An Innovative Neuro-Genetic Algorithm and Geometric Loss Function for Mobility Prediction

Abstract

In this research we design a time series geo-location prediction model based on Long Short-Term Memory (LSTM) with a custom geometric loss function. In order to estimate a close to optimal LSTM Recurrent Neural Network (RNN) architecture we use an innovative Genetic Algorithm (GA) tailored for RNN hypertuning. The proposed Neuro-Genetic Algorithm (Neuro-GA) includes a similarity function for the selection of the RNN that will be recombined and an early stopping criterion for the worse performing RNNs. In addition, we examine the applicability of an incremental learning approach for personalized RNN modeling. Compared with auto-machine learning and deep learning models, the proposed methodology shows substantially better prediction results and the early stopping criterion improves the speed of hypertuning convergence. The experiments also show that the incremental learning approach has significant better accuracy than a generic RNN as the personalized models are retrained to new users location data.