Research Focus:
Human-Machine Teaming for Smart Transportation Planning
Human decision-makers and decision support tools are increasingly interactive in transportation planning processes. My research focuses on utilizing the strengths and avoiding the weaknesses of human planners and artificial intelligence (AI) for guiding the urban dynamics. The objective is to provide people accessibility to opportunities and resources in a sustainable, equitable, and efficient manner.
​​
I approach this research problem through two mutually enhancing tracks. The first track develops and applies reinforcement learning (RL) to assist transportation planners in strategic decision-making and take into account human planning biases and administrative latencies. As the first track requires a model for the environment where a RL agent reside, the second track develops and applies resolution-flexible urban dynamics models.
Track 1: Planners-in-the-Loop Reinforcement Learning
Existing planning models tend to be predictive rather than prescriptive. This track involves developing and applying RL algorithms for recommending optimal decisions and their implementation timing. However, human urban and transportation planners will continue being critical components in the urban development processes. Therefore, it is important for the RL algorithm to understand the performance metrics defined by human planners, common planning biases, and delays (e.g., administrative, political, construction).
​
Which transportation infrastructure projects are implemented affect how the economy and environment evolve and perform in the long term; the implementation schedule and the actual costs of projects are also important. It is a recurrent issue that the implementations of many transportation policies and projects tend to experience significant delays and cost escalations (if not eventually cancelled), due to technical, psychological, political, and managerial challenges. Moreover, systematic underestimation of project expenses, durations, and risks may form a vicious cycle by, in effect, incentivizing larger underestimations and penalizing honest and accurate ones.
​
I study planning biases and delays in transportation projects and propose modeling solutions for capturing various types of delays and path-dependence in smart mobility systems. The studies and modeling efforts set the foundation for the reinforcement learning algorithm developed in the first research track to mitigating planning biases and implementation delays through correctly employing various feedback loops and delays in urban dynamics
Track 2: Resolution-Flexible Urban Dynamics Modeling
This track develops a multi-resolution, multi-paradigm, and multi-timescale modeling system.. Calibration and validation are major challenges in this step -- the developed system of models need to be consistent with the empirical data as well as consistent internally across different model resolutions.
​
​