1, 1, 1, 0.2
Set manually in Section IV-A; they define the handover cost that agents optimize, directly influencing the learned policy.
Throughput/latency classification thresholds = Computed from historical data
Labels for the h2 predictor are derived from time-dependent distributions of observations; thresholds are fitted to training data and are part of the reward pipeline.
RLF anomaly classification thresholds = Computed from historical data
Binary anomaly labels for h3 are derived from historical deviations; the thresholds are fitted to data, not given a priori.
CPDM action step set for CIOs = {-2,-1,0,1,2} dB
The reduced action space is a hand-picked discrete set; the exhaustive search in CPDM depends on this choice.

assumptions (4)

domain assumption The multi-cell system can be modeled as a Markov decision process with discrete time steps and stationary transition dynamics.
Section II defines the MDP tuple and uses it for both CDRL and CPDM; real networks have non-stationary traffic and control loops that may violate the Markov property.
ad hoc to paper Decomposing the value function into sub-critics with prediction losses improves learning and preserves optimality.
Section III-C proposes the compositional critic; no proof is given that the decomposition is lossless or that the auxiliary prediction loss does not bias the value estimate.
domain assumption The self-developed simulator accurately reproduces real-world SON behavior.
Section IV-B uses the simulator as ground truth for all comparisons; no validation against field data or a public simulator is provided.
ad hoc to paper Alternating updates of cell-level and cell-pair-level agents converge to a good joint policy.
Section III-B-2 states 'we only provide empirical studies to show its effectiveness'; no convergence guarantee is established.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Compositional Learning for Modular Multi-Agent Self-Organizing Networks." pith.science (2026). https://pith.science/paper/EPAETFCQ

@misc{pith2026250602616,
  author       = {Pith},
  title        = {Pith review of: Compositional Learning for Modular Multi-Agent Self-Organizing Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EPAETFCQ}},
  note         = {Machine review of arXiv:2506.02616}
}

read the original abstract

Self-organizing networks face challenges from complex parameter interdependencies and conflicting objectives. This study introduces two compositional learning approaches-Compositional Deep Reinforcement Learning (CDRL) and Compositional Predictive Decision-Making (CPDM)-and evaluates their performance under training time and safety constraints in multi-agent systems. We propose a modular, two-tier framework with cell-level and cell-pair-level agents to manage heterogeneous agent granularities while reducing model complexity. Numerical simulations reveal a significant reduction in handover failures, along with improved throughput and latency, outperforming conventional multi-agent deep reinforcement learning approaches. The approach also demonstrates superior scalability, faster convergence, higher sample efficiency, and safer training in large-scale self-organizing networks.

Figures

Figures reproduced from arXiv: 2506.02616 by the authors.

**Figure 1.** CTDE-based two-tier agents. a reduced fixed dimension of features. For instance, an efficient and practical solution is to simply compute the mean of each cell pair KPI over cell pairs Cn [5]: g0 : R L (ψ)Nn → R L (ψ) : ψn 7→ h ψn,l : l = 1, . . . , L(ψ) i where ψn,l := 1 Nn X m∈N n ψn,m,l. (2) Let the action and state spaces of any CA be denoted by A(C) and S (C), respectively. We have A(C) := QL (p) l=1 Pl and S (… view at source ↗

**Figure 2.** Basic architecture of CDRL. impede learning progress. To address the dilemma, we use symlog normalization and apply a function selected from the bisymmetric logarithmic family: symlog(x) := sign(x) ln(|x| + 1). To read out the prediction and compute the loss function of the prediction, the inverse transformation is symexp(x) := sign(x)(exp(|x|) − 1). D. CPDM: Compositional Predictive Decision-Making Despite the be… view at source ↗

**Figure 3.** Simulated network scenario with 10 three-sector s [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Performance comparison: Reward evolution during [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Performance comparison of CDFs. 5) Accuracy of Compositional Predictive Functions: The compositional learning of the prediction functions achieves accurate prediction results, with 0.006 mean absolute error on HO cost, 86.8% precision and 81.1% recall on throughput cla…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

[1]

3GPP, TS 32.500: Self-Organizing Networks (SON); Concepts and re - quirements; Rel-17, 3GPP Technical speciﬁcation (TS), 2022

work page 2022
[2]

Trust and perfor mance in future AI-enabled, open, multi-vendor network management automation,

A. Banerjee, S. S. Mwanje, and G. Carle, “Trust and perfor mance in future AI-enabled, open, multi-vendor network management automation,” IEEE Transactions on Network and Service Management , vol. 20, no. 2, pp. 995–1007, 2023

work page 2023
[3]

Mobility-awa re load balanc- ing for reliable self-organization networks: Multi-agent deep reinforce- ment learning,

A. Mohajer, M. Bavaghar, and H. Farrokhi, “Mobility-awa re load balanc- ing for reliable self-organization networks: Multi-agent deep reinforce- ment learning,” Reliability Engineering & System Safety , vol. 202, p. 107056, 2020

work page 2020
[4]

A service-centric Q-learning algorithm for mobility robustness optimization in LTE,

M. L. Mar´ ı-Altozano, S. S. Mwanje, S. L. Ram´ ırez, M. Toril, H. Sanneck, and C. Gij´ on, “A service-centric Q-learning algorithm for mobility robustness optimization in LTE,” IEEE Transactions on Network and Service Management, vol. 18, no. 3, pp. 3541–3555, 2021

work page 2021
[5]

Inter -cell slicing resource partitioning via coordinated multi-agent deep re inforcement learning,

T. Hu, Q. Liao, Q. Liu, D. Wellington, and G. Carle, “Inter -cell slicing resource partitioning via coordinated multi-agent deep re inforcement learning,” in IEEE ICC , 2022

work page 2022
[6]

Ta sk-driven modular networks for zero-shot compositional learning,

S. Purushwalkam, M. Nickel, A. Gupta, and M. Ranzato, “Ta sk-driven modular networks for zero-shot compositional learning,” i n IEEE/CVF ICCV, 2019, pp. 3593–3602

work page 2019
[7]

Modula r deep learning,

J. Pfeiffer, S. Ruder, I. Vuli´ c, and E. M. Ponti, “Modula r deep learning,” arXiv preprint arXiv:2302.11529 , 2023

arXiv 2023
[8]

Single and multi-agent deep r einforcement learning for AI-enabled wireless networks: A tutorial,

A. Feriani and E. Hossain, “Single and multi-agent deep r einforcement learning for AI-enabled wireless networks: A tutorial,” IEEE Communi- cations Surveys & Tutorials , vol. 23, no. 2, pp. 1226–1252, 2021

work page 2021

Show all 17 references

[9]

Knowledge transfer in deep reinforcement learning for slice-aware mobility robustne ss optimization,

Q. Liao, T. Hu, and D. Wellington, “Knowledge transfer in deep reinforcement learning for slice-aware mobility robustne ss optimization,” in IEEE ICC , 2022, pp. 4262–4268

2022
[10]

Multi-agent actor-critic for mixed cooperative- competitive environments,

R. Lowe et al. , “Multi-agent actor-critic for mixed cooperative- competitive environments,” Advances in neural information processing systems, vol. 30, 2017

2017
[11]

Actor-Critic algorithms,

V . Konda and J. Tsitsiklis, “Actor-Critic algorithms, ” in NIPS, 1999

1999
[12]

Addressing function approximation error in Actor- Critic methods,

S. Fujimoto et al. , “Addressing function approximation error in Actor- Critic methods,” ArXiv, vol. abs/1802.09477, 2018

2018 arXiv
[13]

S. P . Boyd and L. V andenberghe, Convex optimization . Cambridge university press, 2004

2004
[14]

On ADMM in dee p learning: Convergence and saturation-avoidance,

J. Zeng, S.-B. Lin, Y . Y ao, and D.-X. Zhou, “On ADMM in dee p learning: Convergence and saturation-avoidance,” Journal of Machine Learning Research, vol. 22, no. 199, pp. 1–67, 2021

2021
[15]

Fast and scalable network slicing by integrating deep learning with Lagrangi an methods,

T. Hu, Q. Liao, Q. Liu, A. Massaro, and G. Carle, “Fast and scalable network slicing by integrating deep learning with Lagrangi an methods,” in GLOBECOM. IEEE, 2023, pp. 6346–6351

2023
[16]

3GPP, Radio Resource Control (RRC) protocol speciﬁcation; Rel-1 8, 3GPP Technical speciﬁcation (TS), 2024

2024
[17]

Microscopic trafﬁc simulation using SUMO,

P . A. Lopez et al. , “Microscopic trafﬁc simulation using SUMO,” in ITEC. IEEE, 2018. [Online]. Available: https://elib.dlr.de/1 24092/

2018

Pith tools

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

[1] [1]

3GPP, TS 32.500: Self-Organizing Networks (SON); Concepts and re - quirements; Rel-17, 3GPP Technical speciﬁcation (TS), 2022

work page 2022

[2] [2]

Trust and perfor mance in future AI-enabled, open, multi-vendor network management automation,

A. Banerjee, S. S. Mwanje, and G. Carle, “Trust and perfor mance in future AI-enabled, open, multi-vendor network management automation,” IEEE Transactions on Network and Service Management , vol. 20, no. 2, pp. 995–1007, 2023

work page 2023

[3] [3]

Mobility-awa re load balanc- ing for reliable self-organization networks: Multi-agent deep reinforce- ment learning,

A. Mohajer, M. Bavaghar, and H. Farrokhi, “Mobility-awa re load balanc- ing for reliable self-organization networks: Multi-agent deep reinforce- ment learning,” Reliability Engineering & System Safety , vol. 202, p. 107056, 2020

work page 2020

[4] [4]

A service-centric Q-learning algorithm for mobility robustness optimization in LTE,

M. L. Mar´ ı-Altozano, S. S. Mwanje, S. L. Ram´ ırez, M. Toril, H. Sanneck, and C. Gij´ on, “A service-centric Q-learning algorithm for mobility robustness optimization in LTE,” IEEE Transactions on Network and Service Management, vol. 18, no. 3, pp. 3541–3555, 2021

work page 2021

[5] [5]

Inter -cell slicing resource partitioning via coordinated multi-agent deep re inforcement learning,

T. Hu, Q. Liao, Q. Liu, D. Wellington, and G. Carle, “Inter -cell slicing resource partitioning via coordinated multi-agent deep re inforcement learning,” in IEEE ICC , 2022

work page 2022

[6] [6]

Ta sk-driven modular networks for zero-shot compositional learning,

S. Purushwalkam, M. Nickel, A. Gupta, and M. Ranzato, “Ta sk-driven modular networks for zero-shot compositional learning,” i n IEEE/CVF ICCV, 2019, pp. 3593–3602

work page 2019

[7] [7]

Modula r deep learning,

J. Pfeiffer, S. Ruder, I. Vuli´ c, and E. M. Ponti, “Modula r deep learning,” arXiv preprint arXiv:2302.11529 , 2023

arXiv 2023

[8] [8]

Single and multi-agent deep r einforcement learning for AI-enabled wireless networks: A tutorial,

A. Feriani and E. Hossain, “Single and multi-agent deep r einforcement learning for AI-enabled wireless networks: A tutorial,” IEEE Communi- cations Surveys & Tutorials , vol. 23, no. 2, pp. 1226–1252, 2021

work page 2021