Towards Scalable O-RAN Resource Management: Graph-Augmented Proximal Policy Optimization

Duc-Thinh Ngo (STACK; Kandaraj Piamrat; LS2N); Ons Aouedi; Philippe Raipin-Parv\'edy; Thomas Hassan

REVIEW 4 major objections 5 minor 19 references

Towards Scalable O-RAN Resource Management: Graph-Augmented Proximal Policy Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read GPPO claims graph-based reinforcement learning keeps O-RAN placements feasible and 18% cheaper at 64 radio heads.

desk verdict A reasonable engineering combination of GNN and masked PPO for O-RAN split/placement, but the headline generalization claim is contradicted by the paper's own evaluation protocol, and the closest prior work is never benchmarked. read the letter →

arxiv 2509.10499 v1 pith:UGH3WGWJ submitted 2025-09-01 cs.NI cs.AI

Duc-Thinh Ngo (STACK , LS2N) , Kandaraj Piamrat , Ons Aouedi , Thomas Hassan , Philippe Raipin-Parvédy This is my paper · ORCID

classification cs.NIcs.AI

keywords O-RANfunctionalsplitbasebandunitplacementgraphneuralnetworksproximalpolicyoptimizationactionmaskingresourceallocationreinforcementlearning

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 argues that the O-RAN resource-management problem—choosing a functional split and placing each radio head's vDU and vCU across edge and regional-cloud servers—cannot be solved well at scale by reinforcement learning agents that see only a flat vector of network features. It proposes GPPO, a PPO agent whose policy reads the substrate network through a graph neural network (a two-layer GINEConv encoder) and whose action space is pruned by connectivity masking. Across simulated networks of 8 and 64 radio heads, GPPO reports up to 18% lower deployment cost than a masked-PPO baseline, 25% higher reward in generalization tests, and feasible deployments in every training run; plain PPO and DDPG fail to find any feasible policy at 64 radio heads. If these simulation results transfer, the practical consequence is that topology-aware encodings are a necessary ingredient for scalable joint split-and-placement control in O-RAN.

What carries the argument

Two-layer GINEConv (Graph Isomorphism Network with edge attributes) message passing: each node's embedding is updated as MLP(x_i + sum over neighbors ReLU(x_j + e_ij)), so link delays and bandwidths enter the state representation through aggregation. Global mean pooling produces the graph-level embedding fed to PPO's policy and value networks. Action masking, applied to the action logits before softmax, confines the agent to placements that keep a valid RU-DU-CU path, reducing exploration of infeasible actions.

What would settle it

Train GPPO and MPPO on a fixed set of generated topologies and test on a disjoint set of topologies with different node counts and connectivity (never seen in training); if GPPO's success rate or cost advantage disappears, the topology-generalization claim is falsified. A cheaper check is to vary the cost coefficients across orders of magnitude and see whether the 18% gap persists; if the gap only appears for the paper's hand-set coefficients, the scalability result is an artifact of that cost structure.

Watch

Extended reading notes

Core claim

GPPO's central claim is that representing the O-RAN substrate as a graph and encoding node and edge attributes—remaining compute, traffic demand, latency requirement, link delay, bandwidth—with edge-aware graph isomorphism message passing lets a proximal-policy-optimization agent jointly decide functional split, vDU edge server, and vCU regional cloud for every radio head in one forward pass. The graph embedding is pooled into a global state vector feeding both policy and value networks, and action masks remove placements that violate hard connectivity. On the paper's simulated small topology, GPPO improves little over masked PPO alone; on the 64-radio-head topology the gap becomes qualitati

Load-bearing premise

The claimed gains assume the simulated topologies, delay/bandwidth samples, slice request distributions, and hand-set cost coefficients represent real O-RAN economics; the generalization test also trains and evaluates on the same five topologies, so transfer to truly unseen topologies is not demonstrated.

Editorial extensions

If this is right

If GPPO's results hold, joint functional-split and vDU/vCU placement is tractable for O-RAN networks at least up to 64 radio heads, where flat-vector RL baselines fail to produce any feasible policy.
Action masking alone is not enough: MPPO's 50% success rate at 64 RHs shows that topological awareness, not just constraint pruning, carries the scalability.
A single GPPO policy can serve multiple topologies with lower average cost and higher reward than masked PPO, suggesting learned policies are not purely topology-specific.
Perfect reliability at scale means an operator could deploy the policy without fallback heuristics for feasibility, provided the simulation cost structure matches real deployments.
The 18% cost reduction and 25% reward gain quantify the value of GNN encoding compared with the strongest masked-PPO baseline.

Reading between the lines

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

The paper trains the generalization model on five topologies and evaluates on the same five, so 'unseen' transfer is not actually tested; a direct test on topologies excluded from training—especially with different numbers of radio heads—would likely show a smaller or conditional advantage.
Because every cost coefficient and topology distribution is hand-set, the 18% figure is tied to that cost structure; varying the routing-cost and reconfiguration penalties could change which split choices dominate, and would be a cheap sensitivity check.
The same GNN-plus-masking architecture could be applied to other combinatorial resource-allocation problems with a graph substrate, such as service function chain placement or multi-access edge computing, where the paper's difficulty—combinatorial action space plus topology coupling—recurs.
A natural next experiment is to compare GPPO with an exact ILP solver on small instances to measure how far the learned policy is from optimal deployment cost; the paper reports only baselines, not optimality gaps.

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

The core idea here is a sensible engineering combination: use a GNN (GINEConv) to encode the O-RAN topology, feed the graph embedding to a masked PPO agent, and jointly decide functional split, vDU/vCU placement, and routing. The problem formulation is clear and the action-masking trick is a legitimate way to shrink an otherwise enormous action space. The paper also does something useful empirically: on their simulated 64-RH topology, GPPO achieves feasible deployments with decent cost while vanilla PPO and DDPG fail entirely, and masked PPO without the GNN only succeeds half the time. That is a meaningful result, even if it is what you would expect from the architecture.

The soft spots are not minor. The abstract claims "25% higher reward in generalization tests," but the generalization section explicitly says the model was trained on five 8-RH topologies and evaluated on the same five topologies. The text then interprets this as "unseen network structures." That is a direct contradiction. The experiment, as written, tests multi-task learning across five fixed topologies, not generalization to held-out ones. The fix is easy—train on a subset and test on the rest, or draw new topologies from the same generator—but as written, a headline claim is unsupported.

Second, the "state-of-the-art baselines" in the title and abstract are not actually compared. The closest previous methods, [5] and [6], are discussed in the literature review and even cited in Table I, but never benchmarked. The authors compare only against DDPG, vanilla PPO, and masked PPO with an MLP encoder. That is a fair ablation for the GNN component, but it does not show superiority over the prior DRL/GCN approaches that motivated the work.

Other concerns are less severe. The large-scale result (18.2% cost reduction) is based on a single randomly generated topology with hand-set cost coefficients, so external validity is inherently limited. There is no code or data release, which makes reproducibility hard to assess. The reward function includes the same cost components used for evaluation, but that is standard RL, not a hidden circularity.

Overall, the paper is a solid workshop-level or second-tier journal contribution in its current form, with a central method that is plausible and worth testing properly. It deserves a serious referee, but the authors need to fix the generalization experiment and ideally add comparisons to [5] and [6]. I would send it to peer review with a clear request for major revision, not desk reject it.

Referee Report

4 major / 5 minor

Summary. The paper addresses joint functional split selection and vDU/vCU placement in O-RAN via a PPO agent augmented with a GNN feature extractor and invalid-action masking (GPPO). It formulates a constrained optimization problem, recasts it as an MDP with a bounded reward, and evaluates on 8-RH and 64-RH simulated topologies against MLP-based PPO, DDPG, and masked PPO (MPPO). The reported results show GPPO achieving lower deployment costs and higher rewards, with 6/6 successful training runs at 64-RH versus 3/6 for MPPO, and an 18.2% cost reduction over MPPO. A 'generalization' experiment on five 8-RH topologies reports a 19% cost reduction and 25% reward improvement over MPPO, which the text interprets as evidence of transfer to unseen network structures.

Significance. If the results hold, GPPO would be a practically useful scalable method for O-RAN orchestration, since it combines graph-based state encoding with action masking and shows that generic MLP RL fails at 64-RH scale while GNN-augmented PPO remains feasible. Strengths include a clean MDP formulation with explicit constraints, Optuna-based hyperparameter tuning for all baselines, six independent training runs, and a clear internal comparison showing that removing the GNN component (MPPO) degrades reliability at scale. However, the contribution is currently overstated: the generalization experiment evaluates only topologies seen in training, the closest prior methods ([5], [6]) are not benchmarked, and 'perfect reliability' is measured per training run rather than as an episode-level success probability. These issues affect the paper's headline claims and need to be fixed before publication.

major comments (4)

[VI.B.6, VI.C.3, Table VI] The 'generalization' test does not evaluate unseen topologies. Section VI.B.6 states that the DRL model is trained on 5 different 8-RH topologies simultaneously and evaluated on the same 5 topologies. Section VI.C.3 then concludes that GPPO 'generalize[s] learned policies to unseen network structures.' Because every evaluation topology was used during training, the 19% cost reduction and 25% reward improvement in Table VI do not support transfer to unseen structures. Please either hold out topologies during training, resample new topologies from the same generator for evaluation, or remove the 'unseen/generalization' interpretation from the abstract and Section VI.C.3.
[I, II, VI.B.1, Table V] The paper identifies [5] and [6] as the closest prior works (Section I and Table I), but never benchmarks them. The baselines are generic DDPG, PPO, and masked PPO. Consequently, the abstract's claim that GPPO 'consistently outperforms state-of-the-art baselines' is unsupported. Add comparisons with [5] and [6] on the same simulator, or revise the claim to specify that the comparison is against generic RL baselines rather than state-of-the-art O-RAN orchestration methods.
[VI.C.4, Table V] The 'perfect reliability'/'perfect success rate' claim is based on 6/6 independent training runs, not on the fraction of feasible evaluation episodes. For the large-scale topology, MPPO's 3/6 is also a run-level statistic; the text later interprets this as a '50% success rate' suitable for production decisions. Please report episode-level success rates (e.g., over the 10 evaluation episodes per run) and define the reliability metric precisely, since the abstract's 'perfect reliability' is otherwise ambiguous.
[VI.A, VI.C.1] The large-scale result (18.2% cost reduction) is based on a single randomly generated 64-RH topology, uniform link delay/bandwidth samples, and hand-set cost weights (psi, phi_R, phi_L). While not an internal inconsistency, this limits the practical scalability claim. Please include a sensitivity analysis over multiple large-scale topologies and cost-coefficient values, or moderate the claim that GPPO is ready for practical O-RAN deployments.

minor comments (5)

[V.A.3, Eq. (6)] Equation (6) contains a duplicated resource-violation term: 'max(R_i - p_i, 0)' appears twice for i in D∪C. Also, p_i is not defined; it should be the total computing consumption on server i (analogous to p_d and p_c in constraints C5–C6).
[V.C.2] The text refers to 'Section 4.3' for the GNN blocks; this should be Section V.C.2. Also, 'GINEConv' is mentioned with reference [13], but the specific edge-attribute formulation in Eq. (12) should be linked precisely to that reference.
[V.B] The symbols e_rd and e_rc are used in the action-masking description but are not defined in Table III or the text. Please define the substrate connectivity matrix (or list) used for RH–ES and RH–RC adjacency.
[VI.A.6] The units of the cost coefficients are described as 'CCs per Mbps' for psi_DU_s and psi_CU_s, but Eq. (1) multiplies c_s (computing resource requirement) by psi_s. Please clarify the units so the cost computation is unambiguous.
[VI.C.3] Table VI reports 'Mean Cost' and 'Mean Reward' but not standard deviations as fractions; the text reports 19% and 25% improvements. It would be helpful to include the standard deviations in the percentages or in the table for reader interpretation.

Circularity Check

1 steps flagged · score 6.0 of 10

Generalization evidence is circular: train and test topologies coincide, so the 'unseen structures' claim is unsupported.

fitted input called prediction [Section VI.B.6 (Evaluation Protocol) and Section VI.C.3 (Generalization); Table VI]
"In the generalization setup, we trained a DRL model on 5 different 8-RH topologies simultaneously and evaluated them on the same 5 topologies for 100 evaluation episodes in total. ... This demonstrates GPPO’s superior ability to generalize learned policies to unseen network structures"

The evaluation set equals the training set by construction: the same 5 topologies are used for both fitting the policy and measuring 'generalization'. The reported 19% cost reduction and 25% reward improvement are therefore training-set performances, not predictions about unseen structures. The conclusion that GPPO generalizes to unseen network structures is thus not derived from held-out evidence; it reduces to an interpolation check on the very topologies the model was trained on. This is a fitted-input-called-prediction pattern: the policy is fitted to the 5 topologies, then the result on those same topologies is presented as evidence of transfer to unseen inputs.

full rationale

The core derivation chain is otherwise non-circular: the ILP formulation is converted into an MDP, the reward is constructed from the same cost components that the agent optimizes (standard RL design, not a tautological prediction), and the cost coefficients, slice distributions, and resource capacities are imported from external prior work rather than from the paper's own fitted results. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The one genuine circular element is the generalization evaluation: Section VI.B.6 trains and evaluates on the same five topologies, while Section VI.C.3 and the abstract claim this demonstrates transfer to unseen network structures. That makes the abstract's '25% higher reward in generalization tests' an artifact of the evaluation design rather than a supported prediction about new topologies. The remaining large-scale cost and reward comparisons, measured on the training topology, are still legitimate benchmark results, so the circularity is partial rather than total.

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

The central claim rests on hand-set simulation parameters, inherited cost and traffic distributions, and standard RL/GNN machinery. No new physical or conceptual entities are postulated. The main burden is that the objective, the reward, and the evaluation metrics all derive from the same hand-assembled cost model, so simulation-level comparisons are internally consistent but not externally calibrated.

free parameters (5)

Optuna-tuned PPO hyperparameters = lr=1e-4, batch=128, gamma=0.98, GAE lambda=0.97, clip=0.3, entropy=1e-6
Selected by automated search on the same tasks; not justified by analysis or ablations.
Network architecture sizes = MLP 256x256; GNN hidden 1024; 2 GINEConv layers
Chosen by hand; no ablation showing sensitivity to these widths or depths.
Cost weights = phi_R=1, phi_L=1
Hand-set; stated adjustable in practice; directly affects the cost and reward comparison.
Split processing cost coefficients = psi_DU={0.05,0.04,0.00325,0}, psi_CU={0,0.001,0.00175,0.05} CC/Mbps
Taken from [3],[4]; still arbitrary parameters that define the optimization objective.
Topology and resource configuration = 8/64 RH; 3/4 ES; 2 RC; ES=20, RC=100 CC; link bandwidth U(10,40) Gbps; delay U(0,3.6) ms; split-4 direct link prob 10%;
Synthetic problem instances; no real network data or operator traces.

assumptions (5)

domain assumption Synthetic topologies with guaranteed RH-ES-RC connectivity represent O-RAN substrates.
Experiments use randomly generated graphs (Figure 5); no real O-RAN deployment topology is used.
domain assumption Slice request distributions from [14] represent real SLA mixes.
eMBB, mMTC, and uRLLC parameters are sampled from uniform distributions inherited from [14].
domain assumption The linear cost model (processing, routing, reconfiguration) reflects operator deployment costs.
Section VI-A6 defines the costing system; no validation against actual operator cost data.
standard math PPO with invalid action masking produces a valid policy gradient.
Relied on in Section V-B and cited to [12].
domain assumption GNN mean pooling plus node-order feature preserves enough positional information for action indexing.
Permutation-invariant GNN needs node order to map actions to RH indices; the paper does not analyze how this affects generalization across different topologies.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Scalable O-RAN Resource Management: Graph-Augmented Proximal Policy Optimization." pith.science (2026). https://pith.science/paper/UGH3WGWJ

@misc{pith2026250910499,
  author       = {Pith},
  title        = {Pith review of: Towards Scalable O-RAN Resource Management: Graph-Augmented Proximal Policy Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGH3WGWJ}},
  note         = {Machine review of arXiv:2509.10499}
}

read the original abstract

Open Radio Access Network (O-RAN) architectures enable flexible, scalable, and cost-efficient mobile networks by disaggregating and virtualizing baseband functions. However, this flexibility introduces significant challenges for resource management, requiring joint optimization of functional split selection and virtualized unit placement under dynamic demands and complex topologies. Existing solutions often address these aspects separately or lack scalability in large and real-world scenarios. In this work, we propose a novel Graph-Augmented Proximal Policy Optimization (GPPO) framework that leverages Graph Neural Networks (GNNs) for topology-aware feature extraction and integrates action masking to efficiently navigate the combinatorial decision space. Our approach jointly optimizes functional split and placement decisions, capturing the full complexity of O-RAN resource allocation. Extensive experiments on both small-and large-scale O-RAN scenarios demonstrate that GPPO consistently outperforms state-of-the-art baselines, achieving up to 18% lower deployment cost and 25% higher reward in generalization tests, while maintaining perfect reliability. These results highlight the effectiveness and scalability of GPPO for practical O-RAN deployments.

Figures

Figures reproduced from arXiv: 2509.10499 by the authors.

**Figure 2.** Functional split configurations showing protocol layer distribution [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Workflow of the GNN-enhanced PPO agent for O-RAN placement. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Message passing and aggregation of the GINEConv block. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Substrate network topologies (a) Small-scale network with 8 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Performance comparison on the 8-RH topology: (a) deployment cost and (b) training reward. However, deployment costs are only reported for episodes that achieve feasible solutions, since infeasible deployments cannot be physically implemented and thus have no meaningful…

**Figure 8.** Figure 8: Reward evolution during training across methods and network scales. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages

[5]

Deep Reinforcement Learning for Orchestrating Cost-Aware Reconfigurations of vRANs,

F. W. Murti, S. Ali, G. Iosifidis, and M. Latva-aho, “Deep Reinforcement Learning for Orchestrating Cost-Aware Reconfigurations of vRANs,” IEEE Transactions on Network and Service Management , vol. 21, no. 1, pp. 200–216, Feb. 2024

work page 2024
[6]

NetMind: Adaptive RAN Baseband Function Placement by GCN Encoding and Maze-solving DRL,

H. Li, P. Li, K. D. Assis, A. Aijaz, S. Shen, R. Nejabati, S. Yan, and D. Simeonidou, “NetMind: Adaptive RAN Baseband Function Placement by GCN Encoding and Maze-solving DRL,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC) , Apr. 2024, pp. 1–6

work page 2024
[1]

Scalable and cost-efficient algorithms for baseband unit (bbu) function split placement,

N. Mharsi, M. Hadji, D. Niyato, W. Diego, and R. Krishnaswamy, “Scalable and cost-efficient algorithms for baseband unit (bbu) function split placement,” in 2018 IEEE Wireless Communications and Networking Conference (WCNC), 2018, pp. 1–6

work page 2018
[2]

Flexible functional split in 5G networks,

D. Harutyunyan and R. Riggio, “Flexible functional split in 5G networks,” in 2017 13th International Conference on Network and Service Management (CNSM), Oct. 2017, pp. 1–9

work page 2017
[3]

Constrained Deep Reinforcement Based Functional Split Optimization in Virtualized RANs,

F. W. Murti, S. Ali, and M. Latva-Aho, “Constrained Deep Reinforcement Based Functional Split Optimization in Virtualized RANs,” IEEE Transactions on Wireless Communications , vol. 21, no. 11, pp. 9850– 9864, Oct. 2022

work page 2022
[4]

Deep Reinforcement Learning for Robust VNF Reconfigurations in O-RAN,

E. Amiri, N. Wang, M. Shojafar, M. Q. Hamdan, C. H. Foh, and R. Tafa- zolli, “Deep Reinforcement Learning for Robust VNF Reconfigurations in O-RAN,” IEEE Transactions on Network and Service Management , vol. 21, no. 1, pp. 1115–1128, Feb. 2024

work page 2024
[7]

Proactive VNF Scaling and Placement in 5G O-RAN Using ML,

K. Ali and M. Jammal, “Proactive VNF Scaling and Placement in 5G O-RAN Using ML,” IEEE Transactions on Network and Service Management, vol. 21, no. 1, pp. 174–186, Feb. 2024

work page 2024
[8]

PlaceRAN: Optimal placement of virtualized network functions in Beyond 5G radio access networks,

F. Z. Morais, G. M. F. De Almeida, L. L. Pinto, K. Cardoso, L. M. Contreras, R. D. R. Righi, and C. B. Both, “PlaceRAN: Optimal placement of virtualized network functions in Beyond 5G radio access networks,” IEEE Transactions on Mobile Computing , pp. 1–1, 2022

work page 2022

Show all 19 references

[9]

O-RAN- Oriented Approach for Dynamic VNF Placement Focused on Interference Mitigation,

V . H. L. Lopes, G. M. Almeida, A. Klautau, and K. V . Cardoso, “O-RAN- Oriented Approach for Dynamic VNF Placement Focused on Interference Mitigation,” IEEE International Conference on Communications , 2024

2024
[10]

Proximal policy optimization algorithms,

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” CoRR, vol. abs/1707.06347, 2017

2017 arXiv
[11]

Trust region policy optimization,

J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, “Trust region policy optimization,” in International conference on machine learning. PMLR, 2015, pp. 1889–1897

2015
[12]

A closer look at invalid action masking in policy gradient algorithms,

S. Huang and S. Onta ˜n´on, “A closer look at invalid action masking in policy gradient algorithms,” arXiv preprint arXiv:2006.14171 , 2020

2006 arXiv
[13]

Strategies for pre-training graph neural networks,

W. Hu, B. Liu, J. Gomes, M. Zitnik, P. Liang, V . Pande, and J. Leskovec, “Strategies for pre-training graph neural networks,” in International Conference on Learning Representations (ICLR) , 2020

2020
[14]

Slicing Enabled Flexible Functional Split and Multi-Dimensional Resource Provisioning in 5G- and-Beyond RAN,

Y . Wu, L. Liang, Y . Jia, W. Wen, and Z. Chen, “Slicing Enabled Flexible Functional Split and Multi-Dimensional Resource Provisioning in 5G- and-Beyond RAN,” IEEE Transactions on Wireless Communications , vol. 23, no. 2, pp. 1213–1227, Feb. 2024

2024
[15]

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
[16]

Stable baselines3 contrib,

A. Raffin, A. Hill, M. Ernestus, A. Gleave, A. Kanervisto, and N. Dormann, “Stable baselines3 contrib,” https://github.com/ Stable-Baselines-Team/stable-baselines3-contrib, 2021

2021
[17]

Stable-baselines3: Reliable reinforcement learning imple- mentations,

A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-baselines3: Reliable reinforcement learning imple- mentations,” Journal of Machine Learning Research , vol. 22, no. 268, pp. 1–8, 2021

2021
[18]

Optuna: A next-generation hyperparameter optimization framework,

T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyperparameter optimization framework,” in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 2019, pp. 2623–2631

2019
[19]

Understanding the difficulty of training deep feedforward neural networks,

X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proceedings of the thirteenth international conference on artificial intelligence and statistics . JMLR Workshop and Conference Proceedings, 2010, pp. 249–256

2010

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [5]

Deep Reinforcement Learning for Orchestrating Cost-Aware Reconfigurations of vRANs,

F. W. Murti, S. Ali, G. Iosifidis, and M. Latva-aho, “Deep Reinforcement Learning for Orchestrating Cost-Aware Reconfigurations of vRANs,” IEEE Transactions on Network and Service Management , vol. 21, no. 1, pp. 200–216, Feb. 2024

work page 2024

[2] [6]

NetMind: Adaptive RAN Baseband Function Placement by GCN Encoding and Maze-solving DRL,

H. Li, P. Li, K. D. Assis, A. Aijaz, S. Shen, R. Nejabati, S. Yan, and D. Simeonidou, “NetMind: Adaptive RAN Baseband Function Placement by GCN Encoding and Maze-solving DRL,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC) , Apr. 2024, pp. 1–6

work page 2024

[3] [1]

Scalable and cost-efficient algorithms for baseband unit (bbu) function split placement,

N. Mharsi, M. Hadji, D. Niyato, W. Diego, and R. Krishnaswamy, “Scalable and cost-efficient algorithms for baseband unit (bbu) function split placement,” in 2018 IEEE Wireless Communications and Networking Conference (WCNC), 2018, pp. 1–6

work page 2018

[4] [2]

Flexible functional split in 5G networks,

D. Harutyunyan and R. Riggio, “Flexible functional split in 5G networks,” in 2017 13th International Conference on Network and Service Management (CNSM), Oct. 2017, pp. 1–9

work page 2017

[5] [3]

Constrained Deep Reinforcement Based Functional Split Optimization in Virtualized RANs,

F. W. Murti, S. Ali, and M. Latva-Aho, “Constrained Deep Reinforcement Based Functional Split Optimization in Virtualized RANs,” IEEE Transactions on Wireless Communications , vol. 21, no. 11, pp. 9850– 9864, Oct. 2022

work page 2022

[6] [4]

Deep Reinforcement Learning for Robust VNF Reconfigurations in O-RAN,

E. Amiri, N. Wang, M. Shojafar, M. Q. Hamdan, C. H. Foh, and R. Tafa- zolli, “Deep Reinforcement Learning for Robust VNF Reconfigurations in O-RAN,” IEEE Transactions on Network and Service Management , vol. 21, no. 1, pp. 1115–1128, Feb. 2024

work page 2024

[7] [7]

Proactive VNF Scaling and Placement in 5G O-RAN Using ML,

K. Ali and M. Jammal, “Proactive VNF Scaling and Placement in 5G O-RAN Using ML,” IEEE Transactions on Network and Service Management, vol. 21, no. 1, pp. 174–186, Feb. 2024

work page 2024

[8] [8]

PlaceRAN: Optimal placement of virtualized network functions in Beyond 5G radio access networks,

F. Z. Morais, G. M. F. De Almeida, L. L. Pinto, K. Cardoso, L. M. Contreras, R. D. R. Righi, and C. B. Both, “PlaceRAN: Optimal placement of virtualized network functions in Beyond 5G radio access networks,” IEEE Transactions on Mobile Computing , pp. 1–1, 2022

work page 2022