SERGIO CUEVAS DEL VALLE; PABLO SOLANO-LÓPEZ; HODEI URRUTXUA

DOI Number: 10.13009/EUCASS2023-056

This work proposes novel Neural Network (NN) training algorithms and architectures to solve with low­cost general Optimal Control problems: regression of the optimal control policy for a given state trajectory. This is achieved via two novelties: first, a novel cost-effective optimal control solver is used for low-cost data augmentation of optimal state-control trajectories, and then combined with Feedforward and General Regression Neural Networks to solve the so-called direct regression problem. Secondly, Pontryagin’s Maximum Principle is leveraged to modify physics-informed NN to embed the Hamiltonian structure of Optimal Control within the training algorithm for enhanced robustness and generalized performance.

Read the full paper here

In Categories: 1. Aerospace Europe Conference 2023, 1.Events, 5. Flight Dynamics/GNC and Avionics
Email
Print
LinkedIn
The paper above was part of  proceedings of a CEAS event and as such the author has signed a publication agreement to have their paper published in the repository. In the case this paper is found somewhere else CEAS always links to the other source.  CEAS takes great care in making the correct content available to the reader. If any mistakes are found  in the listings please contact us directly at papers@aerospacerepository.org and we will correct the listing promptly.  CEAS cannot be held liable either for mistakes in editorial or technical aspects, nor for omissions, nor for the correctness of the content. In particular, CEAS does not guarantee completeness or correctness of information contained in external websites which can be accessed via links from CEAS’s websites. Despite accurate research on the content of such linked external websites, CEAS cannot be held liable for their content. Only the content providers of such external sites are liable for their content. Should you notice any mistake in technical or editorial aspects of the CEAS site, please do not hesitate to inform us.