Tire Model Optimization via LLM Agents

LLM agent code of the paper "Genetic and agentic symbolic regression of distributed rate-and-state friction models" by Luigi Romano (Linköping University), Alessandro Lucantonio (Aarhus University), and Marco Virgolin (Unlayer AI), published at the Workshop on Symbolic Regression and Equation Discovery. The GP code of the paper is available at https://github.com/cpml-au/SR-Tyre

👀 Note: this paper inspired the creation of an agent SKILL for model discovery, see automodel

Prerequisites

• Python 3.13+
• uv

Setup

uv sync

Structure

• code/tire_model_core.py contains the shared data preparation and split-aware evaluation logic.
• code/tire_model_loss.py is the agent-facing loss entrypoint. It exposes only training and validation metrics.
• code/tire_model_report.py is the human-facing plotting script with final test metrics.
• code/tire_model_experiment.py initializes experiment runs, records each attempt, and finalizes the holdout test evaluation.

Model Loss

uv run python code/tire_model_loss.py

This evaluates the current mu_func.py with no plots and prints JSON containing:

• the training loss,
• the validation loss,
• the visible per-bin losses for the training and validation bins.

The default split is bins 1,2,3 for training and bin 4 for validation (bin 5 is the holdout test bin). The holdout test metric is excluded from this command, but the holdout data and evaluation code are public. The split is therefore honor-based: agents are expected not to inspect or evaluate the test bin before finishing a run. You can override the visible bins:

uv run python code/tire_model_loss.py \
  --training-bins 1 2 3 \
  --validation-bins 4

Experiment Runs

To run the experiment with your agent, simply paste the content of initial_prompt.txt to kick-start the run.

read AGENTS.md and run the optimization as instructed for 20 iterations

Your agent should create a timestamped run directory in logs/ containing:

• run.json with the run manifest,
• agent_prompt.md with the exact prompt and commands for the agent,
• attempts.csv and attempts.jsonl with one record per attempt,
• attempts/attempt_XXX_mu_func.py snapshots with comment headers,
• final_summary.json after the run is finished.

The manifest, per-attempt records, and final summary also store the agent name and model used for the run.

After each change the agent makes to code/mu_func.py, it will record the attempt with:

uv run python code/tire_model_experiment.py record-attempt \
  --run-dir logs/<run-dir> \
  --note "short reason"

When the optimization loop is done, the agent finalizes the run and reveals the holdout test metrics (it is instructed to do so only then, and exactly once):

uv run python code/tire_model_experiment.py finish --run-dir logs/<run-dir>

finish requires at least one recorded attempt and verifies that the current code/mu_func.py exactly matches the latest recorded attempt.

Human Report

uv run python code/tire_model_report.py

For headless checks without opening a window:

uv run python code/tire_model_report.py --no-show

License

This project is licensed under the MIT License. See LICENSE.

Citation

If you use this code in your research, please cite the paper:

@inproceedings{romano2026genetic,
  title={Genetic and agentic symbolic regression of distributed rate-and-state friction models},
  author={Romano, Luigi and Lucantonio, Alessandro and Virgolin, Marco},
  booktitle={Workshop on Symbolic Regression and Equation Discovery (IEEE WCCI/CEC 2026)},
  year={2026}
}