Neural network architecture search and model compression for fast prediction of UAS traffic density

The number of daily small Unmanned Aircraft Systems (sUAS) operations in uncontrolled low altitude airspace is expected to reach into the millions in the future. This makes UAS traffic density prediction a critical and challenging problem. In an Unmanned Aircraft System Traffic Management (UTM) framework, an accurate traffic prediction model is the key to manage congestion in very dense areas and allow us to evaluate different mission and resource provision plans in search for the best solution. In our previous work, a deep neural network (DNN) model has been proposed that predicts the instantaneous traffic density based on mission schedule information. However, one of the main drawbacks of the DNN is the high computational cost, which prevents us from applying the model to search for the best mission plan or best locations of launching and landing zones, because it requires exponentially large numbers of predictions based on different input combinations. In this paper, we aim to reduce the complexity of the neural network model. A neural architecture optimization framework that searches for the best compression ratio for each layer is developed. Overall, we are able to reduce the size of the traffic prediction model by 50%. Furthermore, because the pruning adds more regularization on the model and reduces the potential of overfitting, the compressed model also achieves small improvements in the prediction accuracy.