The machine-learning methods in the asteroids dynamics Evgeny Smirnov, [email protected] FB/Telegram: @smirik Pulkovo observatory, Russia 

The list of numbered asteroids in the Solar system has grown significantly in recent years. 

In asteroid dynamics, many problems require numerical integration 
 of the equations of motion 

This approach is 
 computationally expensive Therefore, fast, novel methods can be useful to work with big data 

AI & ML methods have become popular among the IT 

Google Flu case 

Twitter & Earthquake 

ML in astronomy • Outlier detection techniques for Exoplanets (Goel & Montgomery, 2015); • Cosmological parameter estimation via neural network (Hobson et al., 2014); • Identification & classification of active galactic nuclei (Cavouti et al., 2014); • Visualize & classify a large set of Type Ia Supernova spectra (Sasdelli et al., 2016); • Filtering out a large number of false-positive streak detections of near- Earth asteroid candidates in the Palomar Transient Factory (Waszczak et al., 2017); • A Machine Learns to predict the stability of tightly packed planetary systems (Tamayo et al. 2016); • A lot of others… 

Types of ML • Supervised learning: example inputs and desired outputs are provided; the goal is to create a map that binds inputs to outputs. • Unsupervised learning: no examples are provided, the goal is to discover hidden patterns. • Reinforcement learning: the same as supervised learning but instead of a training set there is an environment that provides the rewards based on the actions 

k-nearest neighbours 

Decision Tree 

Gradient Boosting over Decision Trees, Logistic regression, Neural Networks … 

Smirnov, Markov, MNRAS, 2017 

MMR identification using ML Smirnov E.A., Markov A.B. Identification of asteroids trapped inside three- body mean motion resonances: a machine-learning approach. MNRAS. 469. 2017 

MMR identification using ML Smirnov E.A., Markov A.B. Identification of asteroids trapped inside three- body mean motion resonances: a machine-learning approach. MNRAS. 469. 2017 Recall 98,38 % Precision 91,01 % Accuracy 99,97 % 

� ���� ���� ���� ���� ��� ���� ���� ��� ���� ��� ���� ��� ���� ��� � � ���� (480) Hansa family identification using ML 

Statistics Family Koronis Hansa All Recall 99,91 % 100,00 % 98,01 % Precision 77,93 % 84,04 % 50,22 % Accuracy 99,56 % 99,95 % 99,67 % 

It works