XGBoost Gradient-Boosted Tree Predictions Using Limited Data for Coaxial Rotor Helicopters

Abstract

The use of finite-state methods is critical to the development of accurate and efficient inflow models used in rotorcraft flight dynamics simulation and control. Recent work in the finite-state field has allowed for the application of these models to multirotor systems using the adjoint theorem, which involves time delays and adjoint variables. However, the addition of time delays and adjoint variables drives the necessity for the addition of further inflow states to achieve model accuracy. Computation with a higher numbers of inflow states requires greater computing power and therefore limits the ability of real-time analysis. To help mitigate these issues, this paper explores the use of a gradient booted trees in XGBoost^TM, as well as the use of varied, lower state training data and limited higher state training data, to accurately predict the velocity on the lower rotor of a coaxial rotor helicopter. The investigation involves XGBoost hyperparameter searches to determine the best model, variation in training and testing subset splits, and use of validation subset comparisons for identifying the best performing model.