Simultaneous estimation of flexible models and associated hyperparameters: An application to travel activity-duration modeling

Estimation of flexible-statistical models of travel demand involves tuning varying parameters, hyperparameters, manually and iteratively. Proper tuning of hyperparameters results in superior models. However, considerable expertise including technical knowledge of statistics, data mining or machine learning and experience are required to tune hyperparameters and consequently generate appropriate models. Moreover, tuning hyperparameters is prone to subjective error and consequently produces travel demand models that are difficult to reproduce, and extend; and makes the development more an art than a science. There is a need for methods to reduce or eliminate subjectivity during the tuning process. This study proposed a framework to reduce subjectivity during the tuning of hyperparameters required for the estimation of nonparametric models of activity-duration. That is, a flexible-statistical framework, which leverages state-of-the-art innovations in Bayesian optimization (BO), was proposed to estimate Gaussian process models of activity duration and associated hyperparameters. The framework was applied to estimate duration models for five types of out-of-home non-mandatory activity episodes for household individuals in the greater Los Angeles area. Experiments demonstrate that the accuracy of results from the proposed framework are superior than those from the state-of-art tuning process, and are obtained in a fraction of time. The proposed framework could potentially increase the productivity of modelers by reducing time required to tune hyperparameters