Artificial intelligence (AI) has had a transformative impact in improving the efficiency, safety, and accessibility of transportation systems. Successes in existing deployments have led AI researchers to seek to develop and apply new AI algorithms – particularly those based on deep learning and multi-agent systems – to improve the performance and scalability of transportation. Many of these algorithms have shown strong performance in simulation-based evaluations. However, most of the state-of-the-art algorithms in the AI literature have never been physically deployed.

This thesis argues that a barrier to the deployment of many advanced AI tech- nologies in transportation lies in that their designs are divorced from the key practical considerations of stakeholders. When these AI technologies are to be deployed in existing transportation systems, they face four key categories of challenges:

1. uncertainty in present and projected traffic conditions;
2. heterogeneity among users and deployment contexts;
3. assurance in terms of the understandability and safety of algorithms; and
4. coordination at the individual and system levels.

• Understanding how designs that expose uncertainty in gig driver schedule recommendation can improve users' trust over repeated interactions
• Evaluating the impact of heterogeneity in driver behaviour models and simulation scale on traffic simulation outcomes with statistically rigorous experiments
• Building an algorithmic pipeline for demand modelling in a traffic simulation to incorporate uncertain, heterogeneous detector data and stakeholder feedback
• Imposing safety constraints upon coordinated reinforcement learning (RL)-based TSC policies through a suite of action postprocessing techniques
• Developing a performant but scalable algorithm to distil RL-based TSC policies into coordinated, understandable decision trees

Based on my work, I show that designing AI technologies to better address these challenges also leads to algorithms with technical novelty. Various other problems within transportation (e.g. freight logistics and autonomous driving) and beyond transportation (e.g. robotic navigation and computer networking) are potential applications where my technical contributions could better align AI technologies with stakeholders' needs and preferences. Regardless of the domain, I suggest that the path to the successful deployment of AI technologies lies in designing them with people in mind at every stage of the data-to-deployment pipeline.

214 pages

Thesis Committee:
Fei Fang (Co-chair)
Norman Sadeh (Co-chair)
Sean Qian
Matteo Pozzi
Ryan Shi (University of Pittsburgh)
Peter Stone (The University of Texas at Austin)

Nicolas Christin, Head, Software and Societal Systems Department
Martial Hebert, Dean, School of Computer Science

Creative Commons: CC-BY-NC (Attribution-Non-Commerical)

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu