Controlling the landing of a spaceship is a complex and relevant problem. There are many observable parameters that drastically affect the system state and whose relevance changes at each stage of the landing. There are closed-loop implementations of controllers such as Proportional-Integral-Derivative (PID) and Model Predictive Control (MPC) to steer a spaceship. Unlike PID, MPC needs a model of the real system to work, making the quality of the model crucial.
Firstly, it is complicated to filter out the features that are vital for the model prediction from a plethora of observable values. There are too many features to be generated and then selected. Thus, features’ importance and influence on the model’s outcome are analysed to improve the efficiency and accuracy of the model.
Secondly, the predictor part of the MPC needs to predict the state well and reliably so the optimizer part can determine the right control parameters to land the spaceship using the predictor as a reference. Various regression models, that predict the altitude and vertical velocity at multiple timesteps, are trained and evaluated. Afterwards, they are integrated into the predictor part of MPC and used in landing.
Machine Learning models predicted altitude and vertical velocity quantitatively and qualitatively better than MPC’s standard Newtonian predictor. Furthermore, Machine Learning models demonstrated that it is possible to land spaceships successfully when used as the predictor in MPC.