Estimation of cumulative prospect theory-based passenger behavioral models for dynamic pricing & transactive control of shared mobility on demand

Abstract

This thesis studies the optimal design of large-scale shared mobility on demand services (SMoDS) in urban settings. Specifically, we build upon previous work done in the Active-Adaptive Control Lab lab on the dynamic pricing and routing of ride sharing services. We develop and characterize a novel passenger behavioral model based on Cumulative Prospect Theory (CPT) to more accurately represent decision making in the presence of significant risks and uncertainty associated with SMoDS' travel times. A comprehensive survey was designed to estimate both the mode-choice and CPT models. The mode choice section consisted of a series of discrete choice experiments created via factorial design, while the CPT section involved carefully constructed lottery questions and travel choice scenarios to elicit risk preferences. After conducting a pilot study and going through several iterations, the survey was launched via a panel firm.

Data was collected from 1000+ respondents in the Greater Boston metro area. This was used to fully characterize the model and estimate parameters through methods including maximum simulated likelihood estimation, nonlinear least squares and global optimization tools. I also utilized other techniques like regularization and transfer learning to improve the quality of results obtained. Beyond parameter estimation, the uncertainty associated with such behavioral models was quantified via well-established nonlinear programming methods. Sensitivity and robustness analyses were performed to assess the effects of CPT model parametrization errors on the performance of the SMoDS system and objectives like expected revenue, average waiting times etc. These insights were used to design and simulate a closed loop, feedback control mechanism for the SMoDS system to correct modelling errors in real-time, achieve setpoint tracking and enable parameter estimation.

The scheme uses the dynamic tariff as a transactive control input to influence the passenger's behavior as desired. This was implemented via gradient-descent based control schemes to update the price signals, in order to drive the (i) drive the passengers' probabilities of accepting the SMoDS ride offer towards the desired value while also (ii) learning the true passenger behavioral model parameters.