Neural network approach to the state classification problem implemented with Keras.
"Neural State Classification for Hybrid Systems" ATVA 2018 (https://arxiv.org/pdf/1807.09901.pdf).
The 'models' directory contains three projects: 'helicopter_models', 'parametric_neuron_models', and 'powertrain_models'.
'helicopter_models/networks.ipynb' implements a 1D CNN trained on 1 million example configurations from a hybrid automata model of a helicopter control system.
CNN Architecture: 16 3x1 convolutions => 32 3x16 convolutions => FC 2048 => FC 1024 => FC 512 => FC1
Dropout is applied in all fully connected layers except the output layer. Batch normalization is applied to both convolutional and fully connected layers. Adagrad optimizer is used with He normal initialization.
Accuracy: 98.67%
'parametric_neuron_models/networks.ipynb' contains 5 models trained on samples from a parametrized neuron model where the number of parameters varied between 1 and 5.
Accuracy: 97-99%
'powertrain_models/networks.ipynb' implements DNN and 1D CNN for a powertrain model. Best accuracy trained on 20k samples is ~95%.