RINA - Démonstrateur d’utilisation de l’intelligence artificielle pour la gestion opérationnelle des risques naturels gravitaires géologiques | FEREC 2020

test data. The first is used to train the model based and the second to test the capabilities of the model. In this section, different types of models and the criteria will be explained. 4.1. Model Artificial intelligence contains a multitude of different types of models. Some are more complex than others, but they are all adapted to different types of problems. Different types of models will be created and compared between each other and the rules created to control the flow of the road. This problem could be solved as a regression, predicting the mass that fell each day, or as a classification problem, solving simply for the presence of an event or even bining the number of events. The simplest is creating classification models as it is closer to what is done operationally and also the difficulty of not having any event is even more drastic. Three types of classification models were chosen: nearest neighbors, decision tree and Neural Network. For each one, various tunings are needed in order to optimise the model. Python's library scikit learn and tensorflow were used to generate all predictive models. 4.1.1. Nearest Neighbors This is the simplest of the methods but it is also quite effective . The idea is that when a new point of data is added, the model will analyse the data already available and observe a specified number of points which are closer to it. This can be visualised in a 2D space as in Figure 19 but can be translated to higher planes having each dimension a feature. In order to do so it is important to scale all the variables. There are many variations of this method. The simplest one is predicting the point chosen to be the most current between its neighbors. For example, observing the figure with 3 neighbors (plain circle) it would predict the missing point is red as there are 2 red points and one blue. If we changed and considered 5 neighbors (dotted circle) the point would be blue. Another possibility considered in this work is weighting inversely to their distance. This means that closer points will have a larger weight. This means that, for 5 neighbors the point could be predicted as red as they are closer. This is an additional level of complexity that can help to improve the performance of the model. The third variation adopted was weighting the events. As it may be desirable to predict an event when there isn’t in order not to miss some events and as nothing happening is much more frequent than having an event it could be desirable to give a greater weight to one class rather than another. 26