The University of Tennessee, Knoxville

Transportation Engineering and Science Program

DVCV Research

Research activities revolve around "driving volatility" and they focused on: 1) examining instantaneous driving decisions and trip-level driving volatility at a microscopic level, 2) analyzing location-specific volatility, 3) investigating instantaneous driving decisions related personal optimal policies of drivers, 4) development of Markov Decision Process-based traffic flow algorithm, 5) application of gossip algorithms. The results are summarized.

The results from Markov Switching models reveal that acceleration and braking are two distinct regimes in a typical driving cycle, with braking showing substantially greater volatility (Figure 1). Compared to braking, the acceleration regime typically lasts longer. As expected, some drivers show greater volatility than others especially when driving on local and state routes. Importantly, when more objects surround a vehicle, the tendency is to accelerate even more if a driver is in acceleration regime, and to accelerate or lower the intensity of their braking if a driver is in the braking regime.

A novel framework is developed that helps in monitoring of acceleration and braking at a microscopic level, which can generate alerts and warnings, provided through advanced traveler information systems. The study shows that sequence of acceleration and braking events and vehicular jerk are key driving behavior factors to consider. Results revealed that the amount of warnings/alerts varies significantly between driver groups and it is highly associated with young drivers, new vehicles, two-seat vehicles, AM and PM rush hours, and commuter trips.

A fundamental understanding of instantaneous driving decisions is developed by examining two-dimensional instantaneous accelerations, i.e. longitudinal and lateral accelerations. Instantaneous driving volatility is captured by characterizing extreme driving events (Figure 3), and it clearly shows that driving behavior is strongly associated with driving contexts, whether driving on local roads or freeways. The results from empirical analysis revealed that extreme events identified in BSMs are strongly correlated with trip attributes, driver maneuvers, and driving contexts, which can help in real-time generation of warnings and alerts (Figure 4).

Using unique connected vehicle and crash data at intersections, we found that location-based volatility is positively associated with crash frequency (Figure 5). On average, a one-percent increase in the coefficient of variation in longitudinal accelerations/decelerations is associated with 0.25 increase in crash frequency for signalized intersections, while controlling for other factors. If many drivers behave in a volatile manner at a specific intersection, then such intersections can be identified before accidents happen. Of course, the reasons for volatile behaviors may be related to intersection and environmental conditions, vehicles' and drivers' conditions.

Aplication of Markov Decision Process (MDP) and Inverse Reinforcement Learning (IRL) provide valuable insights about driver-specific optimal policies that maximize the expected sum of rewards obtained by drivers in making instantaneous driving decisions (Figure 6). The MDP and IRL results show that when drivers have no objects surrounds them, acceleration is the prevalent personally optimal policy. With more surrounding objects, drivers choose to (or are constrained) to either brake or maintain constant speed as their personally optimal policy. Analysis of data shows that 65%, 53%, and 38% of drivers' personally optimal policies are to exit 3+, 2, and 1 surrounding object, respectively. In other words, with more surrounding objects, drivers try to avoid them as their personally optimal policy.

In related efforts, drivers in a group of vehicles were modeled as Markov agents and an MDP was solved at each time step for every pair of agents in the set. Each pair of automata negotiated between policy iterations to emulate the ability of human drivers to predict the actions of other drivers. The algorithm successfully replicated driver behavior in situations where no external forces cause vehicles to stop, and was robust in handling lead vehicles as well as middle vehicle stopping suddenly, as part of a platoon. The calibrated model successfully matched naturalistic driving data for a 30 second period.

To model the spread of behaviors in a network of proximate drivers, a particular gossip algorithm called the Voter model was modified and applied. In a simplified setting, each driver was assumed to be repelled by each other driver when making lane choices. The equilibrium solution for the Master equation was analytically obtained and compared with numerical solutions. Behavior of the model regarding dependence of the equilibrium distribution on the number of cars, time taken to reach an equilibrium condition from initial distributions, and the effect of influencers (such as police cars) on the lane adoption behavior were studied.


The University of Tennessee, Knoxville. Big Orange. Big Ideas.

Knoxville, Tennessee 37996 | 865-974-1000
The flagship campus of the University of Tennessee System