Year: 2024
Type: dissertation
Abstract: The discovery of the optimal trajectory has consistently captivated the interest of the scientific community. With the development of Machine Learning and Deep Learning techniques, solutions have been developed to address this challenge, particularly through the use of Artificial Neural Networks. This dissertation focuses on creating a Feed Forward Neural Network to predict the optimal path for Formula 1 and Deutsche Tourenwagen Masters (DTM) tracks. Initially, 76 circuits were generated, augmented to a total of 5533 tracks through the application of Data Augmentation techniques, and then fed into the traditional high-accuracy Optimal Control Problem model. Afterwards, data transformation was applied and preprocessed. Numerous experiments were performed to develop a predictive model that is both simple and effec tive, with a particular focus on varying hidden layers and neurons. During the model training process, observations were made regarding the values of loss, mean squared error (MSE), mean absolute error (MAE), and root mean squared error (RMSE). This analysis led to the determination that a model with a single hidden layer containing 40 neurons produced the best results, yielding an MSE of 0,048. Two final experiments were performed by increasing the number of epochs to 100 and for two different foresight values: 30 and 100. With these modified models, we generated predictions for three tracks, each varying in complexity. The results indicated that for simpler tracks a lower foresight value suffices, resulting in an MSE of 0,076. However, for more complex tracks, it leads to a higher error. On the other hand, a more complex model will result in a higher error on circuits with fewer bends and a lower error on more complex tracks.
Author: Moreira, Inês Mafalda André Ramos
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6998459208