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

28 CHAPTER 3. EXPERIMENTS 3.2.2 Dense Neural Networks (DNNs) First, we test simple DNN with one hidden layer. Targets is the daily number of rockfall NB t at day t binned in 2 or 3 classes, inputs are previous rain R t − 1 ,.. and previous rockfall number NB t − 1 ... As class targets NB t , even grouped in 2 or 3 classes are very unbalanced (see Table 3.1) , we try weighting loss by frequency of each class. All tested configuration settings for DNNs are listed below, as complete enumeration of configuration settings is numerous, we test them by subset of parameters. Different configuration settings for learning NN classifier DATA PRE-PROCESSING - inputs : rain R t − 1 ,..., R t − k with k ∈ { 5 , 10 } or with previous rockfall NB t − 1 ,..., NB t − k - rescale in [True, False]: rescale input data X by X − mean ( X ) std ( X ) to center data around 0 where relu activation are more efficient - grouping labels in K = 2 or 3 classes LEARNING HYPER-PARAMETER - hidden : nb of neurons in hidden layer [32; 64] - weighting_class in [True, False] : weighting sample in loss function by their frequency in training set - batchsize in [32, 64, 128] : number of samples fed together in model - dropout in [0, 0.2, 0.4] : dropout in the hidden layer to avoid over-fitting - learning_rate in [0.001, 0.0005, 0.0001] : - nb epochs for training 1st DNN Series: 3 class FIXED SETTINGS input : rain with 5 days lag hidden : 64 epochs : 60 TESTED SETTINGS batchsize : 32, 64, 128 lr : 0.0001 , 0.0005, 0.001 weighting : True, False dropout : 0.0, 0.2, 0.4 rescale : True, False 2d DNN Series: 3 class FIXED SETTINGS input : rain epochs : 60 lr : 0.001 TESTED SETTINGS nb-lags :5, 10, 15 hidden : 32 or 64 weighting : True, False dropout : 0.0, 0.2 rescale : True, False