Proceedings of the 2009 Winter Simulation Conference M. D. Rossetti, R. R. Hill, B. Johansson, A. Dunkin and R. G. Ingalls, eds. A MARKOV PROCESS BASED DILEMMA ZONE PROTECTION ALGORITHM Pengfei Li Montasir M. Abbas, Via Department of Civil and Environmental Engineering 301-D Patton Hall Virginia Tech Blacksburg, VA 2406, USA Via Department of Civil and Environmental Engineering, 301-A Patton Hall Virginia Tech Blacksburg, VA 24061, USA ABSTRACT Dilemma zone (DZ) is an area at high-speed signalized intersections, where drivers can neither cross safely nor stop comfort- ably at the yellow onset. The dilemma zone problem is a leading cause for crashes at high-speed signalized intersections and is therefore a pressing issue. This paper presents a novel Markov-chain-based dilemma zone protection algorithm that con- siders the number of vehicles caught in DZ as a Markov process. The new algorithm can predict the numbers of vehicles in DZ and determine the best time to end the green so as to reduce the number of vehicles caught in DZ per hour. The algorithm was compared to the traditional green extension system and the results showed that the new algorithm was superior. 1 INTRODUCTION Dilemma zone (DZ) is an area at high-speed signalized intersections, where drivers are indecisive of stopping or crossing when presented with yellow indicator. The dilemma zone problem is a leading cause for rear-end collisions and red-light running. According to latest safety surveys, there are more than one million crashes at intersections a year and most of them occurred at signalized intersections(National Safety Council 2007). As a result, there is a strong need for real time algorithms that can predict the number of vehicles in DZ so as to make the correct decision of when to end the green phase. Traffic controllers typically provide a window of time with a variable length to end the green. If the green continued un- til its maximum time (known in traffic jargon as a max-out), the controller would be forced to end the green regardless of the number of vehicles in DZ. In this paper, we use the Markov process (MP) as a means to predict the number of vehicles in di- lemma zone in the near future, and evaluate the impact of ending the green at any given times before max-out. The Markov process has been proved capable of simulating a wide range of systems (Ross 2006). However, in the field of the traffic sig- nal control, few MP-based signal control applications were reported. In this paper, we considered the number of vehicles in DZ during the green as a discrete Markov variant. In light of this idea, the number of vehicles in DZ is predicted with the cur- rent number of vehicles in DZ and the state transition matrix. Specifically, using the transition matrix, the new algorithm first predicts the numbers of vehicles in DZ per hour for all the time steps before the maximum green. Then the algorithm seeks the best time to end the green to obtain the fewest hourly vehicles caught in DZ. If the best time is now, the algorithm will end the green immediately, if the best time is in future, the algorithm will extend the green one more step. This process is re- peated until either the algorithm considers now is the best time to end the green or the green phase reaches the maximum. In order to elaborate the new algorithm, we structured this paper into three parts: the first part includes the literature re- view on the MP’s applications to the transportation field, the explanation how the new MP-based algorithm works and the description how a VISSIM-based simulation environment was developed in order to test and evaluate the new algorithm; the second part is to evaluate the new algorithm using data from a high-speed signalized intersection in Christiansburg, VA. The geometry of that intersection and its dynamic traffic patterns were modeled into the simulation environment exactly to obtain a close-to-reality traffic network; the third part will analyze the safety performance of the new algorithm. 2 LITERATURE REVIEW The Markov process has been proved capable of simulating highly stochastic, non-linear traffic systems. A typical Markov control model is composed of four items: state space, control actions, states transition probabilities and reward matrix. The state space X: it is a Borel 2 Space and each element in the space is called state. In the context of the traffic system, the state space is defined to reflect traffic dynamics and it can be, for example, queue length, number of vehicles in DZ or control 2436 978-1-4244-5771-7/09/$26.00 ©2009 IEEE