An AI-driven framework for rapid and localized optimizations of urban open spaces

Arman Nikkhah Dehnavi; Maedeh Mirdamadi; Pegah Eshraghi; Riccardo Talami; Zahra-Sadat Zomorodian

REVIEW 3 major objections 6 minor 2 references

An AI-driven framework for rapid and localized optimizations of urban open spaces

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Park redesign answers arrive in 1 minute, not 30

desk verdict A legitimate applied extension of SHAP+CFX to SVF/visibility optimization, with honest simulation validation, but the headline speed claim rests on an unfair GA comparison that needs fixing. read the letter →

arxiv 2501.08019 v1 pith:GHXA5IAQ submitted 2025-01-14 cs.LG cs.AIcs.CY

Pegah Eshraghi , Arman Nikkhah Dehnavi , Maedeh Mirdamadi , Riccardo Talami , Zahra-Sadat Zomorodian This is my paper · ORCID

classification cs.LGcs.AIcs.CY

keywords urbanopenspacesSkyViewFactorvisibilityanalysiscounterfactualexplanationsXGBoostSHAPmorphologygeneticalgorithmoptimization

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 claims that urban open spaces can be optimized through small, targeted changes to surrounding buildings rather than full redesigns. It focuses on two morphology-driven metrics: sky view factor, the fraction of open sky seen from a park, and visibility, how well a park is seen from surrounding buildings. The authors train machine-learning models on 1,152 simulated Tehran-style blocks, find XGBoost most accurate, and then use counterfactual explanations to propose minimal edits such as lowering an eastern building two stories or shifting a southern neighbor farther away. Grasshopper re-simulation of the recommendations gives mean errors between 4.12% and 6.06%, and the full optimization runs in about a minute versus 15–30 minutes for a genetic algorithm. If true, planners gain a fast, interpretable way to retrofit existing open spaces without expensive global search.

What carries the argument

The engine is the counterfactual explanation, a small set of feature changes that flips a model's prediction to a target outcome. The implementation uses a KD-tree, a binary space-partitioning structure, to organize the parameter space of building heights and distances and quickly find the nearest feasible configuration that meets the goal, prioritizing features identified by SHAP and respecting fixed constraints such as orientation and street width. This mechanism replaces the iterative evaluate-and-mutate loop of a genetic algorithm with a single nearest-neighbor-style query on the trained XGBoost surrogate, which is what produces the speed advantage.

What would settle it

Take real parks with surrounding building data, apply the CFX-recommended height and distance changes, and compare predicted SVF and visibility against fisheye photographs or high-fidelity re-simulation; if mean RMSE leaves the reported 4–6% band, or if a genetic algorithm beats CFX on solution quality when both are given the same ten-minute budget, the central claim would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that counterfactual explanation can serve as an optimizer for localized urban morphology. Using XGBoost as a surrogate for Ladybug/Grasshopper simulation, SHAP values to prioritize features, and a KD-tree search to locate the nearest feasible design change, the framework generates minimal adjustments that raise predicted SVF by up to 13 percentage points or lift a park's visibility class by one level. The paper reports that these recommendations are accurate: Grasshopper re-simulation of ten test configurations yields mean RMSE between 4.12% and 6.06%, while the CFX run takes about one minute compared with 15–30 minutes for a genetic algorithm. The claim is that explanation machinery, not global search, is sufficient for practically useful local improvements, with the extra benefit that each suggestion is directly actionable and interpretable.

Load-bearing premise

The load-bearing premise is that 1,152 randomly generated regular-grid blocks built from Tehran-style rules—no vegetation, fixed floor-area ratios, straight streets, and 3-to-10-story buildings—represent real urban open spaces well enough that recommendations trained on them remain valid in other contexts.

Editorial extensions

If this is right

A designer can test five alternative localized retrofit strategies per park in about a minute before committing to detailed simulation.
Feature-importance rankings translate directly into design heuristics: park area and east/west building heights dominate SVF, while distance to southern buildings and building width dominate visibility.
Because SVF and visibility depend only on morphology rather than climate, the same trained pipeline can be reapplied to other cities without re-tuning to local weather, provided block geometry stays within the training envelope.
The 15–30x speed advantage over genetic algorithms makes block-by-block, city-scale optimization feasible on ordinary hardware.
The validated RMSE band of roughly 4–6% supports using CFX suggestions as a screening step, with high-fidelity simulation reserved for the most invasive recommended changes.

Reading between the lines

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

Not directly claimed by the paper: the speed of CFX points to interactive design tools where dragging a building edge updates predicted sky view and visibility instantly, with KD-tree queries replacing repeated simulations.
Not directly claimed by the paper: the same surrogate-plus-counterfactual recipe should extend to other morphology-sensitive metrics such as daylight autonomy, wind comfort, or noise whenever enough simulated samples exist to train an accurate predictor.
Not directly claimed by the paper: because RMSE grows when CFX recommends large height or distance changes, adding uncertainty bands to counterfactual suggestions would shore up the method exactly where it is weakest.
Not directly claimed by the paper: a head-to-head test on real urban blocks with vegetation, irregular parcels, and setbacks would reveal whether the SHAP-dominant features remain dominant outside regular-grid synthetic configurations.

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.

Desk Editor's Note

read the letter

The genuinely new thing here is the packaging: SHAP for feature importance plus counterfactual explanations (CFX) for minimal-change design suggestions, applied to SVF and visibility in urban open spaces, with the CFX outputs checked against Grasshopper simulations. That validation step is real work, and the mean RMSE of 4.12-6.06% across ten configurations is credible evidence that the counterfactuals are not just model hallucinations. The surrogate models are also evaluated on a held-out split, with XGBoost getting R2=0.91 for SVF and F1 up to 0.89 for visibility. No circularity in the target-result sense: the ML model is fit to simulation data, and the CFX suggestions are validated by independent simulations. The authors deserve credit for that. The soft spot is the one the stress-test flagged, and it is load-bearing. The paper's central claim is that CFX gets the same results as GA in about one minute versus 15-30 minutes. But the GA time is wall-clock simulation-based optimization, while the CFX time excludes the 1,152 Grasshopper runs used to build the training set and the XGBoost training and tuning. That is not apples-to-apples. Worse, Section 4.4 never reports the actual SVF or visibility values the GA achieved - it only says 'closest results' were reached in the 3rd or 4th generation. So 'comparable convergence' is asserted, not demonstrated. If the GA actually reaches substantially better values, the speed advantage is irrelevant. This is a fixable but serious flaw: the authors need to report final objective values for both methods, and either include full pipeline time for CFX or clearly scope the claim to inference-time only. Other concerns are minor by comparison. The synthetic scenarios are simplified Tehran-style blocks with no vegetation, fixed FAR ranges, and regular grids, so the generalizability claim in the title is overstated. Lack of code, data, and hyperparameter details would also need addressing. The SHAP interpretation sections read a bit long, but they are not wrong. Overall, this is a coherent applied study, not a breakthrough. The core idea - using counterfactuals for localized design retrofits - builds directly on Regenwetter et al. and existing ML-surrogate workflows. But the demonstration is honest and the validation is a plus. With a proper GA comparison and full timing accounting, the paper could be solid. As is, the speed claim is unsubstantiated, but the underlying methodology deserves a serious referee who can push for that revision.

Referee Report

3 major / 6 minor

Summary. This paper presents a workflow for suggesting minimal, localized changes to urban open-space layouts to improve sky view factor (SVF) and visibility. The authors generate 1,152 synthetic urban block configurations based on simplified Tehran morphology, simulate SVF and visibility with Ladybug/Grasshopper, train five surrogate machine-learning models, select XGBoost as the most accurate, analyze feature importance with SHAP, and generate counterfactual explanations (CFX) via a KD-tree search. The CFX suggestions are re-simulated in Grasshopper and achieve mean RMSE between 4.12% and 6.06%. The paper claims that the CFX approach is 15–30 times faster than a genetic algorithm benchmark while producing comparable results.

Significance. If the speed and accuracy claims hold, the framework would be a useful interpretable surrogate-based design-support tool for early-stage urban retrofitting. The study includes a genuine held-out model comparison (XGBoost SVF R²=0.91; visibility F1 up to 0.89) and a validation of counterfactual suggestions against simulation, which is a good internal consistency check. The integration of SHAP and CFX on a simulation-based surrogate is a meaningful contribution in this applied context. However, the paper's central efficiency claim is not yet supported by the evidence, as discussed in the major comments. No code or dataset is provided, so reproducibility could not be verified.

major comments (3)

[§4.4] The headline speed comparison is not like-for-like. The reported GA time of 15–30 minutes is wall-clock simulation time for a population of 50 over 3–4 generations, whereas the reported CFX time of about 1 minute covers only the counterfactual retrieval step on an already-trained surrogate; the 1,152 Grasshopper simulations used to generate the training set (§3.1) and the XGBoost training and hyperparameter selection (§3.2) are excluded. In addition, the text states that GA reached 'closest results' but never reports the actual SVF or visibility values achieved by GA on the same scenarios, so the claim of 'similar outcomes' is not demonstrated. To support the speed advantage, the comparison should include the full pipeline cost for both methods and report the achieved objective values of the GA solutions.
[§3.4 and Figure 5] The validation of CFX suggestions by Grasshopper simulations is not an independent check on accuracy because the same Ladybug/Grasshopper pipeline generated the training labels in §3.1; it only confirms that the surrogate does not grossly diverge from the simulation engine. The abstract and conclusion claim suitability for 'real-world' urban planning, but no comparison against measured SVF/visibility data is provided. Please add a discussion of this limitation or include field or independent simulation data.
[§3.1 and §5] The claims of generalizability across diverse urban contexts are not supported by the experimental design. The 1,152 configurations are generated from a narrow set of rules (regular grid, no vegetation, fixed FAR values of 4.5 and 6.5, three orientations, two street widths, and 3–10 story heights), all based on a simplified Tehran morphology. Feature-importance rankings and CFX recommendations may therefore be specific to this distribution; testing on an independent urban dataset or a broader generative distribution is needed before claiming generalizability.

minor comments (6)

[Tables 4 and 5] The notation in Tables 4 and 5 is unclear: empty cells, '0', and values such as 'dN 23 -14' make it difficult to see which features are changed and by how much per strategy. Please add explicit deltas or arrows (e.g., '−14 m') and fill all cells.
[Equation (1)] The Shapley formula in §3.3 is incorrectly typeset; the factorial term and the set difference are garbled. Please rewrite it in standard form: φ_i = Σ_{S⊆N∖{i}} (|S|!(n−|S|−1)!/n!) [v(S∪{i})−v(S)].
[§4.2] The text for visibility SHAP values contains a direction typo: it says 'southeast (hSW)' when hSW denotes southwest. Please check directional abbreviations throughout the SHAP section.
[§4.3] The choice of five counterfactuals per scenario is not justified; a sensitivity analysis showing how the number of counterfactuals affects solution quality would strengthen the optimization claims.
[§4.4 and Figure 5] The box plot in Figure 5 is described but not shown in the manuscript text; please include the figure or state that it is available in supplementary materials.
[References] Some references contain typographical errors ('Chen et al., 2012s', 'Gomez-rodriguez' capitalization); please ensure the reference list is cleaned.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CFX outputs are checked against fresh Grasshopper simulations, so the central speed and accuracy claims do not reduce to their inputs.

full rationale

The claimed derivation chain is not circular. The paper generates 1,152 random urban block scenarios (Section 3.1), computes SVF and visibility labels with Ladybug/Grasshopper, and trains five ML models on an 80/20 split (Section 3.2); XGBoost is then selected on held-out metrics (Section 4.1). The CFX stage searches the trained surrogate for minimal feature changes (Section 3.3), and those recommendations are subsequently re-simulated in Grasshopper (Section 4.4). The reported mean RMSE of 4.12 to 6.06 percent is computed between CFX predictions and those fresh simulations, not against the training labels; the validation truth is therefore not the fitted quantity by construction. The GA-vs-CFX timing comparison is asymmetric in that the CFX '1 minute' excludes offline data generation and surrogate training, and final GA solution quality is not reported, but this is an evaluation-design and reporting concern, not circularity. The self-citations (Eshraghi et al. 2024; Talami et al. 2020) appear only as contextual references for SHAP-based building-performance tools and problem-decomposition methods and are not load-bearing premises for the central claims. No equation or fitted parameter is renamed as a prediction, and no result is forced by a self-citation chain.

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

The central claim is an empirical ML pipeline, not a physical derivation, so the ledger records the modeling choices and domain assumptions that carry the result: the simulation ground truth, the synthetic dataset, the classification thresholds, CFX count, and the GA benchmark settings.

free parameters (4)

Visibility class thresholds = 30-50%, 50-75%, 75-100%
Chosen to convert continuous visibility into three classes; the class boundaries determine the classification targets and the CFX goal of moving up one class (Section 3.2).
Number of counterfactuals per scenario = 5
The CFX run generates exactly five suggested strategies; this choice affects the reported 1-minute runtime and the coverage of solutions (Section 4.4).
Genetic algorithm configuration = population=50, max stagnation=50, boost=2, maintain=5%, inbreeding=75%
Benchmark settings selected by the authors; the speed comparison depends on these GA settings, which are described as defaults rather than tuned (Section 3.4).
ML model hyperparameters = not reported
Table 1 was intended to list parameter values but the content is absent from the provided text; tuning procedure is not described, and final accuracy depends on these choices (Section 3.2).

assumptions (4)

domain assumption Ladybug and Grasshopper simulations accurately compute SVF and visibility for real urban spaces
All training labels and validation outcomes come from these tools with no field measurements or comparison to alternative methods (Sections 3.1 and 4.4).
domain assumption The synthetic randomly generated blocks are representative of Tehran's urban form
Blocks follow a regular grid with 6/12 m streets, 60% or 45% coverage, FAR 4.5/6.5, and 0/40/-40 degree orientations, but are not sampled from actual Tehran geometries (Section 3.1).
domain assumption Vegetation has no effect on SVF or visibility
Trees are explicitly excluded to isolate building adjacency effects, which simplifies the problem but removes a major real-world influencer of SVF and visibility (Section 3.1).
standard math Shapley values and KD-tree counterfactual search work as described in cited literature
The paper adopts SHAP and KD-tree CFX algorithms without deriving or modifying them; correctness is inherited from prior work (Sections 3.3 and 3.4).

how reviews work

0 comments

Cite this review

Pith. "Pith review of An AI-driven framework for rapid and localized optimizations of urban open spaces." pith.science (2026). https://pith.science/paper/GHXA5IAQ

@misc{pith2026250108019,
  author       = {Pith},
  title        = {Pith review of: An AI-driven framework for rapid and localized optimizations of urban open spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GHXA5IAQ}},
  note         = {Machine review of arXiv:2501.08019}
}

read the original abstract

As urbanization accelerates, open spaces are increasingly recognized for their role in enhancing sustainability and well-being, yet they remain underexplored compared to built spaces. This study introduces an AI-driven framework that integrates machine learning models (MLMs) and explainable AI techniques to optimize Sky View Factor (SVF) and visibility, key spatial metrics influencing thermal comfort and perceived safety in urban spaces. Unlike global optimization methods, which are computationally intensive and impractical for localized adjustments, this framework supports incremental design improvements with lower computational costs and greater flexibility. The framework employs SHapley Adaptive Explanations (SHAP) to analyze feature importance and Counterfactual Explanations (CFXs) to propose minimal design changes. Simulations tested five MLMs, identifying XGBoost as the most accurate, with building width, park area, and heights of surrounding buildings as critical for SVF, and distances from southern buildings as key for visibility. Compared to Genetic Algorithms, which required approximately 15/30 minutes across 3/4 generations to converge, the tested CFX approach achieved optimized results in 1 minute with a 5% RMSE error, demonstrating significantly faster performance and suitability for scalable retrofitting strategies. This interpretable and computationally efficient framework advances urban performance optimization, providing data-driven insights and practical retrofitting solutions for enhancing usability and environmental quality across diverse urban contexts.

Figures

Figures reproduced from arXiv: 2501.08019 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

[1]

Alfeo, A.L. et al. (2023) ‘From local counterfactuals to global feature importance: efficient, robust, and model -agnostic explanations for brain connectivity networks’, Computer Methods and Programs in Biomedicine , 236, p. 107550. Available at: https://doi.org/10.1016/j.cmpb.2023.107550. Ali, U. et al. (2020) ‘A data-driven approach to optimize urban sc...

arXiv 2023
[22]

and Mohamed, W

Toutou, A., Fikry, M. and Mohamed, W. (2018) ‘The parametric based optimization framework daylighting and energy performance in residential buildings in hot arid zone’, Alexandria Engineering Journal , 57(4), pp. 3595 –3608. Available at: https://doi.org/10.1016/j.aej.2018.04.006. Tsirtsis, S. and Gomez -rodriguez, M. (2020) ‘Decisions , Counterfactual Ex...

work page arXiv 2018

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Alfeo, A.L. et al. (2023) ‘From local counterfactuals to global feature importance: efficient, robust, and model -agnostic explanations for brain connectivity networks’, Computer Methods and Programs in Biomedicine , 236, p. 107550. Available at: https://doi.org/10.1016/j.cmpb.2023.107550. Ali, U. et al. (2020) ‘A data-driven approach to optimize urban sc...

arXiv 2023

[2] [22]

and Mohamed, W

Toutou, A., Fikry, M. and Mohamed, W. (2018) ‘The parametric based optimization framework daylighting and energy performance in residential buildings in hot arid zone’, Alexandria Engineering Journal , 57(4), pp. 3595 –3608. Available at: https://doi.org/10.1016/j.aej.2018.04.006. Tsirtsis, S. and Gomez -rodriguez, M. (2020) ‘Decisions , Counterfactual Ex...

work page arXiv 2018