Pith. sign in

REVIEW 3 major objections 5 minor 16 references

Explainable AI for Radar Resource Management: Modified LIME in Deep Reinforcement Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By replacing LIME's independent perturbations with a deep network that generates correlated tracking costs from target positions, the paper obtains a more faithful and more useful explanation of a deep-reinforcement-learning radar…

desk verdict A plausible LIME variant for correlated radar states, with a solid MAE result but a mislabeled utility metric and missing ablations. read the letter →

arxiv 2506.20916 v1 pith:LABIKLZD submitted 2025-06-26 cs.LG

Ziyang Lu , M. Cenk Gursoy , Chilukuri K. Mohan , Pramod K. Varshney This is my paper · ORCID
classification cs.LG
keywords explainableAILIMEdeepreinforcementlearningradarresourcemanagementdwelltimeallocationfeaturecorrelationDDPGcognitive
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 sets out to show that LIME's local explanations of a deep-reinforcement-learning radar resource manager become both more faithful and more useful when the perturbed states it samples respect the correlations among the radar state features. It proposes DL-LIME, a modification that perturbs only the position and dual-variable components of a state and uses a deep network, pre-trained on states the trained agent actually experienced, to generate the corresponding tracking costs. In the authors' numerical comparison, DL-LIME achieves lower average MAE (1.95 versus 2.27) and higher average utility (4.49e4 versus 4.01e4) than conventional LIME, at the cost of higher runtime. The explanations it produces are interpretable: each target's previous tracking cost is the dominant factor in that target's dwell-time decision, with other targets' costs also contributing.

What carries the argument

The central mechanism is a DNN-based correlated sampler inserted into LIME's perturbation step. Conventional LIME draws each component of a perturbed state independently from a normal distribution around the instance being explained, which can produce physically impossible radar states because tracking costs and target positions are strongly correlated (a far target has higher tracking cost). DL-LIME instead perturbs only the $2N+1$ non-cost state components and feeds them through a pre-trained feedforward DNN that predicts the corresponding tracking costs, so the local neighborhood explored by LIME respects the learned conditional relationships. The weighted linear surrogate model is then fitted on these realistic perturbed states, and its coefficients serve as feature-importance scores.

What would settle it

Compute the true EKF tracking costs for the same perturbed positions DL-LIME feeds its DNN and compare the resulting MAE with DL-LIME's, and also run an ablation that perturbs all $3N+1$ state components while using the DNN only to keep costs consistent with positions. If DL-LIME loses its advantage in these conditions, the claim that correlation modeling is the cause of the improvement is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that replacing LIME's independent feature perturbations with deep-network-generated correlated samples changes both the quality and the content of local explanations. The DNN takes the $2N+1$ state components that exclude tracking costs — target positions and the dual variable — and outputs the $N$ tracking costs, so each perturbed state honors the learned physics of the radar problem rather than an independence assumption. The numerical evidence (Table II) shows DL-LIME with lower MAE and higher utility than LIME, and the explanation plots (Fig. 3) show that for each target $n$, the previous tracking cost $c^n_{t-1}$ dominates the decision for Action $n$, while the other targets' costs also carry weight. The authors interpret this as the DDPG agent balancing targets jointly rather than allocating dwell time to each target in isolation.

Load-bearing premise

The load-bearing premise is that the feedforward DNN, trained on the agent's experienced states, predicts tracking costs that match the true conditional distribution for every perturbed state LIME generates; if the DNN generalizes poorly to local neighborhoods or rare states, DL-LIME's perturbed samples are no more realistic than LIME's independent samples, and the reported gains could be an artifact of perturbing fewer dimensions.

Editorial extensions

If this is right

  • In the tested radar environment, DL-LIME explains the DDPG agent's dwell-time decisions with lower average MAE (1.95 versus 2.27) than conventional LIME.
  • Executing DL-LIME's explained actions in the environment yields higher average utility (4.49e4 versus 4.01e4), so the explanations carry real task value.
  • The runtime cost is higher (1.70 versus 0.42 seconds per decision), but the paper shows it can be traded against fidelity by adjusting the number of perturbed samples.
  • The explanations show that each target's previous tracking cost dominates its own dwell-time decision, with the other targets' costs also contributing, indicating the agent allocates resources jointly.

Reading between the lines

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

  • A direct test of the mechanism would replace the DNN's predicted costs with the exact EKF tracking costs for the same perturbed positions; if the MAE advantage persists, the benefit is correlation, not just the DNN.
  • An ablation that perturbs all 16 state components while using the DNN only to restore cost-position consistency would separate correlation preservation from dimensionality reduction.
  • The same correlated-sampling recipe could be applied to SHAP or other attribution methods whose independent sampling assumption distorts correlated features.
  • The dominant role of previous tracking cost suggests a simple resource-allocation heuristic — allocate dwell time roughly in proportion to each target's last tracking cost — that could be benchmarked against the DDPG agent's utility.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes DL-LIME, a modification of LIME for explaining a DDPG-based radar resource manager. Conventional LIME perturbs all state components independently; DL-LIME instead perturbs only the position, velocity, and dual-variable components and uses a feedforward DNN, pretrained on experienced states, to generate the corresponding tracking-cost components, thereby preserving physical correlations during sampling. The authors evaluate fidelity (MAE), task performance (utility), runtime, and peak performance period, and report that DL-LIME outperforms conventional LIME on fidelity and utility while being slower. They also present qualitative explanations suggesting that a target's previous tracking cost dominates its dwell-time decision.

Significance. If the reported gains are reproducible, the paper makes a useful contribution by adapting LIME's sampling distribution to physically correlated radar state features without abandoning model-agnostic local surrogates. The qualitative findings in Section V-D are plausible and would be of interest to the cognitive-radar and XAI communities. The paper does not release code, error bars, or multiple-seed statistics, so the numerical claims are not independently verifiable at this stage; nevertheless, the core idea of replacing independent perturbations of correlated features with DNN-generated values is worth pursuing.

major comments (3)
  1. [Section V-B, Table II] Table II reports an average utility of 4.49e4 for DL-LIME, which exceeds the DDPG agent's 4.39e4 even though DDPG was trained to maximize the same utility via the Lagrangian reward in Eq. (9). This is unexpected for a local linear surrogate and requires a concrete explanation. Please specify exactly how the utility is computed: whether the LIME actions are executed as one-step replacements at the same checkpoints, whether target trajectories are held fixed across methods, and whether the DDPG utility is evaluated from the same policy that generated the explanations. If the comparison is one-step open-loop, the 'task performance' claim should be reworded; if it is closed-loop, a rollout protocol with identical random seeds and state updates must be described.
  2. [Section V-B] No error bars, multiple seeds, or significance tests are reported for the MAE (1.95 vs. 2.27) or utility (4.49e4 vs. 4.01e4) comparisons. Because DRL training and the target-spawning model in Section V-A2 are stochastic, single-run comparisons are insufficient to support the paper's headline superiority claim. At minimum, the authors should report means and standard deviations over several independent seeds and, if possible, paired per-slot comparisons.
  3. [Section IV-B.2, Section V-B] The fidelity advantage is attributed to the DNN generating tracking costs that preserve correlations, but the DNN's prediction accuracy against true EKF costs is never reported. Without validating the DNN on perturbed states, the lower MAE could simply result from perturbing fewer dimensions (2N+1 instead of 3N+1) rather than from capturing correlations. Please report the DNN's prediction error on held-out states, include an ablation that perturbs only the 2N+1 non-cost components with independent sampling, and, if feasible, compare against a covariance-based Gaussian perturbation of the full state.
minor comments (5)
  1. [Section V-B] The utility metric is defined with an unclosed parenthesis in 'Ut({τ n t }N n=1'; also, the description of executing LIME actions 'in the environment' should state explicitly whether the environment state is advanced between actions or whether a single time step is used.
  2. [Section V-B] The peak performance period values sum to 100% (48.57 + 11.20 + 40.23), implying exactly one method is best in every slot; if ties are possible, the metric definition should state how ties are resolved.
  3. [Section V-B] The runtime comparison in Table II is hard to interpret because the number of LIME perturbation samples for conventional LIME is not stated; the text mentions 10,000 samples only for DL-LIME. Please specify the sample count for both methods so the runtime and MAE comparison is apples-to-apples.
  4. [Section IV-A] Equation (12) takes the square root of an exponential kernel, which is equivalent to redefining the kernel width; this is unnecessarily confusing. Consider writing the kernel directly as exp(-D^2/(2a^2)) or using a consistent definition.
  5. [References] Reference [16] is cited as an arXiv preprint; if a peer-reviewed version exists, it should be cited instead. Also, the abbreviation 'Pf' and 'Pd' in Table I are not defined in the table caption; they should be introduced in the text or table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the fidelity comparison is grounded in the DDPG policy's actual actions, and self-citations only provide background system-model context.

full rationale

The paper's central comparison is the fidelity of DL-LIME and LIME as local surrogates for a trained DDPG radar resource manager. That comparison is not circular: the LIME linear model is fitted to the DDPG agent's actual action outputs via the weighted least-squares loss in Eq. (13), and the reported MAE in Eq. (15) directly measures the difference between the surrogate's actions and the agent's actions. The proposed DL-LIME modification replaces independent perturbation of tracking-cost features with DNN-generated costs, where the DNN is a supervised regressor trained on a collected dataset D of experienced states; this is a modeling choice whose accuracy is not validated in the paper, but it does not make the subsequent LIME fit an identity with its inputs. The self-citations to [9] and [10] are used only to set up the radar system model and the constrained DRL formulation, not to justify the explanation results, so they are not load-bearing in a circular sense. The surprising Table II result that DL-LIME's average utility exceeds the DDPG agent's utility (4.49e4 vs 4.39e4) is an internal consistency or evaluation-protocol concern, not a circularity: it does not show that any predicted quantity is equivalent by construction to its own input. Likewise, the DNN could be inaccurate or the perturbation space could be artificially narrowed, but these are correctness and generalization risks, not instances of fitting a parameter and then renaming it a prediction. No quoted equation reduces to another equation by construction, and no central claim depends on an unverified self-citation chain. The appropriate finding is therefore no significant circularity.

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

The central claim rests on method hyperparameters and on a DNN that is assumed to generate realistic correlated cost features. No new physical entities are introduced. The main unvalidated input is the DNN's ability to act as a conditional generator for tracking costs.

free parameters (5)
  • Perturbation variance sigma_m^2 = Estimated from dataset D
    Per-component variance used for LIME perturbation; estimated from states experienced by the same DRL agent, so the local neighborhoods reflect the agent's operating distribution.
  • LIME kernel width a = 2.5
    Controls the exponential similarity kernel in Eq. (12); chosen by hand, no sensitivity analysis reported.
  • Ridge regularization coefficient c = 1e-3
    Regularization in the LIME linear model, Eq. (13); chosen by hand.
  • Number of perturbation samples K = 10000
    Number of samples per explanation; Fig. 4 shows a runtime/MAE tradeoff, but the main results use a single value.
  • DNN hidden layer sizes and learning rate = 128/64 neurons, lr=1e-4
    Architecture and learning rate for the cost-prediction network; no ablation or accuracy reporting.
assumptions (4)
  • domain assumption Extended Kalman filter and radar measurement model (Eqs. (1) to (3)) accurately describe target motion and tracking error.
    Background from authors' prior work [9],[10]; used without independent validation here.
  • domain assumption The CMDP formulation with utility Ut = -sum costs + beta*Gamma and dual-variable update (Eqs. (7) to (10)) is a valid model for radar resource management.
    The DRL problem definition is inherited from previous papers by the same authors; the paper does not re-derive it.
  • ad hoc to paper A feedforward DNN with inputs consisting of target positions and the dual variable can generate tracking costs that preserve the true correlations among state components.
    This is the load-bearing premise of DL-LIME; the paper provides only a scatter plot (Fig. 1) as motivation and no accuracy metrics for the DNN.
  • domain assumption The dataset D collected by deploying the well-trained agent is representative of the states encountered during the explanation evaluation.
    Both the perturbation statistics and the DNN training rely on D; if D omits rare but important states, explanations may be biased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable AI for Radar Resource Management: Modified LIME in Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/LABIKLZD

@misc{pith2026250620916,
  author       = {Pith},
  title        = {Pith review of: Explainable AI for Radar Resource Management: Modified LIME in Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LABIKLZD}},
  note         = {Machine review of arXiv:2506.20916}
}
read the original abstract

Deep reinforcement learning has been extensively studied in decision-making processes and has demonstrated superior performance over conventional approaches in various fields, including radar resource management (RRM). However, a notable limitation of neural networks is their ``black box" nature and recent research work has increasingly focused on explainable AI (XAI) techniques to describe the rationale behind neural network decisions. One promising XAI method is local interpretable model-agnostic explanations (LIME). However, the sampling process in LIME ignores the correlations between features. In this paper, we propose a modified LIME approach that integrates deep learning (DL) into the sampling process, which we refer to as DL-LIME. We employ DL-LIME within deep reinforcement learning for radar resource management. Numerical results show that DL-LIME outperforms conventional LIME in terms of both fidelity and task performance, demonstrating superior performance with both metrics. DL-LIME also provides insights on which factors are more important in decision making for radar resource management.

Figures

Figures reproduced from arXiv: 2506.20916 by the authors.

Figure 1
Figure 1. Tracking Costs and Target-to-radar Distances [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of DDPG and DL-LIME Time Allocation Strategies [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Component importance analysis of DL-LIME explanations at different time points. (a)-(b) show explanations for the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Tradeoff between MAE and Runtime for DL-LIME [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 10 canonical work pages

  1. [1]

    Continuous control with deep reinforcement learning,

    T. Lillicrap, “Continuous control with deep reinforcement learning,” arXiv preprint arXiv:1509.02971 , 2015

    arXiv 2015

  2. [2]

    Mastering the game of go without human knowledge,

    D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton et al. , “Mastering the game of go without human knowledge,” nature, vol. 550, no. 7676, pp. 354–359, 2017

    2017

  3. [3]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” in International conference on machine learning . PMLR, 2018, pp. 1861–1870

    2018

  4. [4]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  5. [5]

    Deep reinforcement learning control for radar detection and tracking in congested spectral environments,

    C. E. Thornton, M. A. Kozy, R. M. Buehrer, A. F. Martone, and K. D. Sherbondy, “Deep reinforcement learning control for radar detection and tracking in congested spectral environments,” IEEE Transactions on Cognitive Communications and Networking , vol. 6, no. 4, pp. 1335– 1349, 2020

    work page 2020

  6. [6]

    Scene-adaptive radar tracking with deep reinforcement learning,

    M. Stephan, L. Servadei, J. Arjona-Medina, A. Santra, R. Wille, and G. Fischer, “Scene-adaptive radar tracking with deep reinforcement learning,” Machine Learning with Applications , vol. 8, p. 100284, 2022

    work page 2022

  7. [7]

    Quality of service based radar resource management using deep reinforcement learning,

    S. Durst and S. Br ¨uggenwirth, “Quality of service based radar resource management using deep reinforcement learning,” in 2021 IEEE Radar Conference (RadarConf21). IEEE, 2021, pp. 1–6

    work page 2021

  8. [8]

    Time budget management in multifunction radars using reinforcement learning,

    P. Pulkkinen, T. Aittom ¨aki, A. Str ¨om, and V . Koivunen, “Time budget management in multifunction radars using reinforcement learning,” in 2021 IEEE Radar Conference (RadarConf21) , 2021, pp. 1–6

    work page 2021

Show all 16 references
  1. [9]

    Resource allocation for multi-target radar tracking via constrained deep reinforcement learning,

    Z. Lu and M. C. Gursoy, “Resource allocation for multi-target radar tracking via constrained deep reinforcement learning,” IEEE Transac- tions on Cognitive Communications and Networking , vol. 9, no. 6, pp. 1677–1690, 2023

    2023

  2. [10]

    Learning- based cognitive radar resource management for scanning and multi- target tracking,

    Z. Lu, M. C. Gursoy, C. K. Mohan, and P. K. Varshney, “Learning- based cognitive radar resource management for scanning and multi- target tracking,” in ICC 2024 - IEEE International Conference on Communications, 2024, pp. 2785–2790

    2024

  3. [11]

    Explainable artificial intelligence (xai): Concepts, taxonomies, opportu- nities and challenges toward responsible ai,

    A. B. Arrieta, N. D ´ıaz-Rodr´ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garc ´ıa, S. Gil-L ´opez, D. Molina, R. Benjamins et al. , “Explainable artificial intelligence (xai): Concepts, taxonomies, opportu- nities and challenges toward responsible ai,” Informatio...

    2020

  4. [12]

    Explainable machine learning in deployment,

    U. Bhatt, A. Xiang, S. Sharma, A. Weller, A. Taly, Y . Jia, J. Ghosh, R. Puri, J. M. Moura, and P. Eckersley, “Explainable machine learning in deployment,” in Proceedings of the 2020 conference on fairness, accountability, and transparency , 2020, pp. 648–657

    2020

  5. [13]

    ” why should i trust you?

    M. T. Ribeiro, S. Singh, and C. Guestrin, “” why should i trust you?” explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , 2016, pp. 1135–1144

    2016

  6. [14]

    A unified approach to interpreting model predictions,

    S. Lundberg, “A unified approach to interpreting model predictions,” arXiv preprint arXiv:1705.07874 , 2017

    2017 arXiv

  7. [15]

    Christoph, Interpretable machine learning: A guide for making black box models explainable

    M. Christoph, Interpretable machine learning: A guide for making black box models explainable . Leanpub, 2020

    2020

  8. [16]

    A modified perturbed sampling method for local interpretable model-agnostic explanation,

    S. Shi, X. Zhang, and W. Fan, “A modified perturbed sampling method for local interpretable model-agnostic explanation,” arXiv preprint arXiv:2002.07434, 2020

    2002 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.