Self-Explaining Reinforcement Learning for Mobile Network Resource Allocation

Ahmad Terra; Franco Ruggeri; Konrad Nowosadko

REVIEW 4 major objections 5 minor 15 references

Self-Explaining Reinforcement Learning for Mobile Network Resource Allocation

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A self-explaining neural network can serve as the policy of a PPO agent, matching deep-net performance while providing feature-level explanations for every action in a mobile network resource allocation task.

desk verdict Useful, honest application of SENN-as-policy to a small wireless allocation problem; the abstract oversells the explanation correlation, but the architecture and experiments are clear enough to warrant a real review. read the letter →

arxiv 2509.14925 v2 pith:R6QNAMUA submitted 2025-09-18 cs.LG cs.NI

Konrad Nowosadko , Franco Ruggeri , Ahmad Terra This is my paper · ORCID

classification cs.LGcs.NI

keywords self-explainingneuralnetworksreinforcementlearninginterpretabilityPPOmobilenetworkresourceallocationfeatureattributionglobalexplanationslocal

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 claims that a reinforcement learning agent can be made natively interpretable by replacing the policy network of a PPO agent with a Self-Explaining Neural Network (SENN), while still performing within a small margin of a black-box deep network and clearly better than the best deployed heuristic on a mobile network resource allocation task. The SENN actor produces intrinsic local explanations—feature relevance scores and effect values—for every action, and the paper introduces two ways to aggregate these into global explanations: effect distributions and K-means clustering with centroid importances. A trainable bias term is added to the SENN aggregator, which improves accuracy and exposes the model's systematic preferences. The authors argue that these intrinsic explanations are more trustworthy than post-hoc feature-attribution methods, and they report that clustering-based attributions agree in sign with DeepLift and InputXGradient on most features, though effect distributions often disagree. The significance for a general reader is that RL policies in high-stakes domains could become auditable without sacrificing much performance.

What carries the argument

The load-bearing component is the Self-Explaining Neural Network used as the PPO policy actor, with two modifications: an identity conceptizer (no latent feature extraction, so concepts equal raw input features) and a trainable bias in the aggregator. The policy output is f(x)=θ(x)ᵀx+b, where θ(x) is the parametrizer's relevance-score matrix. The robustness loss λ∥∇x f(x)−θ(x)ᵀJ_x^h(x)∥ enforces local linearity and stabilizes explanations, and the paper uses the Lipschitz estimate as a stability metric. Global explanations come from aggregating local explanations via effect distributions and K-means clustering, the latter producing action-specific centroids and importance vectors I(a). This

What would settle it

On a synthetic environment with a known linear optimal policy, train the same SENN-PPO pipeline and compute the Pearson correlation between the extracted global feature attributions and the known policy coefficients; if the correlation is not high (e.g., below 0.9), the faithfulness claim is refuted. Alternatively, on the mobile-network task, recompute the attribution agreement across all ten seeds and report the distribution; if the average sign-agreement with DeepLift/InputXGradient is not significantly above chance, the abstract's correlation claim collapses.

Watch

Extended reading notes

Core claim

SENNs can serve as the actor in PPO. Using an identity conceptizer and a bias-augmented aggregator, the policy output is θ(x)ᵀx + b, so each action's logit decomposes into directly inspectable input-feature contributions. Local explanations are the relevance scores θ(x) and the effect scores θ(x)⊙x. Global explanations are obtained by averaging effect-value distributions or by K-means clustering the relevance scores and computing per-action centroid importances. In the resource allocation task, the biased SENN matches a deep PPO agent almost exactly and outperforms the heuristic, while the clustering-based global attributions mostly align in sign with DeepLift and InputXGradient. The paper t

Load-bearing premise

The load-bearing assumption is that agreement with post-hoc attribution methods (DeepLift and InputXGradient) is a valid measure of explanation correctness—an assumption the paper's own results weaken, since effect distributions often disagree with those methods and no ground-truth attribution exists.

Editorial extensions

If this is right

If the reported performance is representative, SENN-parametrized policies can be deployed in place of black-box DRL actors in low-dimensional settings without a big reward penalty.
Operators can audit individual network actions by reading the relevance and effect scores for each input feature, without running a separate explanation model.
The proposed global-explanation techniques give a model-intrinsic alternative to post-hoc attribution, with clustering-based attributions closer to DeepLift/InputXGradient than effect distributions are.
The robustness-loss coefficient λ provides a single tuning knob that trades predictive return against local stability of explanations.
Visualizing the bias vector reveals the model's inherent preference among actions, complementing feature-level attributions.

Reading between the lines

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

A stricter test would compare SENN attributions against known ground-truth feature effects in a synthetic environment; the paper's reliance on post-hoc methods as a proxy leaves that open.
Because the identity conceptizer requires low-dimensional inputs, extending this approach to high-dimensional states (images, raw radio signals) would need a conceptizer whose concepts remain human-meaningful—something the paper does not address.
The discrepancy between effect distributions and post-hoc methods suggests the 'correlate strongly' claim may be limited to clustering-based attributions; a reader should not extrapolate it to all intrinsic explanations.
The method's global explanations are computed over inference-time transitions, so they describe the trained policy's behavior in the simulation distribution, not necessarily its behavior in deployment if the environment shifts.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a clean, honest application paper—SENN as the policy network in PPO for a small mobile network resource allocation problem—with two modest architectural tweaks and two global-explanation aggregation schemes. The main thing to know: the performance claim is plausible but statistically unsupported, and the abstract’s “correlate strongly” overstates what the body actually shows.

What is actually new: an identity conceptizer (sensible for low-dimensional states), a trainable bias in the aggregator, and the two aggregation methods (effect distributions and k-means-based centroids). None of these is a new learning paradigm, but the combination is not in the prior work they cite, and the paper describes the architecture and experimental setup clearly enough to reimplement. The honest parts are welcome: Section VI-D openly shows the λ trade-off and the outlier at λ=0.001, and the conclusion says the global explanation methods “performed reasonably well but did not achieve particularly strong results.” That matches the data better than the abstract does.

Soft spots, in order of size. First, the headline correlation claim. Section VI-C gives no correlation coefficient, no significance test, and explicitly says ED “positively attributes each feature, disagreeing with post-hoc methods.” The clustering agreement is described only qualitatively. So the abstract’s “correlate strongly” is unsupported. Second, the performance comparison in Figure 2 has no error bars, confidence intervals, or significance tests, so “significantly outperforms” is not established. Third, using agreement with DeepLift/InputXGradient as a correctness yardstick is a weak validation: those post-hoc methods are not ground truth, and the paper itself notes the scope difference. That does not sink the paper—the intrinsic explanations stand on their own—but it means the evaluation does not prove correctness. Minor: no code released, and the clustering hyperparameters (τ, k) are set with limited sensitivity analysis.

Overall, this is a legitimate incremental contribution. It would be useful to readers working on XRL in networking who want a concrete SENN-PPO recipe. It deserves a serious referee, but the abstract needs to be aligned with the evidence and the statistics need to be added. I’d engage with it.

Referee Report

4 major / 5 minor

Summary. The paper introduces a self-explaining reinforcement learning framework in which the actor of a PPO agent is a Self-Explaining Neural Network (SENN) with an identity conceptizer and an additive bias. It derives local explanations (relevance and effect scores) and two global-explanation aggregation procedures: effect distributions and K-means clustering of relevance vectors. The method is evaluated on the mobile-env resource allocation problem, comparing episodic reward against a dense-network PPO baseline and a heuristic, and comparing global attributions with GradSHAP, InputXGradient, and DeepLift. The paper claims competitive performance and strong correlation of the extracted global explanations with post-hoc methods, but the reported support is largely qualitative.

Significance. This is a potentially useful contribution to Explainable RL: replacing the PPO actor with a SENN and aggregating its local relevance vectors into global explanations is a natural and reasonably novel idea, and the mobile-env testbed is appropriate for a low-dimensional demonstration. The identity conceptizer and bias term are sensible modifications, and the comparison against multiple post-hoc explanation methods is a good instinct. However, the paper does not currently supply the quantitative evidence (statistical performance tests, agreement metrics, sensitivity analysis) needed to substantiate its headline claims, so the significance of the demonstrated results is uncertain.

major comments (4)

[Abstract; §VI-A; Fig. 2] The abstract claims the approach 'significantly outperforms the best deployed heuristic' and performs within a small margin of the state-of-the-art DRL method. However, Fig. 2 is a plot of aggregated returns with no error bars, confidence intervals, or statistical tests, and §VI-A only says 'all agents outperformed the best heuristic' and 'just a bit worse' than the PPO baseline. Across 10 seeds, a paired test such as Wilcoxon should be reported before using 'significantly' in the abstract.
[Abstract; §VI-C; Fig. 6] The abstract asserts that 'the extracted global explanations correlate strongly with DeepLift and InputXGradient.' The body provides no correlation coefficient, sign-agreement rate, or any numerical agreement metric. In §VI-C the authors state that K-means 'agrees with other methods on the character of the contribution of most of the features' but with 'discrepancies in the contributions’ values,' and that ED 'positively attributes each feature, disagreeing with post-hoc methods.' This is at most qualitative sign-level agreement, not a demonstrated strong correlation. Quantify the agreement (e.g., Spearman ρ, sign match, normalized error) or soften the claim.
[§VI-C] The evaluation treats agreement with post-hoc attribution methods as validation that the global explanations are correct, but this yardstick is not justified. The paper itself notes that post-hoc methods use an input-output perspective while the proposed methods use a relevance-score-output perspective, so disagreement is expected under the stated design. Without an external ground truth or a clearly defined faithfulness property, the comparison in Fig. 6 is descriptive, not a correctness validation. Please state the validation criterion explicitly and frame the comparison accordingly.
[§IV-C/§V-B] The global-explanation results depend on several unstated or arbitrary choices: the λ value used for the model whose explanations are shown is not reported; k=14 is chosen in §V-B; and τ=0.6 in Eq. (6) is fixed without analysis. Since Fig. 7 shows that λ materially changes policy behavior and stability, and since the clustering results are central to the explanation claims, the paper should specify the configuration and provide a sensitivity analysis over τ, k, and the selected λ.

minor comments (5)

[Throughout] Typographical errors: 'Artificial Intelligance' (I), 'etimate' (VI-A), 'Casuality' (V-C), 'accurracy' (III).
[References] References [9] and [14] are duplicates (same Schneider et al. paper); merge and cite once.
[Abstract] The sentence 'Our approach strong performance on par...' is ungrammatical, and the level of performance is stated more cautiously in the body; make abstract consistent.
[Eq. (6); §V-B] Equation (6) overloads n for number of input features and number of actions; use distinct symbols. Also clarify whether K-means is applied to relevance scores or to input features; §V-B says 'relevance scores' but §V-C averages 'centroid values' as feature attributions.
[Fig. 6] Figure 6 has no numerical axis labels or table of attribution values; the qualitative comparison is hard to verify. Consider adding a table with mean attributions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: global explanations are aggregated from the model's own relevance scores and compared against external post-hoc benchmarks; there are no self-cited load-bearing results and no fitted-quantity-renamed-as-prediction.

full rationale

The paper's central derivations are self-contained. The SENN architecture and loss are taken from external prior work [3], not from the authors' own publications. The modifications (identity conceptizer, trainable bias) are architectural choices that do not define the explanation target in terms of itself. Local explanations are the relevance scores θ(x) and effects θ(x)⊙x produced by the trained SENN; global explanations are arithmetic aggregations of these local values (means for ED, centroids for clustering). No parameter is fitted to the post-hoc attribution values and then reported as a prediction: the post-hoc methods (DeepLift, InputXGradient, GradSHAP) are used only as a comparative benchmark in Section VI-C, and the paper explicitly reports that ED disagrees with them and that clustering only matches 'on the character of the contribution of most of the features.' The abstract's phrasing 'correlate strongly' is not supported by the quantitative evidence in the body, and the use of post-hoc agreement as a ground-truth proxy is an unvalidated evaluation assumption. However, these are correctness/evidence concerns, not circularity: the claimed explanation outputs are not defined in terms of the benchmark, and no derivation step reduces to its own input. There is also no self-citation chain, no uniqueness theorem imported from the authors, and no ansatz smuggled in via self-citation. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard algorithmic components (PPO, SENN, K-means) plus three hand-tuned meta-parameters (lambda, tau, k) and two domain assumptions about the simulator. No new particles, forces, or constructs are introduced.

free parameters (3)

lambda (robustness loss scale) = 0.0, 0.001, 0.1 (and others)
Controls the strength of the local stability regularizer and the accuracy-stability trade-off; chosen by hand and varied across runs (Section VI-D).
tau (cluster purity threshold) = 0.6
Hand-set threshold in Equation (6) that determines whether a cluster is assigned to an action; affects which centroids are used for global explanations.
k (number of K-means clusters) = 14
Chosen via silhouette score and Davies-Bouldin index to balance cluster purity and interpretability (Section V-B, VI-C).

assumptions (4)

standard math PPO provides stable policy updates for this environment
Assumed as a black-box training algorithm; cited from Schulman et al. [10] and used throughout.
standard math The SENN robustness loss enforces local linearity of the parametrizer
Assumed from Alvarez Melis and Jaakkola [3]; the paper relies on this to justify the explanations.
domain assumption Okumura-Hata path loss model computes SNR in mobile-env
The environment's SNR observations are computed with this model, cited from [13]; the paper does not question it.
domain assumption Mobile-env with 3 UEs and 3 BSs is a representative resource allocation problem
The paper evaluates only this small instance and generalizes conclusions from it; see Section V-A.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Self-Explaining Reinforcement Learning for Mobile Network Resource Allocation." pith.science (2026). https://pith.science/paper/R6QNAMUA

@misc{pith2026250914925,
  author       = {Pith},
  title        = {Pith review of: Self-Explaining Reinforcement Learning for Mobile Network Resource Allocation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6QNAMUA}},
  note         = {Machine review of arXiv:2509.14925}
}

read the original abstract

Deep reinforcement learning (DRL) methods, though powerful, often lack transparency, which limits their adoption in critical domains. We apply Self-Explaining Neural Networks (SENNs) to RL by parametrizing the policy of a PPO agent with a SENN, producing intrinsic local explanations, and propose a method for aggregating them into global explanations. We evaluate our approach on a mobile network resource allocation problem, our approach performs within a small margin of the state-of-the-art deep learning method and significantly outperforms the best deployed heuristic, while the extracted global explanations correlate strongly with DeepLift and InputXGradient, making SENNs a promising candidate for high-stakes RL.

Figures

Figures reproduced from arXiv: 2509.14925 by the authors.

**Figure 3.** Observation values (purple) and corresponding rele [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Distribution of effect values, the red dot marks the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: Comparing attributions of each input feature toward [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Relation between, different λ values, Lipschitz constant and performance for the resource allocation problem in mobile networks. higher λ values to underperform in terms of episodic return can be observed. Showing that the regularization effect introduced with robustne…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 2 canonical work pages

[1]

Explainable reinforcement learning: A survey and comparative review,

S. Milani, N. Topin, M. Veloso, and F. Fang, “Explainable reinforcement learning: A survey and comparative review,”ACM Comput. Surv., vol. 56, no. 7, Apr. 2024. [Online]. Available: https://doi.org/10.1145/3616864

doi:10.1145/3616864 2024
[2]

Conceptual challenges for interpretable machine learning,

D. S. Watson, “Conceptual challenges for interpretable machine learning,”Synthese, vol. 200, 2022. [Online]. Available: https://doi.org/10.1007/s11229-022-03485-5

work page doi:10.1007/s11229-022-03485-5 2022
[3]

Towards robust interpretability with self-explaining neural networks,

D. Alvarez Melis and T. Jaakkola, “Towards robust interpretability with self-explaining neural networks,” inAdvances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018

2018
[4]

C-senn: Contrastive self-explaining neural network,

Y . Sawada and K. Nakamura, “C-senn: Contrastive self-explaining neural network,” 2022. [Online]. Available: https://arxiv.org/abs/2206.09575

arXiv 2022
[5]

Q-senn: quantized self-explaining neural networks,

T. Norrenbrock, M. Rudolph, and B. Rosenhahn, “Q-senn: quantized self-explaining neural networks,” inProceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence, ser. AAAI’24/IAAI’24/...

work page doi:10.1609/aaai.v38i19.30145 2024
[6]

A unified approach to interpreting model predictions,

S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., vol. 30. Curran Associates, Inc., 2017

2017
[7]

Learning important features through propagating activation differences,

A. Shrikumar, P. Greenside, and A. Kundaje, “Learning important features through propagating activation differences,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. PMLR, 06–11 Aug 2017, pp. 3145–3153. [Online]. Available: https://proceedings.m...

2017
[8]

Notions of explainability and evaluation approaches for explainable artificial intelligence,

G. Vilone and L. Longo, “Notions of explainability and evaluation approaches for explainable artificial intelligence,” Information Fusion, vol. 76, pp. 89–106, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1566253521001093

2021

Show all 15 references

[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. [Online]. Available: http://dblp.uni- trier.de/db/journals/corr/corr1707.htmlSchulmanWDRK17

2017 arXiv
[11]

Molnar,Interpretable Machine Learning, 3rd ed., 2025

C. Molnar,Interpretable Machine Learning, 3rd ed., 2025. [Online]. Available: https://christophm.github.io/interpretable-ml-book

2025
[12]

Some methods for classification and analysis of multi- variate observations,

J. MacQueen, “Some methods for classification and analysis of multi- variate observations,” inProceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics, vol. 5. University of California press, 1967, pp. 281–298

1967
[13]

mobile- env: An open platform for reinforcement learning in wireless mobile networks,

S. Schneider, S. Werner, R. Khalili, A. Hecker, and H. Karl, “mobile- env: An open platform for reinforcement learning in wireless mobile networks,” inNOMS 2022-2022 IEEE/IFIP Network Operations and Management Symposium, 2022, pp. 1–3

2022
[14]

Multi-agent deep reinforcement learning for coordinated multipoint in mobile networks,

S. Schneider, H. Karl, R. Khalili, and A. Hecker, “Multi-agent deep reinforcement learning for coordinated multipoint in mobile networks,” IEEE Transactions on Network and Service Management (TNSM), 2023

2023
[15]

Not just a black box: Learning important features through propagating activation differences,

A. Shrikumar, P. Greenside, A. Shcherbina, and A. Kundaje, “Not just a black box: Learning important features through propagating activation differences,” 2017. [Online]. Available: https://arxiv.org/abs/1605.01713

2017 arXiv
[16]

Axiomatic attribution for deep networks,

M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” inProceedings of the 34th International Conference on Machine Learning - Volume 70, ser. ICML’17. JMLR.org, 2017, p. 3319–3328

2017

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Explainable reinforcement learning: A survey and comparative review,

S. Milani, N. Topin, M. Veloso, and F. Fang, “Explainable reinforcement learning: A survey and comparative review,”ACM Comput. Surv., vol. 56, no. 7, Apr. 2024. [Online]. Available: https://doi.org/10.1145/3616864

doi:10.1145/3616864 2024

[2] [2]

Conceptual challenges for interpretable machine learning,

D. S. Watson, “Conceptual challenges for interpretable machine learning,”Synthese, vol. 200, 2022. [Online]. Available: https://doi.org/10.1007/s11229-022-03485-5

work page doi:10.1007/s11229-022-03485-5 2022

[3] [3]

Towards robust interpretability with self-explaining neural networks,

D. Alvarez Melis and T. Jaakkola, “Towards robust interpretability with self-explaining neural networks,” inAdvances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018

2018

[4] [4]

C-senn: Contrastive self-explaining neural network,

Y . Sawada and K. Nakamura, “C-senn: Contrastive self-explaining neural network,” 2022. [Online]. Available: https://arxiv.org/abs/2206.09575

arXiv 2022

[5] [5]

Q-senn: quantized self-explaining neural networks,

T. Norrenbrock, M. Rudolph, and B. Rosenhahn, “Q-senn: quantized self-explaining neural networks,” inProceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence, ser. AAAI’24/IAAI’24/...

work page doi:10.1609/aaai.v38i19.30145 2024

[6] [6]

A unified approach to interpreting model predictions,

S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., vol. 30. Curran Associates, Inc., 2017

2017

[7] [7]

Learning important features through propagating activation differences,

A. Shrikumar, P. Greenside, and A. Kundaje, “Learning important features through propagating activation differences,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. PMLR, 06–11 Aug 2017, pp. 3145–3153. [Online]. Available: https://proceedings.m...

2017

[8] [8]

Notions of explainability and evaluation approaches for explainable artificial intelligence,

G. Vilone and L. Longo, “Notions of explainability and evaluation approaches for explainable artificial intelligence,” Information Fusion, vol. 76, pp. 89–106, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1566253521001093

2021