Investigation Of Head-on Collisions Of2D Internal Solitary Waves Using Supervised And Semi-supervised Learning

The research on internal solitary waves (ISWs) is a very important part inphysical oceanography. ISWs have significant effects on many things, such as marineecosystems, sea transportation equipments and the navigations of submarines. ISWsare often found propagating in different directions and interacting with each other.Moreover, many researchers have found that both of the interacted ISWs suffered aphase shift after an interaction. Therefore, the studies of tracking and forecasting onISWs focus on the phase shifts. In addition, laboratory studies on ISWs cost a largeamount of time and materials to repeat complex operations to obtain sufficient data,so the efficiency is very low. Hence, it is very promising and meaningful to propose amethod that using small samples to predict the phase shifts of the two interactedISWs.Machine learning is an important field of research in artificial intelligence. It hasgood quality on data mining. In recent decades, machine learning has been widelyused in plenty of areas, such as medical research, network technology, stock marketforecast, and computer vision applications. However, no research has yet been foundon ISWs by machine learning.In this thesis, we studied the head-on collisions of2D ISWs using supervised andsemi-supervised learning. Firstly, the experiment on head-on collisions and theintroduction of machine learning are elaborated. This part provided important samplesand theoretical knowledge for the later sections of this paper. Then, the articlesystematically studied several supervised and semi-supervised learning algorithms,including support vector machines, k-nearest neighbor method and two kinds ofself-training algorithms. We used these algorithms to predict the phase shift of eachISW that suffered a head-on collision. According to the complete experiments and good results, we can confirm that using machine learning to study head-on collisionsof2D ISWs is feasible and effective. Finally, we summarized the advantages anddisadvantages of the above algorithms, and proposed a new semi-supervised methodusing self-training algorithm and SVM-KNN method.