Markov Chain Monte Carlo Extensions to Gaussian Processes Approach on Orbit Predictions
Hao Peng, Xiaoli Bai · AIAA Scitech 2020 Forum · 2020
The previously proposed machine learning (ML) approach can improve the satellite orbit prediction accuracy via learning from historical data. This study attempts to extend the ML approach from a new perspective. Specifically, Gaussian Processes (GPs) with the Markov chain Monte Carlo (MCMC) sampling technique are used. We compared several different combinations of GPs, whether it is a full or sparse GP, inferred using maximum a posterior (MAP) algorithm or MCMC integration. We observed that MCMC can provide better performance than MAP in most cases, but it also requires more computational time. Other improvements to GPs and MCMC will be studied in the near future.