Pith. sign in

REVIEW 4 major objections 5 minor 17 references

PPO-EPO: Energy and Performance Optimization for O-RAN Using Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A PPO reinforcement-learning agent decides which O-RAN cells to switch off, and on TeraVM Viavi RIC tester data it improves both energy efficiency and downlink throughput compared with SARSA and random cell selection.

desk verdict A reasonable O-RAN cell-shutdown RL application undermined by a self-referential validation: every reported gain is recomputed by the same analytic model used to train the policy. read the letter →

arxiv 2504.14749 v1 pith:VVTTBZMU submitted 2025-04-20 cs.NI

Rawlings Ntassah , Gian Michele Dell'Aera , Fabrizio Granelli This is my paper · ORCID
classification cs.NI
keywords O-RANreinforcementlearningPPOenergyefficiencycellshutdowntrafficsteeringTeraVMViaviRICnetworkoptimization
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 a reinforcement-learning controller for Open RAN that learns when to shut down underutilized cells to save radio-access energy without violating throughput, resource-utilization, and interference constraints. The agent uses PPO (Proximal Policy Optimization) to pick the cell to deactivate, then redistributes affected users to neighboring cells through a handover rule. On the TeraVM Viavi RIC tester dataset, the trained policy yields higher downlink throughput and better energy efficiency than SARSA and random selection. The sympathetic reading is that learned cell-shutdown policies are a viable tool for O-RAN energy saving, provided the underlying power and throughput models are dependable.

What carries the argument

The core mechanism is a Proximal Policy Optimization agent operating on a per-cell state composed of throughput, PRB utilization, interference, and UE distribution. Its action is the choice of a cell to shut down, and the reward combines a performance-gain term and a power-gain term subject to three constraints: average throughput must not fall below a fraction $\delta$ of its prior value, PRB utilization must not increase, and interference must stay below a threshold. The handover machinery uses 3GPP A3 events and reassigns users to neighboring cells via a multinomial distribution weighted by PRB availability and inverse distance. The power model $P_k = P_{\mathrm{idle}} + \aleph_{\mathrm{DL},k}P_{\mathrm{tx}}/\eta$ is what converts a shutdown decision into an energy-efficiency figure.

What would settle it

Measure the actual power draw of the radio units during a PPO-selected shutdown in the TeraVM setup (or comparable O-RAN testbed) and compare it with the value predicted by eq. (7); if realized energy saving is materially smaller, or realized throughput differs from eq. (2), the claimed energy-efficiency gains are not supported. Alternatively, replay the same dataset through a system-level simulator with full link adaptation and check whether PPO still beats SARSA and Random Selection in the downlink-throughput CDF.

Watch

Extended reading notes

Core claim

The central claim is that a PPO-trained policy can solve the O-RAN cell-shutdown optimization problem: given the current distribution of UEs, PRB usage, interference, and throughput per cell, the agent decides which cell to switch off to maximize a weighted objective of performance gain and power gain while respecting constraints. The authors model per-cell power as $P_k = P_{\mathrm{idle}} + \aleph_{\mathrm{DL},k}P_{\mathrm{tx}}/\eta$, and energy efficiency as average throughput divided by that power. Throughput is computed from Shannon capacity with PRB allocation, and SINR from a 3GPP urban-macro pathloss model. The validation on the TeraVM Viavi RIC tester data shows PPO outperforming SARSA and Random Selection in the energy-gain ratio and in the CDF of downlink throughput. The authors interpret this as evidence that a PPO-based traffic-steering and cell-shutdown policy can deliver both energy savings and throughput gains in O-RAN when the objective weights are balanced ($\omega_{\mathrm{perf}}=0.4$, $\omega_{\mathrm{power}}=0.6$).

Load-bearing premise

The load-bearing premise is that the Section III equations (pathloss, SINR, PRB allocation, and the linear power model $P_k = P_{\mathrm{idle}} + \aleph_{\mathrm{DL},k} P_{\mathrm{tx}} / \eta$) faithfully predict real energy and throughput, yet the paper computes these from formulas rather than measuring actual power draw or user throughput after a shutdown.

Editorial extensions

If this is right

  • If validated in hardware, a PPO-trained policy could let mobile operators automate energy-saving cell shutdowns within the O-RAN RIC framework, cutting operational costs without manual reconfiguration.
  • The same RL approach could be extended to more granular shutdown decisions, such as switching off individual radio units or carriers, because the action space and constraints are modular.
  • The comparison against SARSA suggests that off-policy or hybrid RL methods could offer a stability-adaptability trade-off, which the authors themselves propose as future work on hybrid and hierarchical RL.
  • The objective weights ($\omega_{\mathrm{perf}}=0.4$, $\omega_{\mathrm{power}}=0.6$) directly encode the operator's preference for energy saving over throughput, so tuning them yields a family of policies spanning the energy-performance trade-off.

Reading between the lines

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

  • The validation is entirely simulation-based: energy and throughput are computed from the same equations the controller uses, so the reported gains do not yet establish real-world performance; a hardware-in-the-loop test measuring actual power draw is the natural next step.
  • Because the state includes interference, the learned policy may be sensitive to the spatial layout of cells, and the Turin-inspired 12-RU topology in a 400 m by 400 m area may not generalize to irregular deployments without retraining.
  • The CDF analysis shows PPO maintains higher throughput for a larger fraction of users, which suggests the model implicitly relieves congestion; one could test this directly by comparing PRB utilization distributions before and after shutdown decisions.
  • The reward formulation assumes Shannon-capacity throughput, which ignores finite modulation-and-coding tables and scheduling granularity; replaying the policy against a system-level simulator with realistic link adaptation would test robustness.
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

4 major / 5 minor

Summary. The paper proposes PPO-EPO, a PPO-based reinforcement learning agent for O-RAN traffic steering and cell shutdown. The system model includes pathloss, SINR, PRB allocation, throughput, interference, and a linear cell power model, leading to an optimization objective that trades performance gain against power gain under throughput, PRB, and interference constraints (Eq. 13). The agent is trained over five million episodes and evaluated on a dataset generated with the TeraVM Viavi RIC tester, comparing PPO against SARSA and Random Selection. The paper claims that PPO significantly improves network energy efficiency and downlink throughput.

Significance. If the claimed results were properly supported, the paper would be a useful contribution to O-RAN energy management, showing a constrained PPO policy for cell shutdown with a real RIC tester dataset and baseline comparisons. The system model is simple and mostly consistent, and the choice of PPO with explicit performance/power trade-off weights is reasonable. However, the current evaluation does not substantiate the central claim: the reward function is never specified, the constraints of Eq. (13) are not shown to be enforced, and all reported gains are computed from the same closed-form equations used to build the environment and the objective. The use of TeraVM data is a strength, but no actual measured power draw or user throughput after a shutdown decision is reported. The significance is therefore conditional on a substantial revision of the evaluation methodology.

major comments (4)
  1. [Section IV.A] The reward function is never specified. The text states that after taking an action 'the agent calculates the corresponding rewards based on the outcomes of its actions,' but no equation, pseudo-code, or verbal description of the reward is given. Without the reward, the PPO implementation is not reproducible, and it is impossible to determine whether the reported performance and power gains are actually aligned with the training objective. This is load-bearing because every claim of PPO superiority depends on the trained policy.
  2. [Section IV.B and Eqs. (9)-(13)] The validation is self-referential. The paper computes throughput gain, power gain, and energy efficiency using the same analytic model (pathloss, SINR, PRB allocation, and the linear power model of Eq. (7)) that defines the environment and the optimization objective of Eq. (13). Figures 3-6 report these model-computed quantities, not measurements obtained from the TeraVM Viavi RIC tester. No actual power draw, SINR, or user throughput measurements after a shutdown decision are reported. The abstract's claim that the method 'significantly improves' energy efficiency and downlink throughput is therefore not supported by independent ground truth. The authors should either report real measurements from the tester after applying the policy or explicitly frame the results as simulations and validate the underlying model against measurements.
  3. [Eq. (13) and Section IV.B] The constraints of the optimization problem are not shown to be enforced. The problem statement lists a throughput degradation constraint (R_avg,after >= delta * R_avg,before), a PRB utilization constraint, and an interference threshold constraint, but the paper never states the values of delta or I_threshold, nor does it explain how these constraints are incorporated into the PPO training (e.g., penalty terms, action masking, or constrained policy optimization). The observation in Section IV.B that negative throughput gains occur 'aligns with the allowable throughput degradation constraints' is not sufficient evidence, because the constraint might have delta = 0 or could be violated. Without this information, the claim that the policy 'effectively prevents unnecessary performance deterioration' is unsubstantiated.
  4. [Eq. (9) and Eq. (13)] The notation for the performance gain is inconsistent and affects the objective. Eq. (9) defines g_k as a per-cell gain and then defines G_perf as the average of g_k over K cells, but Eq. (13) uses G_perf,k as if it were a per-cell quantity. Additionally, after shutting down a cell, the averaging in G_perf should presumably be over K-1 cells, but the text does not clarify this. The parameter alpha in Eq. (9) is also never defined or assigned a value. These ambiguities make the exact objective being optimized unclear and should be corrected.
minor comments (5)
  1. [Section III, Eq. (5)] The interference weighting factor alpha in Eq. (5) is described as 'a weighting factor based on spatial proximity' but its value or how it is computed is not given in Table I or the text.
  2. [Section III, handover model] The handover redistribution model uses a multinomial distribution with weights xi = PRB_available * lambda_i, but the epsilon in the interference weight lambda_i = 1/(||nu_i - nu_k||^2 + epsilon) is not specified, and the line 'ni ~ Multinomial(,xi)' contains a typo.
  3. [Section IV.A and Fig. 2] The training curve in Fig. 2 has no axis labels or numerical scale, and the text does not state the convergence criterion or the number of training steps per episode. This makes it difficult to assess the claim that the policy converged to an optimal strategy.
  4. [Section IV.B and Figs. 4-6] The figures are described qualitatively (e.g., 'PPO consistently outperforms'), but no numerical values, confidence intervals, or statistical significance tests are reported. Adding quantitative results, such as mean and standard deviation of energy efficiency and throughput across episodes or seeds, would strengthen the comparison.
  5. [Table I] The table lists network and training parameters but omits several quantities used in the model, including the noise power p_noise, the PRB bandwidth B_prb, the power amplifier efficiency eta, the idle power P_idle, the maximum transmit power P_tx, the throughput demand R_demand_u distribution, and the constraint parameters delta and I_threshold.

Circularity Check

2 steps flagged · score 6.0 of 10

Validation reports the same analytic objective the agent optimizes, so the claimed energy/throughput improvements are not independently established.

  1. fitted input called prediction [Section IV.B 'Model Validation', second paragraph; Eqs. (9), (12), (13)]
    "Following the predicted shutdown actions, we analyzed the relationship between throughput gain and power gain, both of which are key parameters in our optimization objective function defined in 13."

    The paper's validation metrics are the objective terms themselves. Eq. 13 maximizes Σ x_k(ω_perf·G_perf,k + ω_power·P_gain,k), where G_perf,k (Eq. 9) is normalized average throughput gain and P_gain,k (Eq. 12) is normalized power gain. Both are computed from the model's own equations: throughput from Eq. (2) and power from Eq. (7). The TeraVM dataset is only used as input features for the trained policy; no measured power draw or user throughput is reported. The reported 'improvement' is therefore an evaluation of the trained objective on held-out inputs, not an independent measurement. The reward used in Sec. IV.A is not specified, so the reader cannot rule out that the agent was trained to maximize exactly the same quantity used as the validation outcome.

  2. self definitional [Section III 'Problem definition', Eqs. (7)-(8), (11), (13); Section IV.B, Fig. 4 caption]
    "The energy efficiency (EEk) of the cell k is denoted by EEk = R̄k / Pk ... The objective function is therefore defined as max xk Σ xk (ωperf·Gperf, k + ωpower·Pgain, k), (13)"

    Energy efficiency, the headline result, is defined by Eq. (8) as the model throughput (Eq. 2) divided by the closed-form power model (Eq. 7). Figure 4's caption calls the reported result 'The energy gain (The ratio of the rate and the power for the three methods)', which is the same R/P ratio. Thus the central claim that PPO 'significantly improves the network's energy efficiency and downlink throughput' is, by construction, a statement about the analytic quantity the optimization problem maximizes; it is not a measurement of real network energy or throughput after a shutdown decision.

full rationale

The derivation chain is internally consistent but self-referential at the validation step. The environment, the objective function (Eq. 13), and the validation figures all use the same closed-form throughput equation (Eq. 2), SINR/interference equations (Eqs. 4-5), and power model (Eq. 7). The training section never states the reward function, while the validation section explicitly says the reported throughput and power gains are 'key parameters in our optimization objective function defined in 13.' Consequently, the claimed 'experimental validation using TeraVM Viavi RIC tester data' is not an external check: the TeraVM data supplies features, but the outcome metrics are recomputed from the same model the agent was trained against. This warrants a partial circularity score of 6. The comparison between PPO, SARSA, and Random Selection still has some independent algorithmic content—the baselines are not literally forced to lose—so the paper is not a 8 or 10. There is no load-bearing self-citation or imported uniqueness theorem; the circularity is entirely in the evaluation loop, not in the related-work citations.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the unvalidated system model (eqs. 2, 7, 13) and on a proprietary dataset processed through the same pipeline used in training. The reward function is not stated, so the listed free parameters and domain assumptions are the load-bearing choices.

free parameters (6)
  • omega_perf and omega_power = omega_perf=0.4, omega_power=0.6
    Weight assignments in the objective function (13) that trade throughput against power. Chosen by hand in Section IV.B, not tuned or justified, and they drive the policy behavior.
  • delta (throughput degradation constraint) = not specified
    Constraint delta in eq. (13) sets the allowable average throughput drop after shutdown. The paper never gives its value or how it is enforced in training, yet it defines what 'acceptable' means.
  • alpha in eq. (9) for gk = not specified
    A smoothing constant in the denominator of the per-cell performance gain gk; no value or role is given.
  • epsilon in interference weight lambda_i = not specified
    Constant in lambda_i = 1/(||nu_i - nu_k||2 + epsilon) used in the handover multinomial weights; no value given.
  • alpha in interference eq. (5) = not specified
    Weighting factor based on spatial proximity in the interference calculation; undefined.
  • R_demand,u distribution = min 0.01 Gb, max 0.1 Gb
    Throughput demand per UE is drawn from an unstated distribution between these bounds, determining PRB demands in eq. (3).
assumptions (5)
  • domain assumption 3GPP TR 38.901 urban macro pathloss model governs received power.
    Used in eq. (1) and cited as [15]; standard propagation model but still a model assumption.
  • domain assumption UE throughput follows the Shannon formula Ru = Bu log2(1+zeta_u) with PRB allocation.
    Eq. (2) idealizes link behavior; no coding overhead, MCS limits, or scheduling granularity beyond integer PRBs.
  • ad hoc to paper Cell power consumption is Pk = Pidle + PRB_total * Ptx / eta.
    Eq. (7) is a simple linear model with no empirical validation; Pidle and eta are not given values.
  • ad hoc to paper Handover redistribution follows a multinomial distribution with weights xi = PRB_available * lambda_i.
    This is an invented assignment rule unstated in 3GPP and not validated against real handover behavior.
  • domain assumption TeraVM Viavi RIC tester dataset is a representative source of O-RAN network states.
    The dataset is proprietary and described only as 'generated using the TeraVM Viavi RIC Tester'; its coverage, size, and relationship to real traffic are not disclosed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PPO-EPO: Energy and Performance Optimization for O-RAN Using Reinforcement Learning." pith.science (2026). https://pith.science/paper/VVTTBZMU

@misc{pith2026250414749,
  author       = {Pith},
  title        = {Pith review of: PPO-EPO: Energy and Performance Optimization for O-RAN Using Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VVTTBZMU}},
  note         = {Machine review of arXiv:2504.14749}
}
read the original abstract

Energy consumption in mobile communication networks has become a significant challenge due to its direct impact on Capital Expenditure (CAPEX) and Operational Expenditure (OPEX). The introduction of Open RAN (O-RAN) enables telecommunication providers to leverage network intelligence to optimize energy efficiency while maintaining Quality of Service (QoS). One promising approach involves traffic-aware cell shutdown strategies, where underutilized cells are selectively deactivated without compromising overall network performance. However, achieving this balance requires precise traffic steering mechanisms that account for throughput performance, power efficiency, and network interference constraints. This work proposes a reinforcement learning (RL) model based on the Proximal Policy Optimization (PPO) algorithm to optimize traffic steering and energy efficiency. The objective is to maximize energy efficiency and performance gains while strategically shutting down underutilized cells. The proposed RL model learns adaptive policies to make optimal shutdown decisions by considering throughput degradation constraints, interference thresholds, and PRB utilization balance. Experimental validation using TeraVM Viavi RIC tester data demonstrates that our method significantly improves the network's energy efficiency and downlink throughput.

Figures

Figures reproduced from arXiv: 2504.14749 by the authors.

Figure 1
Figure 1. Network system design. The O-RUs are positioned at a fixed distance from each other. We define an O-RU as k ∈ K where K is the total number of O-RUs. The network contains a fixed number of user equipment (UEs), represented as U, which are randomly distributed across the cells. Each UE is admitted to a cell based on the highest reference signal power (pk(u)), determined using the pathloss equation (P Lk) specified by… view at source ↗
Figure 2
Figure 2. We trained our model over 5 million episodes, and the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Distribution of both power and throughput gain showing the heatmap [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: CDF of the throughput of the network for the different methods [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 4
Figure 4. Figure 4: The energy gain (The ratio of the rate and the power for the three [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages

  1. [1]

    Energy efficiency enhancements in radio access networks,

    T. Edler and S. Lundberg, “Energy efficiency enhancements in radio access networks,” Ericsson review, vol. 81, no. 1, pp. 42–51, 2004

    work page 2004

  2. [2]

    Energy-and spectral-efficiency trade-off in ofdma-based cooperative cognitive radio networks,

    X. Chen and S. Ouyang, “Energy-and spectral-efficiency trade-off in ofdma-based cooperative cognitive radio networks,” International Jour- nal of Distributed Sensor Networks , vol. 10, no. 2, p. 782648, 2014

    work page 2014

  3. [3]

    A survey of energy-efficient wireless communications,

    D. Feng, C. Jiang, G. Lim, L. J. Cimini, G. Feng, and G. Y . Li, “A survey of energy-efficient wireless communications,” IEEE Communications Surveys & Tutorials, vol. 15, no. 1, pp. 167–178, 2013

    work page 2013

  4. [4]

    Holistic approach for future energy effi- cient cellular networks,

    E. C. Strinati and L. H ´erault, “Holistic approach for future energy effi- cient cellular networks,” e & i Elektrotechnik und Informationstechnik , vol. 11, no. 127, pp. 314–320, 2010

    work page 2010

  5. [5]

    Towards energy efficient ran: From industry standards to trending practice,

    L. Kundu, X. Lin, and R. Gadiyar, “Towards energy efficient ran: From industry standards to trending practice,” arXiv preprint arXiv:2402.11993, 2024

    arXiv 2024

  6. [6]

    Energy saving in 6g o-ran using dqn-based xapp,

    Q. Wang, S. Chetty, A. Al-Tahmeesschi, X. Liang, Y . Chu, and H. Ah- madi, “Energy saving in 6g o-ran using dqn-based xapp,” arXiv preprint arXiv:2409.15098, 2024

    arXiv 2024

  7. [7]

    Achieving energy efficiency in open radio access networks (oran) using xapps,

    S. S. Malik, S. Sil, K. Ranjan, M. S. Duhan, and A. Singh, “Achieving energy efficiency in open radio access networks (oran) using xapps,” Quantum Scientific & Academic Research Consortium , vol. 1, no. 1, 2024

    2024

  8. [8]

    Energy efficiency in wireless: Ran and beyond,

    M. Paolini and S. Fili, “Energy efficiency in wireless: Ran and beyond,” 2023

    work page 2023

Show all 17 references
  1. [9]

    A survey of recent advances in optimization methods for wireless communications,

    Y .-F. Liu, T.-H. Chang, M. Hong, Z. Wu, A. M.-C. So, E. A. Jorswieck, and W. Yu, “A survey of recent advances in optimization methods for wireless communications,” IEEE Journal on Selected Areas in Communications, 2024

    2024

  2. [10]

    J. A. Sheikh, T. Khan, and B. K. Kanaujia, Intelligent Signal Processing and RF Energy Harvesting for State of Art 5G and B5G Networks . Springer Nature, 2024

    2024

  3. [11]

    Cell-free massive mimo in o-ran: Energy-aware joint orchestration of cloud, fronthaul, and radio resources,

    ¨O. T. Demir, M. Masoudi, E. Bj ¨ornson, and C. Cavdar, “Cell-free massive mimo in o-ran: Energy-aware joint orchestration of cloud, fronthaul, and radio resources,” IEEE Journal on Selected Areas in Communications, 2024

    2024

  4. [12]

    Design and evaluation of deep rein- forcement learning for energy saving in open ran,

    M. Bordin, A. Lacava, M. Polese, S. Satish, M. A. Nittoor, R. Sivaraj, F. Cuomo, and T. Melodia, “Design and evaluation of deep rein- forcement learning for energy saving in open ran,” arXiv preprint arXiv:2410.14021, 2024

    2024 arXiv

  5. [13]

    Enhancing energy efficiency in o-ran through intelligent xapps deployment,

    X. Liang, A. Al-Tahmeesschi, Q. Wang, S. Chetty, C. Sun, and H. Ah- madi, “Enhancing energy efficiency in o-ran through intelligent xapps deployment,” arXiv preprint arXiv:2405.10116 , 2024

    2024 arXiv

  6. [14]

    Landscape-enabled al- gorithmic design for the cell switch-off problem in 5g ultra-dense networks,

    J. Galeano-Brajones, F. Luna, J. Carmona-Murillo, A. J. Nebro, C. A. Coello Coello, and J. F. Valenzuela-Vald ´es, “Landscape-enabled al- gorithmic design for the cell switch-off problem in 5g ultra-dense networks,” Engineering Optimization, pp. 1–23, 2024

    2024

  7. [15]

    5g; study on channel model for frequencies from 0.5 to 100 ghz (3gpp tr 38.901 version 15.0.0 release 15),

    3GPP, “5g; study on channel model for frequencies from 0.5 to 100 ghz (3gpp tr 38.901 version 15.0.0 release 15),” 2018

    2018

  8. [16]

    3rd generation partnership project (3gpp), “nr; radio resource control (rrc) protocol specification (release 18),

    ——, “3rd generation partnership project (3gpp), “nr; radio resource control (rrc) protocol specification (release 18),” (technical specification group radio access network, tech),” 2023

    2023

  9. [17]

    Stable-baselines3: Reliable reinforcement learning implementa- tions,

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

    2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.