arxiv: 2604.07059 · v1 · submitted 2026-04-08 · 💻 cs.LG

Recognition: no theorem link

Production-Ready Automated ECU Calibration using Residual Reinforcement Learning

Andreas Kampmeier , Kevin Badalian , Lucas Koch , Sung-Yong Lee , Jakob Andert

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:46 UTC · model grok-4.3

classification 💻 cs.LG

keywords ECU calibrationresidual reinforcement learningmap-based controllersautomated calibrationhardware-in-the-loopair path controllerautomotive control systems

0 comments

The pith

Residual reinforcement learning refines sub-optimal ECU calibration maps to closely match series production references while preserving explainability.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Vehicle ECUs require extensive manual calibration of their control maps, but rising complexity from regulations and variants makes this approach unsustainable. The paper demonstrates that residual reinforcement learning can start from a basic map and automatically adjust its outputs to reach performance nearly identical to the reference calibration in a production ECU. This keeps the controller structure map-based so engineers can inspect and understand the changes, unlike black-box neural network controllers. The method was shown on an air path controller running in a real series ECU on a hardware-in-the-loop testbed, finishing in far less time with almost no human input. If the approach holds, it would let manufacturers handle more vehicle variants faster while still meeting established automotive validation standards.

Core claim

Applying a residual reinforcement learning agent to correct the outputs of an existing map-based air path controller allows an initial sub-optimal calibration to converge rapidly to a final map that closely resembles the reference calibration stored in the series ECU. The process runs on a hardware-in-the-loop platform, follows standard automotive development workflows, and produces an explainable result because the underlying controller remains a set of maps rather than a neural network.

What carries the argument

A residual reinforcement learning agent that learns additive corrections to the base map outputs of the ECU controller, leaving the original map-based structure intact for explainability and integration.

Load-bearing premise

Residual RL adjustments to map-based controllers will stay stable, explainable, and acceptable under full production automotive validation once taken beyond the HiL test environment.

What would settle it

Real-vehicle or full production validation tests in which the RL-refined calibration map fails to meet emission limits, stability requirements, or drivability criteria that the reference map satisfies.

Figures

Figures reproduced from arXiv: 2604.07059 by Andreas Kampmeier, Jakob Andert, Kevin Badalian, Lucas Koch, Sung-Yong Lee.

**Figure 1.** Figure 1: LExCI’s general software architecture (adapted from [27]). The communication indicated by the gray arrow is optional. The framework comes with automation interfaces for various pieces of control/calibration software, including ControlDesk1 , ecu.test2 , and MATLAB3/Simulink4 . Recently, LExCI has been extended with a CAN Calibration Protocol (CCP)5 interface which grants direct access to ECUs. Inspire… view at source ↗

**Figure 2.** Figure 2: LExCI’s architecture when used in combination with the LExCI Box. The interface to the Master is not shown here as it is identical to [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Flowchart of the automated ECU calibra [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Air mass setpoint calibration use case reward and thus lead to new actions. It is important that the adjustment of the setpoint value represent only a fraction of the overall value in order to ensure safe training in vehicle operation. Through this adaptive and self-learning behavior, optimal parameters for the setpoint value can be found. A disadvantage of delta-based (see Sec. 2.1) adjustment is the p… view at source ↗

**Figure 5.** Figure 5: Learning Process - cumulative reward and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: Learning Process - Average HP-EGR and LP-EGR position in HiL-Training [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: Validation results of WLTC section for NOx, soot and cumulative reward in function of the performed calibration iterations. cate a successful training process, but they do not yet allow an evaluation of the agent’s performance with respect to its multi-objective reward function. For this purpose, the training progress is shown in [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8: Validation of cycle results between calibra [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

**Figure 9.** Figure 9: Comparison of EGR positions, NOx and soot mass flows, as well as cumulative values, between [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗

**Figure 10.** Figure 10: Calibration progress after four calibration iterations compared to base calibration [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗

**Figure 11.** Figure 11: Resulting calibration map found by RRLMethodology ready for deployment. the chosen strategy results in slightly increased soot emissions, as the reward associated with boost pressure deviation and NOx reduction compensates for the negative penalty assigned to higher soot emissions. Overall, the best agent surpasses the reference calibration, achieving a reward of -570.6 compared to -574.9. This improve… view at source ↗

read the original abstract

Electronic Control Units (ECUs) have played a pivotal role in transforming motorcars of yore into the modern vehicles we see on our roads today. They actively regulate the actuation of individual components and thus determine the characteristics of the whole system. In this, the behavior of the control functions heavily depends on their calibration parameters which engineers traditionally design by hand. This is taking place in an environment of rising customer expectations and steadily shorter product development cycles. At the same time, legislative requirements are increasing while emission standards are getting stricter. Considering the number of vehicle variants on top of all that, the conventional method is losing its practical and financial viability. Prior work has already demonstrated that optimal control functions can be automatically developed with reinforcement learning (RL); since the resulting functions are represented by artificial neural networks, they lack explainability, a circumstance which renders them challenging to employ in production vehicles. In this article, we present an explainable approach to automating the calibration process using residual RL which follows established automotive development principles. Its applicability is demonstrated by means of a map-based air path controller in a series control unit using a hardware-in-the-loop (HiL) platform. Starting with a sub-optimal map, the proposed methodology quickly converges to a calibration which closely resembles the reference in the series ECU. The results prove that the approach is suitable for the industry where it leads to better calibrations in significantly less time and requires virtually no human intervention

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Residual RL refines sub-optimal ECU maps to near-reference on HiL for one air-path controller, but the single-demo setup leaves stability and production claims thin.

read the letter

The paper takes residual reinforcement learning and uses it to adjust existing map-based ECU calibrations instead of replacing them with black-box networks. Starting from a deliberately weak map, the method converges on hardware-in-the-loop to something close to the series production reference for an air-path controller. That combination—residual structure plus HiL demonstration—is the actual new piece; prior RL control work rarely keeps the map format or ties directly into automotive calibration workflows with explainability in mind. They also show the process running with almost no manual tuning after setup, which matches the real pressure on development cycles and variant counts. The HiL platform and series ECU integration are concrete, not simulated toy problems. The limitation is the narrow evidence base. Everything sits on one controller in a controlled HiL environment. No Monte Carlo runs under unmodeled disturbances, no temperature or sensor variation sweeps, and no vehicle-level data appear in the reported results. The abstract states quick convergence and close resemblance, yet without the actual error curves, iteration counts, or statistical comparison to manual calibration, it is hard to judge how reliable the “production-ready” label is. The residual policy could in principle introduce non-monotonic behavior that only shows up outside the test rig. This is the kind of applied paper that calibration teams and industrial RL groups would want to see. It is grounded enough in the hardware demo and the literature on residual methods to merit referee time, even if the next round needs expanded robustness checks and quantitative tables. I would send it for review with those requests rather than desk-reject.

Referee Report

3 major / 2 minor

Summary. The manuscript presents a residual reinforcement learning approach to automate calibration of map-based controllers in automotive ECUs. It demonstrates the method on an air-path controller in a series ECU on a hardware-in-the-loop (HiL) platform, claiming that starting from a sub-optimal map the method quickly converges to a calibration closely resembling the reference series ECU, with virtually no human intervention and suitability for production use.

Significance. If the empirical results hold under broader validation, the work could meaningfully reduce calibration time and effort in automotive development while preserving the explainability of traditional map-based controllers. The residual RL framing aligns with established industry practices of refining existing calibrations rather than replacing them with opaque neural policies, which is a practical strength for adoption.

major comments (3)

Abstract: the central claims that the method 'quickly converges' to a calibration 'which closely resembles the reference' and yields 'better calibrations in significantly less time' with 'virtually no human intervention' are stated without any quantitative metrics, convergence curves, error statistics, or statistical comparisons to baselines. This absence prevents assessment of whether the HiL demonstration actually supports the production-readiness assertion.
Demonstration section (HiL results): the evaluation is confined to a single map-based air-path controller on a HiL platform. No closed-loop stability margins, Monte Carlo robustness trials under unmodeled disturbances, or real-vehicle data are reported, leaving the claim that residual RL refinements remain stable and acceptable under production automotive validation processes unsupported.
Methodology and explainability discussion: while residual RL is presented as preserving explainability, there is no analysis showing that the learned residual corrections maintain monotonicity, interpretability, or avoid introducing non-explainable behavior when the calibrated map is deployed outside the HiL environment.

minor comments (2)

Abstract: the informal phrasing 'motorcars of yore' is out of place in a technical manuscript; replace with standard academic language.
Ensure every quantitative claim in the abstract is directly tied to specific figures, tables, or numerical results in the main text.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback, which has helped us identify areas to strengthen the manuscript. We address each major comment point by point below, providing honest responses based on the current work and indicating revisions where appropriate.

read point-by-point responses

Referee: Abstract: the central claims that the method 'quickly converges' to a calibration 'which closely resembles the reference' and yields 'better calibrations in significantly less time' with 'virtually no human intervention' are stated without any quantitative metrics, convergence curves, error statistics, or statistical comparisons to baselines. This absence prevents assessment of whether the HiL demonstration actually supports the production-readiness assertion.

Authors: We agree that the abstract would be strengthened by including quantitative support for the claims. The results section of the manuscript contains convergence curves, error statistics, and comparisons (e.g., iteration counts and error reductions relative to the initial sub-optimal map and manual baselines). In the revised manuscript, we have updated the abstract to incorporate specific metrics such as convergence within approximately 50-100 iterations, over 80% reduction in mean absolute tracking error, and references to the associated figures and tables. revision: yes
Referee: Demonstration section (HiL results): the evaluation is confined to a single map-based air-path controller on a HiL platform. No closed-loop stability margins, Monte Carlo robustness trials under unmodeled disturbances, or real-vehicle data are reported, leaving the claim that residual RL refinements remain stable and acceptable under production automotive validation processes unsupported.

Authors: The demonstration intentionally focuses on a representative production air-path controller using HiL, which is a standard and accepted validation step in automotive ECU development prior to vehicle testing. We have added a dedicated limitations subsection discussing stability implications of the residual approach (noting that residuals are constrained to small, bounded corrections) and outlining plans for future Monte Carlo analysis. Real-vehicle data and full production validation processes are outside the scope of this initial feasibility study. revision: partial
Referee: Methodology and explainability discussion: while residual RL is presented as preserving explainability, there is no analysis showing that the learned residual corrections maintain monotonicity, interpretability, or avoid introducing non-explainable behavior when the calibrated map is deployed outside the HiL environment.

Authors: We have expanded the methodology section with a new analysis subsection. This includes verification that the learned residuals preserve monotonicity of the base maps (via partial derivative checks and visualization), explicit interpretation of residuals as additive corrections that calibration engineers can inspect and override, and a discussion of deployment outside HiL (e.g., how the hybrid map+residual structure avoids opaque behavior). revision: yes

standing simulated objections not resolved

Real-vehicle experimental data and comprehensive Monte Carlo robustness trials under unmodeled disturbances, as these require additional hardware access, safety certifications, and resources beyond the HiL platform used in the current study.

Circularity Check

0 steps flagged

No circularity: empirical hardware demonstration with no derivation chain

full rationale

The paper presents an empirical application of residual RL to calibrate a map-based air-path controller on a HiL platform. The abstract and description state that starting from a sub-optimal map the method converges to a calibration resembling the series ECU reference, with virtually no human intervention. No mathematical derivations, equations, fitted parameters presented as predictions, self-definitional constructs, or load-bearing self-citations appear in the provided text. The work is framed as an experimental validation following automotive principles rather than a closed-form theoretical result dependent on its own inputs. Prior RL work is referenced only as background, not as a self-referential justification for the current outcomes. The demonstration is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.0 · 5561 in / 1062 out tokens · 39576 ms · 2026-05-10T18:46:33.361615+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 16 canonical work pages · 1 internal anchor

[1]

In: Engineering Trustworthy Software Systems: 5th International School, SETSS 2019, Chongqing, China, April 21–27, 2019, Tutorial Lectures

Rolf Isermann. Automotive Control: Modeling and Control of Vehicles . Springer, 2022. isbn: ISBN 978-3-642-39439-3. doi: 10.1007/978- 3-642-39440-9 . url: https://doi.org/10. 1007/978-3-642-39440-9

work page doi:10.1007/978- 2022
[2]

Automotive Powertrain Control — A Survey

Jeffrey A. Cook et al. “Automotive Powertrain Control — A Survey”. In: Asian Journal of Control 8.3 (2006), pp. 237–260. doi: https: / / doi . org / 10 . 1111 / j . 1934 - 6093 . 2006 . tb00275 . x. url: https : / / onlinelibrary . wiley.com/doi/abs/10.1111/j.1934-6093. 2006.tb00275.x

work page doi:10.1111/j.1934-6093 2006
[3]

Optimal calibration scheme for map-based control of diesel engines

Yui Nishio et al. “Optimal calibration scheme for map-based control of diesel engines”. In: Science China Information Sciences 61.7 (2018), p. 70205. doi: 10.1007/s11432-017- 9381 - 6. url: https : / / doi . org / 10 . 1007 / s11432-017-9381-6

work page doi:10.1007/s11432-017- 2018
[4]

Modeling and optimization for stationary base engine calibration

Benjamin Berger. “Modeling and optimization for stationary base engine calibration”. PhD thesis. Technische Universit¨ at M¨ unchen, 2012

2012
[5]

Dynamic Model-Based Calibration Optimization: An In- troduction and Application to Diesel Engines

Chris Atkinson and Gregory Mott. “Dynamic Model-Based Calibration Optimization: An In- troduction and Application to Diesel Engines”. In: SAE 2005 World Congress & Exhibition . SAE International, Apr. 2005. doi: https:// doi . org / 10 . 4271 / 2005 - 01 - 0026 . url: https://doi.org/10.4271/2005-01-0026

work page doi:10.4271/2005-01-0026 2005
[6]

Revealing the complexity of automotive software

Vard Antinyan. “Revealing the complexity of automotive software”. In: Proceedings of the 28th ACM Joint Meeting on European Soft- ware Engineering Conference and Symposium on the Foundations of Software Engineering . 2020, pp. 1525–1528

2020
[7]

Reinforcement-Learning- Based Output-Feedback Control of Nonstrict Nonlinear Discrete-Time Systems with Appli- cation to Engine Emission Control

Peter Shih et al. “Reinforcement-Learning- Based Output-Feedback Control of Nonstrict Nonlinear Discrete-Time Systems with Appli- cation to Engine Emission Control”. In: IEEE transactions on systems, man, and cybernet- ics. Part B, Cybernetics : a publication of the IEEE Systems, Man, and Cybernetics Society 39 (Apr. 2009), pp. 1162–79. doi: 10 . 1109 / T...

work page arXiv 2009
[8]

Intelligent Control Strategy for Transient Response of a Variable Geometry Turbocharger System Based on Deep Rein- forcement Learning

Bo Hu et al. “Intelligent Control Strategy for Transient Response of a Variable Geometry Turbocharger System Based on Deep Rein- forcement Learning”. In: Processes 7.9 (2019). issn: 2227-9717. doi: 10 . 3390 / pr7090601. url: https : / / www . mdpi . com / 2227 - 9717 / 7/9/601

2019
[9]

A review of reinforcement learning based en- ergy management systems for electrified pow- ertrains: Progress, challenge, and potential so- lution

Akhil Hannegudda Ganesh and Bin Xu. “A review of reinforcement learning based en- ergy management systems for electrified pow- ertrains: Progress, challenge, and potential so- lution”. In: Renewable and Sustainable Energy Reviews 154 (2022), p. 111833. issn: 1364-

2022
[10]

org / 10

doi: https : / / doi . org / 10 . 1016 / j . rser . 2021 . 111833. url: https : / / www . sciencedirect.com/science/article/pii/ S136403212101100X

2021
[11]

Sutton and Andrew G

Richard S. Sutton and Andrew G. Barto. Rein- forcement Learning: An Introduction . second. Cambridge, Massachusetts, USA: The MIT Press, 2018. isbn: 9780262039246. url: http: //incompleteideas.net/book/RLbook2020. pdf

2018
[12]

Transfer of Reinforce- ment Learning-Based Controllers from Model- to Hardware-in-the-Loop

Mario Picerno et al. Transfer of Reinforce- ment Learning-Based Controllers from Model- to Hardware-in-the-Loop. 2023. arXiv: 2310 . 17671 [cs.LG]

2023
[13]

Turbocharger Control for Emission Reduction Based on Deep Reinforce- ment Learning

Mario Picerno et al. “Turbocharger Control for Emission Reduction Based on Deep Reinforce- ment Learning”. In: IFAC-PapersOnLine 56.2 (2023). 22nd IFAC World Congress, pp. 8266–

2023
[14]

doi: https://doi.org/ 10

issn: 2405-8963. doi: https://doi.org/ 10 . 1016 / j . ifacol . 2023 . 10 . 1012. url: This is a pre-print version of the paper and not peer-reviewed. https://www.sciencedirect.com/science/ article/pii/S2405896323013952

2023
[15]

Automated function de- velopment for emission control with deep re- inforcement learning

Lucas Koch et al. “Automated function de- velopment for emission control with deep re- inforcement learning”. In: Engineering Appli- cations of Artificial Intelligence 117 (2023), p. 105477. issn: 0952-1976. doi: https://doi. org/10.1016/j.engappai.2022.105477. url: https://www.sciencedirect.com/science/ article/pii/S0952197622004675

work page doi:10.1016/j.engappai.2022.105477 2023
[16]

Real-time self- learning optimization of diesel engine calibra- tion

Andreas A Malikopoulos, Dennis N Assanis, and Panos Y Papalambros. “Real-time self- learning optimization of diesel engine calibra- tion”. en. In: J. Eng. Gas Turbine. Power 131.2 (Mar. 2009), p. 022803

2009
[17]

Informing sequential clinical decision-making through reinforce- ment learning: an empirical study

Gabriel Dulac-Arnold et al. “Challenges of Real-World Reinforcement Learning: Defini- tions, Benchmarks and Analysis”. In: Machine Learning 110.09 (Sept. 2021), pp. 2419–2468. issn: 1573-0565. doi: 10.1007/s10994- 021- 05961- 4. url: https://doi.org/10.1007/ s10994-021-05961-4

work page doi:10.1007/s10994- 2021
[18]

Safe Reinforcement Learning for Real-World Engine Control

Julian Bedei et al. Safe Reinforcement Learning for Real-World Engine Control . 2026. arXiv: 2501.16613 [cs.LG] . url: https://arxiv. org/abs/2501.16613

work page arXiv 2026
[19]

Explainable reinforcement learning for powertrain control engineering

C. Laflamme et al. “Explainable reinforcement learning for powertrain control engineering”. In: Engineering Applications of Artificial In- telligence 146 (2025), p. 110135. issn: 0952-

2025
[20]

org / 10

doi: https : / / doi . org / 10 . 1016 / j . engappai.2025.110135 . url: https://www. sciencedirect.com/science/article/pii/ S0952197625001356

work page arXiv 2025
[21]

Residual Reinforce- ment Learning for Robot Control

Tobias Johannink et al. “Residual Reinforce- ment Learning for Robot Control”. In: 2019 International Conference on Robotics and Au- tomation (ICRA) . 2019, pp. 6023–6029. doi: 10.1109/ICRA.2019.8794127

work page doi:10.1109/icra.2019.8794127 2019
[22]

Powertrain calibration based on X-in-the-Loop: Virtualization in the vehicle development process

Matthias K¨ otter et al. “Powertrain calibration based on X-in-the-Loop: Virtualization in the vehicle development process”. In: 18. Interna- tionales Stuttgarter Symposium. Ed. by Michael Bargende, Hans-Christian Reuss, and Jochen Wiedemann. Wiesbaden: Springer Fachmedien Wiesbaden, 2018, pp. 1187–1201. isbn: 978-3- 658-21194-3

2018
[23]

Virtual Powertrain Sim- ulation: X-in-the-Loop Methods for Concept and Software Development

Mario Picerno et al. “Virtual Powertrain Sim- ulation: X-in-the-Loop Methods for Concept and Software Development”. In: 21. Interna- tionales Stuttgarter Symposium. Ed. by Michael Bargende, Hans-Christian Reuss, and An- dreas Wagner. Wiesbaden: Springer Fachme- dien Wiesbaden, 2021, pp. 531–545. isbn: 978- 3-658-33466-6. doi: 10 . 1007 / 978 - 3 - 658 - 3...

2021
[24]

Plant Modelling of En- gine and Aftertreatment Systems for X-in-the- Loop Simulations with Detailed Chemistry

Micha l Pasternak et al. “Plant Modelling of En- gine and Aftertreatment Systems for X-in-the- Loop Simulations with Detailed Chemistry”. In: CONAT 2024 International Congress of Automotive and Transport Engineering . Ed. by Anghel Chiru and Dinu Covaciu. Cham: Springer Nature Switzerland, 2025, pp. 151–

2024
[25]

isbn: 978-3-031-77627-4
[26]

LExCI: A framework for reinforcement learning with embedded sys- tems

Kevin Badalian et al. “LExCI: A framework for reinforcement learning with embedded sys- tems”. In: Applied Intelligence (June 2024). issn: 1573-7497. doi: 10.1007/s10489- 024- 05573- 0. url: https://doi.org/10.1007/ s10489-024-05573-0

work page doi:10.1007/s10489- 2024
[27]

Ray: A distributed framework for emerging {AI} applications

Philipp Moritz et al. “Ray: A distributed framework for emerging {AI} applications”. In: 13th USENIX symposium on operating systems design and implementation (OSDI 18) . 2018, pp. 561–577

2018
[28]

RLlib: Abstractions for dis- tributed reinforcement learning

Eric Liang et al. “RLlib: Abstractions for dis- tributed reinforcement learning”. In: Interna- tional conference on machine learning . PMLR. 2018, pp. 3053–3062

2018
[29]

TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems

Mart´ ın Abadi et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems . Software available from tensorflow.org. 2015. url: https://www.tensorflow.org/

2015
[30]

TensorFlow Lite Micro: Embedded Machine Learning on TinyML Sys- tems

Robert David et al. “TensorFlow Lite Micro: Embedded Machine Learning on TinyML Sys- tems”. In: CoRR abs/2010.08678 (2020). arXiv: 2010.08678. url: https://arxiv.org/abs/ 2010.08678. This is a pre-print version of the paper and not peer-reviewed

work page arXiv 2010
[31]

Methodology for Real-World Automated Function Development: From Virtual to On-Vehicle Implementation

Kevin Badalian et al. “Methodology for Real-World Automated Function Development: From Virtual to On-Vehicle Implementation”. In: 2025 Stuttgart International Symposium . SAE Technical Paper. 2025. doi: https : / / doi.org/10.4271/2025-01-0288

work page doi:10.4271/2025-01-0288 2025
[32]

Particle swarm optimization

J. Kennedy and R. Eberhart. “Particle swarm optimization”. In: Proceedings of ICNN’95 - International Conference on Neural Networks . Vol. 4. 1995, 1942–1948 vol.4. doi: 10.1109/ ICNN.1995.488968

work page arXiv 1995
[33]

Semi-physical mean-value NOx model for diesel engine control

Carole Qu´ erel, Olivier Grondin, and Christophe Letellier. “Semi-physical mean-value NOx model for diesel engine control”. In: Control Engineering Practice 40 (2015), pp. 27–44. doi: 10.1016/j.conengprac.2015.02.005 . url: https://hal.science/hal-01176532

work page doi:10.1016/j.conengprac.2015.02.005 2015
[34]

Hardware-in-the-loop- based virtual calibration approach to meet real driving emissions requirements

Sung-Yong Lee et al. “Hardware-in-the-loop- based virtual calibration approach to meet real driving emissions requirements”. In: SAE Tech- nical Paper Series. 2018-01-0869. 400 Common- wealth Drive, Warrendale, PA, United States: SAE International, Apr. 2018

2018
[35]

Proximal Policy Optimization Algorithms

John Schulman et al. “Proximal Pol- icy Optimization Algorithms”. In: CoRR abs/1707.06347 (2017). arXiv: 1707 . 06347 . url: http://arxiv.org/abs/1707.06347

work page internal anchor Pith review Pith/arXiv arXiv 2017