Machine Learning-Based Tire Model Optimization

Abstract

When developing new vehicles, accurate tire models are crucial for simulating realworld driving behavior. However, inconsistencies in measurement and fitting methods between tire manufacturers introduce bias, leading to different models for the same physical tire.

This thesis proposes a method to correct such bias by translating biased tire models into unbiased ones using supervised learning and reinforcement learning (RL). Initially, three supervised classifiers, Artificial Neural Network, Gaussian Naive Bayes, and Random Forest, were trained to identify the origin (manufacturer) of a tire model, serving as a proxy for the unbiased data. RL models were then formulated to optimize towards this classifier proxy, where the agents aimed to reach states the classifier predicted as “unbiased”. To find a good policy, two different RL algorithms were investigated.

The Random Forest classifier demonstrated the highest accuracy, predicting the origin of a tire model with 97.0% confidence. However, the translation results, evaluated by Mean Squared Error, were mixed. While the RL agents learned to improve the classifier’s assessed probability, only some translated curves showed improvement.

Future work may improve the results through systematic hyperparameter tuning of the RL environment and agent. Additionally, access to more labeled tire model data would enable a more precise assessment of variance and bias, supporting the identification and analysis of underlying measurement errors. Despite the limitations, this thesis demonstrates the potential of combining supervised learning and RL for bias correction in tire modeling.