Pith. sign in

REVIEW 5 major objections 9 minor 44 references

Optimal Reward Shaping: Autonomous Car Parking Case Study

T0 review · 5 major / 9 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Reward weights and learning hyperparameters must be tuned together, or parallel-parking RL collapses into paralysis or cowardice.

desk verdict Solid, reproducible parking-RL engineering case study with a real tuned-vs-untuned gap; the “joint HPO is mandatory for non-holonomic control” claim outruns one custom env and weak baselines. read the letter →

arxiv 2607.23617 v1 pith:ZUES2QTW submitted 2026-07-26 cs.LG math.OC

Emre Özkaya , Nicolas R. Gauger This is my paper
classification cs.LGmath.OC
keywords rewardshapingreinforcementlearningautonomousparkingnon-holonomiccontrolDeepQ-NetworkBayesianhyperparameteroptimizationcoverage-gatedalignmentdrive-directionregularization
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

Model-free reinforcement learning for non-holonomic tasks like parallel parking often fails not because the algorithm is weak, but because the reward signal creates bad local minima: the car freezes, jitters in place, or refuses to enter the stall to avoid collisions. This paper argues that a carefully parameterized reward—coverage-gated alignment, a penalty on drive-direction switches, and success termination tied to the same continuous score used for evaluation—can remove those traps. The decisive claim is that those reward constants and the DQN’s own hyperparameters are tightly coupled, so they have to be meta-optimized jointly with Bayesian optimization. On a custom 2D Ackermann parking environment, the co-optimized agent reaches a mean score of 95.6/100 versus about 26.7 for uncalibrated baselines, with smoother trajectories and fewer failure modes.

What carries the argument

A parameterized hybrid reward plus aligned termination: step reward combines time penalty, drive-direction switch cost, lane-coverage standing/delta terms, distance potential, and coverage-gated angular alignment, while early success pays a terminal bonus only when a multiplicative score S (proximity × alignment × lane containment) exceeds a tunable threshold—then jointly optimized with DQN hyperparameters via Bayesian optimization.

What would settle it

Train the same reward structure with fixed default DQN hyperparameters (no joint HPO), and separately train a strong published parking baseline or another algorithm under the same evaluation protocol; if mean score on 1000 roadway spawns stays near 95 without joint tuning, or if variance across seeds swallows the 95.6 vs 26.7 gap, the co-dependency claim fails.

Watch

Extended reading notes

Core claim

Environmental reward parameters and algorithmic hyperparameters are deeply co-dependent; only joint meta-optimization yields stable convergence. With a parameterized reward that gates alignment on lane coverage, regularizes drive-direction switches, and terminates success on the same continuous multi-criteria score used for evaluation, a DQN trained under surrogate-based Bayesian optimization resolves policy paralysis and over-conservative hazard avoidance and lifts mean parking score from roughly 27 to 95.6 out of 100.

Load-bearing premise

Showing that one Bayesian-optimized 14-parameter setup beats uncalibrated and random mixtures on this single custom parking simulator is enough to call joint meta-optimization a mandatory requirement for non-holonomic control in general.

Editorial extensions

If this is right

  • Uncalibrated dense rewards for non-holonomic docking will systematically produce either jittery immobilization or roadway idling unless reward and learning rates are searched together.
  • Binding episode success to the same continuous score used for meta-optimization removes the ‘dead zone’ where agents farm step rewards without finishing.
  • Drive-direction switch penalties can be treated as tunable design variables that suppress gear chatter without hand-coded action masks.
  • Coverage-gated alignment lets coarse navigation and fine heading share one potential field instead of fighting each other far from the stall.
  • Sample-efficient Bayesian HPO becomes part of the training pipeline, not an optional afterthought, for constrained vehicle control.

Reading between the lines

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

  • The same co-dependency pattern likely appears in other tight non-holonomic skills (reverse bay parking, dock-and-load) even if the geometric constants differ.
  • If joint search is mandatory, reported parking-RL results that only tune network hyperparameters or only hand-tune rewards are hard to compare across papers.
  • Aligned termination suggests a general recipe: define one continuous task score and make both the sparse bonus and the outer optimizer objective identical to it.
  • A useful next measurement would be whether the optimized reward transfers to continuous-action or on-policy methods without re-running the full 14-D search.
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

5 major / 9 minor

Summary. The paper studies reward shaping for model-free RL on a custom 2D parallel-parking task with Ackermann (non-holonomic) kinematics. A parameterized dense reward is proposed — coverage-gated alignment penalty, drive-direction switch regularization, lane-coverage induction — and the terminal bonus, early-stop criterion, and HPO objective are all tied to one multiplicative score S∈[0,100] (Eq. 22). A 14-dimensional design space (11 reward + 3 DQN hyperparameters) is searched by LHS-initialized Bayesian optimization: 500 evaluations, each comprising a 10^7-step DQN training run and a 1,000-episode evaluation from randomized roadway spawns with the training curriculum disabled. The optimized configuration attains mean S=95.6 versus 26.7 averaged over uncalibrated/random configurations; characteristic failure modes (policy paralysis, over-conservative hazard avoidance) are described qualitatively. The authors conclude (§I.A.3, §VIII) that joint reward/algorithm meta-optimization is a "mandatory structural requirement" for non-holonomic control.

Significance. If the headline result is robust, this is a useful engineering demonstration for the reward-design and AutoRL communities: the failure-mode taxonomy (policy paralysis vs. hazard aversion) is instructive, and binding termination, terminal bonus, and meta-objective to a single continuous score is a clean, reusable design pattern. Concrete strengths: the environment and reward are fully specified; evaluation is decoupled from training (frozen weights, curriculum disabled, 1,000 randomized spawns); and code plus reproduction instructions are public. The individual components are incremental relative to potential-based shaping and HPO-for-RL, and the single-environment/single-algorithm scope plus a closed-source optimizer limit generality; the generality claim in §VIII exceeds what a case study can establish.

major comments (5)
  1. [§VII.A, §VI (headline result statistics)] The headline 95.6 rests on one training run per design point; no seed variance is reported anywhere in the paper. Since the manuscript itself argues the optimization landscape is highly sensitive and non-convex (§I), Bayesian optimization may have selected a lucky training seed rather than a robust configuration. Retrain the final design under at least 5 seeds and report mean±std. Also report the per-episode distribution, not only the mean: with S hard-zeroed on collision (Eq. 22), a mean of 95.6 on [0,100] is consistent with roughly 4–5% catastrophic failures, and for a parking task the collision rate and success rate must be disclosed separately.
  2. [§I.A.3, §VII.B, §VIII (Contribution 3)] The central structural claim — that reward parameters and algorithm hyperparameters are co-dependent and must be optimized jointly — is asserted rather than demonstrated. The only evidence is joint-optimized vs. the DoE average, which shows that HPO helped, not that the coupling is binding. The needed experiment is a 2x2 factorial: reward-only HPO (algorithm at defaults), algorithm-only HPO (fixed reasonable reward), joint, and neither. If that is out of scope, 'proving that joint optimization is a mandatory structural requirement' (Contribution 3, §VIII) should be softened to what the data actually support.
  3. [§VII.A (baseline construction and attribution)] The 26.7 comparator is the average over uncalibrated/random configurations, which conflates 'HPO helps' with 'random parameters are bad'; the best DoE point is the honest baseline and should be reported. Attribution is also missing: each proposed component should be ablated at the optimized design (drop R_switch; remove the coverage gate on R_align, Eq. 20; replace the S-bound termination, Eq. 28, with a fixed tolerance) to show the structure, not just the tuning, matters. One non-DQN baseline (e.g., PPO or TD3 on the same environment) would bound how much of the result is algorithm-specific.
  4. [§IV.E and Abstract (endogenous evaluation metric)] Every stage of the evidence chain — terminal training reward (Eq. 29), early termination (Eq. 28), meta-objective (Eq. 30), and the reported metric — is the same designed scalar S, and the success gate S_threshold in [94, 99.9] is itself an optimized variable (§VI.B.2). There is therefore no independent verification of task competence. Define a fixed, Φ-independent success criterion (position/heading tolerances, lane containment, collision-free) and report it for the final policy. Relatedly, the Abstract claims outperformance 'across both success rate and trajectory smoothness', but §VII reports only mean S; report these metrics or delete the claim.
  5. [§IV.E vs §VI.B.2 (Eq. 29, w_success contradiction)] The text states w_success = 1.0, while §VI.B.2 lists w_success in [50, 300] as an optimized variable. This matters materially: with S≈95 and w_success up to 300, the terminal windfall (~10^4) exceeds all per-step terms (~10^0) by 3–4 orders of magnitude, changing the effective return scale and its interaction with γ=0.99 and the TD update. State which value was actually used, justify the scale, and make the reward specification internally consistent — as written the central result cannot be reproduced from the text alone.
minor comments (9)
  1. [§II.B vs §II.C] Timestep inconsistency: Eq. (4)–(6) discretize with Δt=0.1 s, but §II.C states a physics increment of Δt=0.016 s per environment step, and Eq. (8) uses Δt again. Clarify decision timestep vs. integration timestep.
  2. [§IV.A vs §IV.D] Notation: R_time is −c_baseline in Eq. (16) but −w_time in the final paragraph of §IV.D; §VI.B.2 uses c_baseline. Unify.
  3. [§IV.C–D] Eq. (18) (R_lane) and Eq. (21) (R_coverage) define the same quantity with the same weights, while Eq. (15) lists R_lane but not R_coverage. Remove the duplication or state explicitly that only one term is applied, to rule out double counting.
  4. [§VII.A vs §VII.C] §VII.A reports a DoE average of 26.7 while §VII.C cites S̄≈34.4 for the baseline drawn from 'the initial DoE sampling phase'; clarify which set each number summarizes, and state how many of the 500 evaluations are LHS initialization vs. BO iterates.
  5. [§I, §IV.C] Related work: potential-based reward shaping (Ng, Harutyunyan & Russell 1999; Wiewiora 2003) is not cited, although the ΔC term in Eq. (18)/(21) is a potential-difference term whose policy-invariance properties deserve comment. The HPO-for-RL/AutoRL literature is also largely absent; positioning against it would sharpen the novelty claim.
  6. [§V.D] Evaluation protocol: state the spawn-perturbation range used in the 1,000 evaluation games (same ±25 px as curriculum stage 1, or wider?), whether the dynamic traffic vehicle (§II.A) is active during evaluation, and move the ε=0 statement from §VII.C.2 into §V.D. Note also that the fixed target (630, 415) is baked into the observation (Eq. 11), so the claims are limited to this stall geometry.
  7. [§VI.A (reproducibility)] rodopt+ is closed-source. Since the search trace is cheap to publish, release the 500 (Φ_k, f_k) pairs and the final Φ* as data; ideally replicate the search with an open-source BO library (e.g., Optuna/BoTorch) to show the finding is not optimizer-specific.
  8. [§II–VII] Presentation: '¤x' in Eqs. (1)–(3) should be ẋ; 'tracking tracking' typo in §III; Eq. (26) mixes max over corners with max against 0 — rewrite as clamped terms; informal phrases ('bankrolls', 'cautious coward') would benefit from a more consistent register.
  9. [Figures 4 and 6] Fig. 4 would be more informative with the (x, y, θ) parameters of each pose; Fig. 6 should state the smoothing window and add seed-variance shading once multi-seed runs exist.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circular derivation; optimizing and reporting the same score S is standard HPO, not definitional circularity.

  1. self definitional [§IV.E (Eqs. 22–29) and §VI (Eq. 30)]
    "an early success condition (terminated=True) is triggered automatically if and only if the absolute evaluation score exceeds the optimized quality cutoff threshold... R_terminal=S·w_success ... The global black-box objective function f(Φ) is defined as the empirical mean evaluation score across these N=1,000 randomized trials."

    Terminal success payout and the outer meta-objective are both defined directly in terms of the same scalar S. This is intentional reward–metric alignment, not a false prediction: the scientific claim is the empirical gap vs uncalibrated baselines after joint HPO, which remains a non-tautological comparison. Flagged only as mild definitional coupling of objective and report metric (standard HPO), not as collapse of the main result.

full rationale

The paper’s central claims are empirical engineering results on a custom parallel-parking MDP: a parameterized reward (coverage-gated alignment, switch regularization, aligned termination) plus joint Bayesian HPO of 14 reward and DQN hyperparameters yields mean evaluation score 95.6/100 versus 26.7 for uncalibrated DoE/random/default configurations, and qualitatively resolves paralysis and over-conservative avoidance. That comparison is a real train-and-evaluate experiment, not a quantity forced by construction from its inputs. Binding early termination and R_terminal to the same continuous score S used as the outer meta-objective (§IV.E, §VI) is an intentional design choice the authors advertise as eliminating reward–success dead zones; it does not make the 95.6 figure tautological relative to the unoptimized baseline, nor does it smuggle a uniqueness theorem or rename a known closed-form result. There are no self-citations that carry the argument, no fitted parameter re-presented as an independent prediction of a distinct observable, and no ansatz imported from prior author work. The only mild circularity-adjacent feature is the ordinary HPO pattern that the quantity maximized (mean S) is also the headline metric—standard practice and not a derivation collapse. Score 1 reflects that minor design alignment without elevating it to a circular proof.

Assumptions & free parameters 11 free parameters · 5 assumptions · 3 invented entities

The central claim rests on a standard MDP/DQN stack, a hand-built 2D Ackermann simulator, a multi-term shaped reward with many free weights, a continuous score S used as both terminal signal and HPO objective, a three-stage training spawn curriculum, and a proprietary Bayesian optimizer. No new physical entities; load-bearing content is engineering assumptions and fitted hyperparameters.

free parameters (11)
  • c_collision (collision penalty) = BO-selected in [0, 50]; exact optimum not tabulated in text
    Sparse terminal crash cost; bounds [0, 50]; jointly fitted by BO to balance dense gradients vs risk aversion.
  • w_dist (distance penalty weight)
    Scales Euclidean distance potential; bounds [0, 2]; fitted in outer loop.
  • w_align (alignment penalty weight)
    Scales heading error only when coverage > C_high; bounds [0, 3]; key to avoiding premature spin minima.
  • c_baseline / w_time (per-step time penalty)
    Constant negative pressure against idling; bounds ~[0.001, 0.1].
  • c_switch (drive-direction switch penalty)
    Penalizes changes among forward/brake/reverse; bounds [0, 5]; intended to kill gear chatter.
  • w_standing, w_delta (lane coverage standing and delta weights)
    Existence bonus and coverage-progress gradient for lane entry; bounds [0,1] and [0,5].
  • C_low, C_high (coverage gate thresholds)
    When distance vs alignment potentials activate; bounds [0,0.3] and [0.8,1.0].
  • S_threshold (early success gate)
    Episode success iff S > threshold; bounds [94.0, 99.9]; directly couples termination to eval metric.
  • w_success (terminal success windfall scale) = stated both as optimized in [50,300] and as 1.0 in §IV.E.3
    Multiplies S on success; bounds [50, 300]; text also states w_success=1.0 in one place—internal inconsistency in manuscript.
  • DQN learning rate α, exploration fraction, ε_final
    Three algorithmic knobs co-optimized with rewards; α∈[1e-5,1e-3], frac∈[0.2,0.5], ε_final∈[0.01,0.05].
  • Fixed env/physics constants (L=60px, v=±100px/s, δ_max=π/6, Δt, arena geometry, D_max=300, etc.) = as listed in §II–IV
    Not BO variables but hand-chosen scales that define the MDP the claim is measured on.
assumptions (5)
  • domain assumption Vehicle motion is adequately modeled by the discrete-time kinematic bicycle model with no slip and clamped steering (Eqs. 1–8).
    Entire policy and reward evaluation assume this non-holonomic simulator; no dynamics, tire forces, or sensing noise.
  • domain assumption The task is a discounted MDP and DQN with MLP(256,256), replay 200k, target update 10k, γ=0.99 can represent near-optimal discrete parking policies.
    §V.A; claim that co-optimization resolves failure modes is tied to this learner class only.
  • ad hoc to paper Mean continuous score S over 1,000 randomized roadway spawns with training curriculum and dense shaping disabled is a valid generalization metric for parking competence.
    §V.D, §VI; S’s multiplicative form and thresholds (D_max, Θ_max, E_max, lane bounds) are paper-specific design choices.
  • domain assumption Latin Hypercube + GP Bayesian optimization with Expected Improvement over 500 expensive evaluations sufficiently maps the 14-D co-dependency landscape.
    §VI; standard BO assumptions (smooth-enough response, GP surrogate fidelity) for the necessity claim.
  • standard math Standard math of atan2 heading wrap, ray-AABB LIDAR, and Euler integration is correct as implemented.
    Used throughout §II–III for state and reward features.
invented entities (3)
  • Coverage-gated alignment feedback (R_align only if C_t > C_high)
    purpose: Delay heading penalties until the car is mostly in the stall to avoid spin local minima.
    Core shaped-reward mechanism claimed to fix objective mismatch; mechanism is a design rule, not an external physical entity.
  • Drive-direction switch regularization R_switch
    purpose: Penalize high-frequency forward/brake/reverse chatter under discrete actions.
    Action-smoothing term parameterized by c_switch; standard regularization idea specialized to drive state.
  • Unified multiplicative evaluation score S and aligned termination (success iff S > S_threshold)
    purpose: Remove mismatch between dense shaping, early stop, and meta-objective; kill cautious-coward dead zones.
    S is defined in §IV.E and used as both payout and BO objective; falsifiable only inside this scoring convention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal Reward Shaping: Autonomous Car Parking Case Study." pith.science (2026). https://pith.science/paper/ZUES2QTW

@misc{pith2026260723617,
  author       = {Pith},
  title        = {Pith review of: Optimal Reward Shaping: Autonomous Car Parking Case Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUES2QTW}},
  note         = {Machine review of arXiv:2607.23617}
}
read the original abstract

Designing effective reward functions for model-free reinforcement learning under non-holonomic constraints remains a persistent challenge, often resulting in severe local minima such as policy paralysis or over-conservative hazard avoidance. In this work, we present a parameterized reward shaping framework featuring coverage-gated alignment feedback, drive-direction switch regularization, and an aligned episode termination mechanism evaluated on an autonomous parallel parking task. Crucially, we show that environmental reward parameters and algorithmic hyperparameters are deeply co-dependent, requiring joint meta-optimization to achieve stable convergence. By employing surrogate-based Bayesian optimization, our co-optimized Deep Q-Network (DQN) agent resolves characteristic control failure modes, significantly outperforming uncalibrated baselines across both success rate and trajectory smoothness.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 3 canonical work pages

  1. [1]

    Sensors , volume=

    Data efficient reinforcement learning for integrated lateral planning and control in automated parking system , author=. Sensors , volume=. 2020 , publisher=

    2020

  2. [2]

    Sensors , volume=

    A Hierarchical NMPC and TD3-Based Framework for Seamless Cruise-to-Park Automated Valet Parking , author=. Sensors , volume=. 2026 , publisher=

    2026

  3. [3]

    Journal of Intelligent and Connected Vehicles , volume=

    Development of Deep-Learning-Based Autonomous Agents for Low-Speed Maneuvering in Unity , author=. Journal of Intelligent and Connected Vehicles , volume=. 2024 , publisher=

    2024

  4. [4]

    arXiv preprint arXiv:2006.14171 , year=

    A Closer Look at Invalid Action Masking in Policy Gradient Algorithms , author=. arXiv preprint arXiv:2006.14171 , year=

    arXiv 2006

  5. [5]

    2018 , publisher=

    Reinforcement learning: An introduction , author=. 2018 , publisher=

    2018

  6. [6]

    Proceedings of the International Conference on Automated Planning and Scheduling , volume=

    Goal-Based Action Priors , author=. Proceedings of the International Conference on Automated Planning and Scheduling , volume=

  7. [7]

    AIAA Journal , volume=

    Reverse Flow Radius in Vortex Chambers , author=. AIAA Journal , volume=. 1986 , publisher=

    1986

  8. [8]

    An Introduction to the Adjoint Approach to Design , volume =

    Giles, Michael and Pierce, Niles , year =. An Introduction to the Adjoint Approach to Design , volume =. Flow, Turbulence and Combustion , doi =

Show all 44 references
  1. [9]

    Nature , volume=

    Human-level control through deep reinforcement learning , author=. Nature , volume=. 2015 , publisher=

    2015

  2. [10]

    Journal of Machine Learning Research , year =

    Antonin Raffin and Ashley Hill and Adam Gleave and Anssi Kanervisto and Maximilian Ernestus and Noah Dormann , title =. Journal of Machine Learning Research , year =

  3. [11]

    Peter and Richard P

    Jacques E.V. Peter and Richard P. Dwight , keywords =. Numerical sensitivity analysis for aerodynamic optimization: A survey of approaches , journal =. 2010 , issn =. doi:https://doi.org/10.1016/j.compfluid.2009.09.013 , url =

    2010 doi

  4. [12]

    A Consistent and Robust Discrete Adjoint Solver for the SU2 Framework—Validation and Application , isbn =

    Albring, Tim and Sagebaum, Max and Gauger, Nicolas , year =. A Consistent and Robust Discrete Adjoint Solver for the SU2 Framework—Validation and Application , isbn =

  5. [13]

    ACM Transactions on Mathematical Software (TOMS) , volume =

    High-Performance Derivative Computations using CoDiPack , author =. ACM Transactions on Mathematical Software (TOMS) , volume =. 2019 , date =

    2019

  6. [14]

    Aviation Week & Space Technology , volume=

    Planetary Flight Surge Faces Budget Realities , author=. Aviation Week & Space Technology , volume=

  7. [15]

    Space News , volume=

    NASA Considers Switch to Delta 2 , author=. Space News , volume=

  8. [16]

    Computational Methods for Fluid Flow , edition=

    Peyret, Roger and Taylor, Thomas D , year=. Computational Methods for Fluid Flow , edition=

  9. [17]

    1984 , publisher=

    Aerothermodynamics of Gas Turbine and Rocket Propulsion , series=. 1984 , publisher=

    1984

  10. [18]

    Teleoperation and Robotics in Space , series=

    Techniques for Collision Prevention, Impact Stability, and Force Control by Space Manipulators , author=. Teleoperation and Robotics in Space , series=. 1994 , publisher=

    1994

  11. [19]

    AIAA Guidance, Navigation, and Control Conference , series=

    Spacecraft Thermal Control, Design, and Operation , author=. AIAA Guidance, Navigation, and Control Conference , series=. 1989 , pages=

    1989

  12. [20]

    Fluid Mechanics Proceedings , editor=

  13. [21]

    AIAA Paper 2016--3690 , year=

    Nonequilibrium Radiation for Earth Entry , author=. AIAA Paper 2016--3690 , year=

    2016

  14. [22]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  15. [23]

    Advances in neural information processing systems , volume=

    BoTorch: A framework for efficient Monte-Carlo Bayesian optimization , author=. Advances in neural information processing systems , volume=

  16. [24]

    2016 , publisher=

    SU2: An open-source suite for multiphysics simulation and design , author=. 2016 , publisher=

    2016

  17. [25]

    arXiv preprint arXiv:2010.02502 , year=

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

    2010 arXiv

  18. [26]

    AIAA journal , volume=

    Geometric comparison of aerofoil shape parameterization methods , author=. AIAA journal , volume=. 2017 , publisher=

    2017

  19. [27]

    International Conference on Artificial Intelligence and Statistics , pages=

    Scalable constrained Bayesian optimization , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2021 , organization=

    2021

  20. [28]

    Visualizing the loss landscape of neural nets , volume =

    Hao Li and Zheng Xu and Gavin Taylor and Christoph Studer and Tom Goldstein , issue =. Visualizing the loss landscape of neural nets , volume =. Advances in Neural Information Processing Systems , pages =

  21. [29]

    Foundations of Computational Mathematics , volume=

    A Riemannian view on shape optimization , author=. Foundations of Computational Mathematics , volume=. 2014 , publisher=

    2014

  22. [30]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  23. [31]

    arXiv preprint arXiv:2403.13219 , year=

    Diffusion model for data-driven black-box optimization , author=. arXiv preprint arXiv:2403.13219 , year=

    arXiv

  24. [32]

    2014 , publisher=

    Optimal shape design using an unsteady continuous adjoint approach , author=. 2014 , publisher=

    2014

  25. [33]

    International conference on machine learning , pages=

    Improved denoising diffusion probabilistic models , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  26. [34]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    arXiv

  27. [35]

    , title =

    Selig, Michael S. , title =

  28. [36]

    Mathematical programming , volume=

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming , author=. Mathematical programming , volume=. 2006 , publisher=

    2006

  29. [37]

    Forschungsbericht- Deutsche Forschungs- und Versuchsanstalt fur Luft- und Raumfahrt , year=

    A software package for sequential quadratic programming , author=. Forschungsbericht- Deutsche Forschungs- und Versuchsanstalt fur Luft- und Raumfahrt , year=

  30. [38]

    Proper Orthogonal Decomposition, surrogate modelling and evolutionary optimization in aerodynamic design , journal =

    Emiliano Iuliano and Domenico Quagliarella , keywords =. Proper Orthogonal Decomposition, surrogate modelling and evolutionary optimization in aerodynamic design , journal =. 2013 , issn =. doi:https://doi.org/10.1016/j.compfluid.2013.06.007 , url =

    2013 doi

  31. [39]

    2016 , issn =

    Improving kriging surrogates of high-dimensional design models by Partial Least Squares dimension reduction , journal =. 2016 , issn =. doi:https://doi.org/10.1007/s00158-015-1395-9 , url =

    2016 doi

  32. [40]

    17th AIAA/ISSMO multidisciplinary analysis and optimization conference , pages=

    Efficient aerodynamic design using the discrete adjoint method in SU2 , author=. 17th AIAA/ISSMO multidisciplinary analysis and optimization conference , pages=

  33. [41]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  34. [42]

    arXiv preprint arXiv:2410.08727 , year=

    Losing dimensions: Geometric memorization in generative diffusion , author=. arXiv preprint arXiv:2410.08727 , year=

  35. [43]

    Forty-first International Conference on Machine Learning , year=

    Diffusion models encode the intrinsic dimension of data manifolds , author=. Forty-first International Conference on Machine Learning , year=

  36. [44]

    International conference on machine learning , pages=

    Deep unsupervised learning using nonequilibrium thermodynamics , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.