Modeling the Dynamics of Driver’s Dilemma Zone Perception Using Agent Based Modeling Techniques

Several research efforts investigated and targeted the issue of reducing dilemma zone (DZ) related crashes (where drivers upstream of intersections are uncertain about their decision to stop or go at the onset of yellow). One of the important unanswered questions in the literature is whether driver's perception of DZ changes individually as a function of their safe and unsafe past experience at similar situations. This paper investigates the use of agent-based methods in capturing the effect of driver's learning/dynamic perception of DZ. A driving simulator was used to collect driver behavior data. An actor-critic reinforcement learning algorithm was implemented to model the dynamic behavior of driver in dilemma zone. Fuzzy logic is used to partition traffic state variables and a reinforcement learning technique is used in policy calibration and update. The study results show a close matching between the driver's action from the driving simulator and the model output. The research reported here contributes to improved modeling of driver definition and behavior in dilemma zone, which will have significant impacts on the design of optimal control methods and the assessment of intersection safety. Moreover, it lays the groundwork for several subsequent simulator studies and scenario development in driving simulator to investigate the drivers' behavior at signalized intersections. Keywords: Dilemma zone; machine learning; dynamic perception; driving simulator Copyright c 2016 SERSC Language: en