A study compares Gaussian processes, recurring networks and a model with normalizing waves, FloTS, to anticipate the image quality of the sky, useful for telescopes as well as optical links in free space.
The "seeing", which measures the blur introduced by atmospheric turbulence, conditions the quality of astronomical observations as well as the performance of optical communications in free space. Mary-Joe Medlej, Rahul Srinivasan, Simon Prunet, Aziz Ziad and Christophe Giordano compared several approaches to machine learning to predict it up to two hours in advance, emphasizing the quantification of uncertainty. Two probabilistic methods — Gaussian processes and a normalizing float model called FloTS — are faced with conventional recurring networks (RNN and LSTM), all of which are trained and evaluated on measurements at the top of Maulakea in Hawaii.
The models were judged on both deterministic metrics (average quadratic error, Pearson correlation) and on the quality and calibration of their uncertainty estimates. The LSTM and FloTS achieve the best overall accuracy. Gaussian processes provide a useful probabilistic reference, but their Gaussian hypothesis limits the shape of predicted distributions. FloTS combines competitive precision and more flexible uncertainty: it is it that delivers the most informative probabilistic predictions, what better to plan the use of instruments over the night.
Read the full article on the original source
Go to the source →