Transportation Engineering and Science Program
The principal idea behind pursuing multi-dimensional activities undertaken in this NSF project is
to understand (and where possible reduce) "driving volatility" in instantaneous driving decisions and increase driving
and locational stability. Consequently, driving volatility was operationalized at both the driver level, when drivers
undertake trips, and at the network level. Driving volatility helps us understand instantaneous driving decisions in
a connected vehicle environment.
The project team has expanded knowledge about instantaneous driving decisions by making conceptual, methodological,
and empirical contributions. Methodologically, we introduce the applications of Dynamic Markov Switching models,
Markov-Decision Processes, Inverse Reinforcement Learning, and hierarchical models to large-scale CAV data. Such
techniques are contributing to the scientific analysis of real-world connected vehicle data, and extracting actionable
information embedded in such data.
Initially the concept of driving volatility was operationalized at the trip level as the percentage of time a driver's
acceleration or braking is more than the mean plus one or two standard deviations above the accelerations or braking
of a large sample at that speed. In terms of developing new ideas, we have worked on driving volatility by understanding
acceleration and braking regimes in instantaneous trip-level driving decisions. A rich CAV database was used to
empirically explore regimes using Markov switching models and more broadly the microscopic decisions made by drivers.
Questions answered include when these regimes change and quantifying the volatility associated with each regime.
Another new idea is extending volatility to specific locations in a network, termed "location-based volatility."
For location-based volatility, a unique database was assembled by integrating CAV data with crash and inventory
data at intersections. This enabled associating location-based volatility with real-life crashes, generating new
insights. Continuing analysis of high resolution connected vehicle data, the Markov Decision Process and Inverse
Reinforcement Learning are applied to explain observed driving behavior. Personally optimal policies that maximize
the expected sum of rewards for individual drivers are explored in terms of drivers making decisions about acceleration,
braking, and maintaining speed.
The new ideas are meant to enhance technology development, by developing tools and algorithms that address research
needs in connected and automated vehicle space. The study is focusing on science and engineering of new CAV systems
and results-oriented technology development (Figure 1). In particular, by developing and applying new data analytic
techniques, the study enables new discoveries and generation of actionable information for innovative real-life
applications. Ultimately, the fundamental knowledge generated can provide a strong behavioral basis for CAV applications
including hazard notification and real-time or predictive driver feedback. Overall, the basic methods for extracting
useful information from CAV and related data are developed that will enhance safety, fuel efficiency and reduce emissions.
In this study, the dynamics of driving regimes extracted from Basic Safety Messages transmitted between connected
vehicles are analyzed at a microscopic level. We generate new knowledge for connected vehicle technologies by explicitly
investigating time-series instantaneous driving decisions of connected vehicle drivers, and mapping such decisions to
instantaneous driving contexts. This analysis is important because driving decisions primarily depend on surrounding
traffic states, and a detailed understanding of driving decisions can significantly help us with better anticipating
hazardous situations and providing warnings and alerts to drivers. By introducing application of Markov-Switching models
to large-scale CAV data, we characterize typical driving cycle into different regimes. The exhaustive analysis generates
meaningful information about the existence of different regimes, when do the regimes change and how long they last, and
key correlates associated with each regime. Finally, the study provides a new fundamental framework for making short-term
instantaneous driving and regime predictions at specific instances in time, which is critical for developing connected
vehicle hazard anticipation and notification systems.
A key idea explored in this study is to establish context-relevant alert, warning, and control assist thresholds based
on extreme event information. A rigorous analytical methodology introduces speed varying thresholds to account for
different driving situations and contexts. The study generates new knowledge by making sense of high-frequency
geo-referenced connected vehicle data, and extracts, through data analytic techniques, critical information about
extreme events from connected vehicle data.
From a driver behavior perspective, an in-depth understanding of drivers' policies related to instantaneous driving
decisions made on different roadway functional classifications can help improve the effectiveness of transportation
systems. As such, by examining trip-level microscopic instantaneous driving decisions, we study and analyze drivers'
personally optimal policies on a specific trip. Given the availability of large-scale high-resolution connected vehicle
data, instantaneous driving decisions are treated as realization of an optimal driver policy in a Markov Decision Process
framework. The proposed framework allows defining the progression of different driving states over time in terms of
objects surrounding a vehicle. Different actions (i.e. acceleration, braking, or maintaining constant speed) can be
taken by drivers, and a value is derived for each action, which can be quantified in terms of accumulated discounted
rewards (Figure 6). Personally optimal policies are inferred from data when the expected sum of rewards is maximized
for individual drivers, given the progression of different states. To the best of our knowledge, the Markov Decision
Process and inverse reinforcement techniques have not been applied to naturalistic CV data to help extract human drivers'
personally optimal policies, with the aim of answering fundamental questions regarding individual driver decision
mechanisms. In a broader sense, the study outcomes will eventually assist drivers in making more informed and calmer
driving decisions, helping to improve the effectiveness of the transportation system.
As a proactive safety measure and a leading indicator of safety, we expand the concept of volatility to specific
locations by quantifying variability in instantaneous driving decisions at intersections. Traditionally, evaluation
of intersection safety has been largely reactive, based on historical crash frequency data. The developed conceptual
framework uses emerging CAV data and historical crash data for proactive identification of intersections with high
levels of variability in instantaneous driving behaviors prior to the occurrence of crashes. As an empirical contribution,
significant efforts went into assembling a unique database that integrates intersection crash and inventory data with
more than 65 million real-world Basic Safety Messages transmitted between connected vehicles and roadside units. Such
a unique empirical database facilitates rigorous analysis for making new discoveries. Additionally, the techniques used
to extract useful information from integrated CAV, crash, and inventory data provide useful knowledge for enhancing
proactive intersection safety.
Overall, by studying driving volatility from different perspectives, the team is creating a new analytical framework
to extract useful information from raw CAV data with the purpose of enhancing safety and fuel consumption efficiency.
As the project progresses further, apart from the new knowledge generated regarding instantaneous driving volatility
and key associated factors, newly developed algorithms will generate substantial new knowledge based on CAV data science.