Pith. sign in

REVIEW 3 major objections 4 minor 20 references

Learning-based Multi-agent Race Strategies in Formula 1

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read A reinforcement-learning agent can adapt F1 pit stops, tire choices, and energy use to an opponent's live strategy, drawing only on race-day-observable information.

desk verdict The architecture is sensible and the qualitative results are suggestive, but the evaluation never isolates what the interaction module contributes, so the central claim remains unverified. read the letter →

arxiv 2602.23056 v2 pith:HZHVZZVQ submitted 2026-02-26 cs.AI cs.SYeess.SY

classification cs.AIcs.SYeess.SY
keywords Formula1racestrategyreinforcementlearningmulti-agentself-playpit-stopoptimizationenergymanagementtiredegradation
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 a single reinforcement-learning agent can handle the strategic side of a Formula 1 race against a responding opponent. It builds on an already-trained single-agent policy that decides fuel and battery allocation, pit-stop timing, and tire compound, and adds a lightweight interaction module that takes the opponent's observable state (tire age, pit-stop indicator, gap time) and outputs a correction to the nominal action. A custom self-play scheme — first training against the frozen single-agent policy, then against a pool of previously trained agents — generates a small tournament of distinct policies that are ranked by a chess-style rating system. In head-to-head races on a Bahrain-circuit simulation, the top agent beats all rivals by adapting its pit-stop laps and energy allocation to each opponent while keeping a stable two-stop soft-tire plan. The stated payoff is practical: because the agent uses only information a real pit wall would have, it can support race strategists during a race.

What carries the argument

The key mechanism is a small trainable interaction module appended to a frozen single-agent policy; it takes the ego state plus the opponent's observable variables (tire age, pit-stop indicator, compound-used flag, gap time) and outputs an additive correction to the nominal pit-wall action. The physical coupling between cars is represented by a fitted linear lap-time penalty ΔT_int = a·t_gap + b active for gap times between 0.2 and 1.5 s, with a<0 meaning that following another car slows the trailing car at Bahrain. The reward is total race time plus a small winner bonus, and training alternates the learning agent against a pool of previously trained opponents; a chess-style rating system ra

What would settle it

Replace Eq. (6) with a fitted curve from real F1 telemetry of cars following within 1.5 s at Bahrain, and check whether the champion agent still exists and still beats all rivals: if the linear penalty's slope or intercept changes sign or magnitude, the learned pit-timing adjustments should shift accordingly; if they do not, the agent has memorized the artificial curve rather than adapting to the opponent.

Watch

Extended reading notes

Core claim

The paper claims that a reinforcement-learning agent can produce competitive Formula 1 race strategies against a responding opponent by combining a pre-trained single-agent policy with an interaction module that outputs a correction to the nominal action. The interaction module sees only what a real pit wall would see — tire age, pit-stop indicator, compound-used flag, and gap time — and the agent is trained against a pool of previous selves via self-play. In head-to-head tests on a Bahrain simulation, the resulting champion always adopts a two-stop soft-tire plan but shifts its pit-stop laps by several laps to counter each opponent, and beats all three rivals by large margins. Lower-ranked

Load-bearing premise

The load-bearing premise is the fitted linear aerodynamic interaction model (Eq. 6), a lap-time penalty that depends only on gap time with coefficients fitted to the authors' own game-theoretic simulation; the paper reports no values, no uncertainty, and no validation against real F1 telemetry, so the learned adaptive behaviors are only as trustworthy as that curve.

Editorial extensions

If this is right

  • A race strategist could use the learned agent as a live second opinion, querying it for the next pit-stop, tire-compound, and energy-allocation decision as a race unfolds.
  • Because only race-day-observable inputs are used, the trained policies can be run with real-world telemetry without needing access to competitors' internal states.
  • The chess-style rating system lets new agents be added to the battle arena and ranked against existing ones solely by wins and losses, independent of lap-time reward.
  • The champion's race time barely changes across duels, which the paper presents as evidence that the strategy is not an artifact of one particular opponent.

Reading between the lines

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

  • Editorial extension: the linear interaction model is circuit-specific; at a slipstream-heavy track the slope sign would flip, likely inverting the learned behavior (conserving fuel for an undercut may become less valuable than staying in the wake).
  • Editorial extension: the champion always sticks to a two-stop soft-tire plan; a human opponent could exploit this predictability by forcing an early extra stop. The paper does not explore mixed or deceptive strategies.
  • Editorial extension: the rating gap of about 1000 points may partly reflect the fixed starting handicap (the champion always starts 0.5s behind) rather than pure strategic dominance.
  • Editorial extension: because the reward prioritizes minimizing race time over winning, the agents are unlikely to learn blocking or forced-error tactics; a pure win-maximizing reward could produce more adversarial, but less realistic, strategies.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes a multi-agent reinforcement learning framework for Formula 1 race strategy. Building on a frozen single-agent policy from prior work [10], it adds a trainable interaction module that receives ego-vehicle observations and race-day-observable opponent information (tire compound, tire age, gap time, pit-stop indicators) and outputs a correction to the nominal action. A self-play scheme with random opponent selection and Elo-based ranking is used to generate four agents. In experiments on a Bahrain-like simulation, one agent (A) consistently beats the others, adapting pit-stop timing, tire compound selection, and energy allocation in response to its opponent. The paper claims the framework can support race strategists because it relies only on information available during real races.

Significance. If the central claims are supported, the paper would make a useful contribution by bridging single-agent and multi-agent F1 race-strategy optimization. The architecture—frozen single-agent backbone plus a small trainable interaction module—is pragmatic and likely to stabilize training, and the restriction to observations available to a real pit wall (Eq. (5)) is a strength. The self-play/Elo ranking scheme is a reasonable way to generate and compare agents. However, the contribution is currently demonstrated only qualitatively: there is no baseline against the frozen single-agent policy, no ablation of the interaction module, and the quantitative evidence consists of a handful of deterministic races. The aerodynamic interaction model is fitted to the authors' own simulation and is used for both training and evaluation. If the missing control experiments are added and support the claims, this would be a solid contribution to the RL-for-motorsport literature.

major comments (3)
  1. [Section III-A and IV] The paper's central claim—stated in the abstract and Section I-B—is that the combination of the interaction module and self-play training generates competitive multi-agent policies. This causal claim is not tested. The single-agent policy from [10] appears as the first opponent in Figure 4, but no results are reported for that policy in the same multi-agent environment after training, and there is no ablation with the interaction module disabled (i.e., Δa=0 in Eq. (9)) or with opponent observations removed. Without such controls, the adaptive behaviors in Figures 5 and 6 could arise from the frozen pre-trained backbone or from the reward function alone. The sentence in Section IV-B that the interaction module 'works as intended' is therefore unsupported. Please add a comparison against the frozen single-agent policy and at least one ablated variant.
  2. [Section IV-A, IV-B and Table II] The quantitative evidence is extremely thin. Figure 5 reports a single deterministic duel between A and B, and Figure 6 shows only one race per pairwise matchup. There are no repeated random seeds, no sampled distribution of initial gap times, no error bars, and no distribution of race outcomes. Yet the abstract claims 'robust and consistent race performance,' and Section IV-B asserts consistency for A, B, and C. The Elo scores in Table II are presented without the number of matches, the K-factor, initial ratings, or confidence intervals, and the phrase 'even with different initial gap times (not shown)' is not evidence. The authors should run many episodes per matchup (varying initial gap and any stochastic components) and report outcome distributions, mean/median race times with error bars, and rating uncertainties.
  3. [Section II-C, Eq. (6)] The aerodynamic interaction model—the physical coupling that drives the multi-agent behavior—is a linear function of gap time fitted to the authors' own game-theoretic simulation [15]. No coefficient values (a, b), fit residuals, or validation against real telemetry or an independent simulator are provided. Since the same model is used for both training and evaluation, the learned opponent-aware strategies (undercut, wake avoidance, fuel conservation) could be artifacts of an arbitrary fitted curve. This is a correctness-risk concern, not a circularity claim per se. Please report the fitted coefficients and their uncertainty, and perform a sensitivity analysis (e.g., varying the slope/intercept within plausible ranges) or validate the model against an external source, especially because the paper motivates the framework as deployable for real race strategists.
minor comments (4)
  1. [Section II-B] The model equations are not presented; the text says they are the same as in [10] and 'for the sake of space' omits them. Since [10] is a preprint and the manuscript is intended to be self-contained, the key equations should be included in an appendix or supplementary material.
  2. [Section II-E and III-A] Several important parameters are missing: the reward offset T_c, the winner reward c_win, the SAC hyperparameters, the training schedule, and the details of the self-play opponent-selection procedure. These are needed for reproducibility. The same applies to the Elo rating: K-factor, initial ratings, and number of matches are not specified.
  3. [Section IV-B] The text states that 'the energy management is adjusted accordingly, although omitted here for brevity.' Energy allocation is one of the three central adaptive behaviors claimed in the abstract, so it should be shown for the pairwise duels in Figure 6, not just for the A–B duel in Figure 5.
  4. [Throughout] Minor technical issues: the notation 't_gap ∈ [0.2 s, . . . ,1.5 s]' is non-standard; Eq. (3) has a missing space in 'TWb outlap'; and the references include several preprints that are cited as if published. Please clean up formatting and check all references.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned adaptive behavior is an emergent RL output, not a rename of the fitted interaction model.

full rationale

The paper's central claim is that the interaction module plus self-play training produces agents that adapt pit timing, tire choice, and energy allocation to opponents. That behavior is generated by optimizing the stated reward (Eqs. 14-16) in the MDP of Section II-E; it is not obtained by fitting a target or by definitionally encoding the reported result. The aerodynamic interaction model in Eq. (6) is a fitted input to the environment, but the resulting strategies are emergent outputs of reinforcement learning. The coefficients a and b are not renamed as predictions, and the reported adaptive behaviors are not equal to those coefficients by construction. The paper does rely on self-citations [10] (pre-trained single-agent policy and model equations) and [15] (game-theoretic interaction model), but those provide the simulation backbone rather than the learned result itself. The evaluation is internal to the same simulation, and the lack of an ablation against the frozen single-agent policy or of real-world telemetry validation is a correctness external-validity concern, not a circularity under the stated rules. No step satisfies the requirement of exhibiting a specific reduction of Eq. X to Eq. Y by construction or a fitted parameter renamed as a prediction. Therefore the circularity score is 0.

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

The central claim rests on a chain of modeling assumptions inherited from the authors' earlier work ([10], [15]): the car dynamics, tire degradation, lap-time maps, and the fitted aerodynamic interaction curve. The free parameters are mostly undisclosed constants and training hyperparameters. No genuinely new physical entity is proposed.

free parameters (5)
  • Aerodynamic interaction coefficients a, b = not reported
    Equation (6) fits the extra lap time from slipstream as a·t_gap + b using numerical data from [15]. Both coefficients are fitted, equal for both cars, and no values, confidence intervals, or validation are given.
  • Reward offset T_c = not reported
    The step reward r_k = T_c - T_lap uses a constant offset chosen to keep the reward well-scaled. Its value affects the optimization landscape but not the physical model.
  • Winner reward c_win = not reported
    The final reward is c_win if the agent is ahead at the end, chosen to be about one order of magnitude smaller than the cumulative step reward. This hand-tuning shapes how much agents prioritize winning versus raw pace.
  • SAC hyperparameters and training schedule = not reported
    Section III-A says SAC is used and training takes 3 hours, but gives no learning rates, batch sizes, network widths, or number of self-play iterations. These choices materially affect the final policies and Elo rankings.
  • Initial gap time sampling distribution = not reported
    The initial gap time is randomly sampled before each episode to promote robustness, but the distribution is not specified. For the battle-arena evaluation, agent A always starts 0.5 s behind.
assumptions (5)
  • domain assumption The race car model of [10] is an adequate representation of F1 dynamics
    Section II-B states the same model equations as [10] are used, but the equations are not included and no validation against real race data is performed.
  • domain assumption Slipstream interaction is captured by the linear fitted model in Eq. (6)
    The multi-agent coupling—undercuts, wake avoidance, and reaction to the opponent—is driven entirely by this fitted function. If the linear fit is unrepresentative, the learned strategies are artifacts of the simulator.
  • domain assumption The reward decomposition (lap-time reward plus small win bonus) yields realistic strategies
    Section II-E claims a pure winner reward would encourage detrimental interference; the chosen form assumes minimizing race time with a secondary win incentive is the right objective.
  • domain assumption Self-play with an intermittent random opponent pool produces diverse, comparable agents
    The custom self-play scheme in Section III-A has no convergence guarantee or formal analysis; the quality of the final Elo ranking depends on this unexamined assumption.
  • domain assumption Elo scores computed from races in this simulated environment reflect true policy strength
    Elo ratings are updated from simulated matches, so the ranking is only meaningful if the race outcomes in the simulation are a faithful proxy for policy quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-based Multi-agent Race Strategies in Formula 1." pith.science (2026). https://pith.science/paper/HZHVZZVQ

@misc{pith2026260223056,
  author       = {Pith},
  title        = {Pith review of: Learning-based Multi-agent Race Strategies in Formula 1},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZHVZZVQ}},
  note         = {Machine review of arXiv:2602.23056}
}
read the original abstract

In Formula 1, race strategies are adapted according to evolving race conditions and competitors' actions. This paper proposes a reinforcement learning approach for multi-agent race strategy optimization. Agents learn to balance energy management, tire degradation, aerodynamic interaction, and pit-stop decisions. Building on a pre-trained single-agent policy, we introduce an interaction module that accounts for the behavior of competitors. The combination of the interaction module and a self-play training scheme generates competitive policies, and agents are ranked based on their relative performance. Results show that the agents adapt pit timing, tire selection, and energy allocation in response to opponents, achieving robust and consistent race performance. Because the framework relies only on information available during real races, it can support race strategists' decisions before and during races.

Figures

Figures reproduced from arXiv: 2602.23056 by the authors.

Figure 1
Figure 1. Schematic of the agent-environment interaction. Agent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Schematic of the Race car model. Inputs are the agent’s action a, the gap time to the opponent tgap and the additional lap time caused by the aerodynamic interaction ∆Tint. Output are the observation of the ego car o and the available observations for the opponent ˜o. For a detailed mathematical description, the reader is referred to [10]. where Tnom is the nominal lap time map, ∆Tj is the additional time given by t… view at source ↗
Figure 3
Figure 3. Schematic of the agent’s structure. The single-agent policy is taken [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Custom self-play training scheme. The training agent is shown on [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Pit stop and tire compound strategies for duels between different [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 2 linked inside Pith

  1. [10]

    Towards learning-based Formula 1 race strategies,

    G. Fieni, J. W ¨uthrich, M.-P. Neumann, M. M. Moradi, and C. H. Onder, “Towards learning-based Formula 1 race strategies,”arXiv preprint arXiv:2512.21570, 2025

  2. [15]

    Game theory in Formula 1: From physical to strategic interactions,

    G. Fieni, M.-P. Neumann, F. Furia, A. Caucino, A. Cerofolini, V . Ravaglioli, and C. H. Onder, “Game theory in Formula 1: From physical to strategic interactions,”arXiv preprint arXiv:2503.05421, 2025

  3. [1]

    Virtual strategy engineer: Using artificial neural networks for making race strategy decisions in circuit motorsport,

    A. Heilmeier, A. Thomaser, M. Graf, and J. Betz, “Virtual strategy engineer: Using artificial neural networks for making race strategy decisions in circuit motorsport,”Applied Sciences, vol. 10, no. 21, p. 7805, 2020

  4. [2]

    Minimum-race-time energy allocation strategies for the hybrid- electric Formula 1 power unit,

    P. Duhr, D. Buccheri, C. Balerna, A. Cerofolini, and C. H. On- der, “Minimum-race-time energy allocation strategies for the hybrid- electric Formula 1 power unit,”IEEE Transactions on V ehicular Technology, vol. 72, no. 6, pp. 7035–7050, 2023

  5. [3]

    Maximum-distance race strategies for a fully electric endurance race car,

    J. van Kampen, T. Herrmann, and M. Salazar, “Maximum-distance race strategies for a fully electric endurance race car,”European Journal of Control, vol. 68, p. 100679, 2022

  6. [4]

    Evolutionary F1 race strategy,

    A. Bonomi, E. Turri, and G. Iacca, “Evolutionary F1 race strategy,” inProceedings of the Companion Conference on Genetic and Evolu- tionary Computation, 2023, pp. 1925–1932

  7. [5]

    On the optimization of pit stop strategies via dynamic programming,

    O. F. C. Heine and C. Thraves, “On the optimization of pit stop strategies via dynamic programming,”Central European Journal of Operations Research, vol. 31, no. 1, pp. 239–268, 2023

  8. [6]

    Strategic co-design in Formula 1: Balancing physical configuration and race tactics,

    M.-P. Neumann, G. Fieni, F. Furia, A. Cerofolini, V . Ravaglioli, C. H. Onder, and G. Zardini, “Strategic co-design in Formula 1: Balancing physical configuration and race tactics,” 2024

Show all 20 references
  1. [7]

    Mastering Nordschleife–A comprehen- sive race simulation for AI strategy decision-making in motorsports,

    M. Boettinger and D. Klotz, “Mastering Nordschleife–A comprehen- sive race simulation for AI strategy decision-making in motorsports,” arXiv preprint arXiv:2306.16088, 2023

  2. [8]

    Explainable reinforcement learning for Formula One race strategy,

    D. Thomas, J. Jiang, A. Kori, A. Russo, S. Winkler, S. Sale, J. McMillan, F. Belardinelli, and A. Rago, “Explainable reinforcement learning for Formula One race strategy,” inProceedings of the 40th ACM/SIGAPP Symposium on Applied Computing, 2025, pp. 1090– 1097

  3. [9]

    Formula-E race strategy development using distributed policy gradient reinforcement learning,

    X. Liu, A. Fotouhi, and D. J. Auger, “Formula-E race strategy development using distributed policy gradient reinforcement learning,” Knowledge-Based Systems, vol. 216, p. 106781, 2021

  4. [11]

    Competitors- aware stochastic lap strategy optimisation for race hybrid vehicles,

    L. Paparusso, M. Riani, F. Ruggeri, and F. Braghin, “Competitors- aware stochastic lap strategy optimisation for race hybrid vehicles,” IEEE Transactions on V ehicular Technology, vol. 72, no. 3, pp. 3074– 3089, 2022

  5. [12]

    Model predictive control strategies for electric endurance race cars accounting for competitors’ interactions,

    J. van Kampen, M. Moriggi, F. Braghin, and M. Salazar, “Model predictive control strategies for electric endurance race cars accounting for competitors’ interactions,”IEEE Control Systems Letters, vol. 8, pp. 1799–1804, 2024

  6. [13]

    Optimizing pit stop strategies in Formula 1 with dynamic programming and game theory,

    F. Aguad and C. Thraves, “Optimizing pit stop strategies in Formula 1 with dynamic programming and game theory,”European Journal of Operational Research, vol. 319, no. 3, pp. 908–919, 2024

  7. [14]

    Formula-E multi-car race strategy development—a novel approach using reinforcement learning,

    X. Liu, A. Fotouhi, and D. Auger, “Formula-E multi-car race strategy development—a novel approach using reinforcement learning,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 8, pp. 9524–9534, 2024

  8. [16]

    Interaction-aware multi-agent reinforcement learning for mobile agents with individual goals,

    A. Mohseni-Kabir, D. Isele, and K. Fujimura, “Interaction-aware multi-agent reinforcement learning for mobile agents with individual goals,” in2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 3370–3376

  9. [17]

    Highly accurate protein structure prediction with alphafold,

    J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ron- neberger, K. Tunyasuvunakool, R. Bates, A. ˇZ´ıdek, A. Potapenko et al., “Highly accurate protein structure prediction with alphafold,” nature, vol. 596, no. 7873, pp. 583–589, 2021

  10. [18]

    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. Boltonet al., “Mastering the game of go without human knowledge,”nature, vol. 550, no. 7676, pp. 354–359, 2017

  11. [19]

    A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepelet al., “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,”Science, vol. 362, no. 6419, pp. 1140–1144, 2018

  12. [20]

    The rating of chessplayers, past and present,

    A. E. Elo, “The rating of chessplayers, past and present,”(No Title), 1978

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.