Coping with Real-World Challenges in Real-Time Urban Traffic Control

In urban road networks, the use of real-time adaptive traffic signal control systems faces two typical challenges. First, various sources of uncertainty and disturbance can significantly degrade the accuracy of real-time flow predictions. Second, the optimization of vehicle flows must also give active attention to other transportation modes such as bus transit and pedestrian flows. In this paper, these challenges are investigated in the context of a recently implemented system called SURTRAC (Scalable URban TRAffic Control), which has now been running continuously in an actual urban environment for more than a year. SURTRAC takes a decentralized, schedule-driven approach to real-time traffic control and its design aims at urban (grid-like) networks with multiple, competing dominant flows that shift through the day. Motivated by observations of the system in operation, several strategies are proposed for strengthening the basic SURTRAC algorithm to better deal with real-world uncertainties and disruptive events, as well as multi-modal traffic demands. We evaluate the effectiveness of these strategies using both simulations and analysis of data collected from the pilot deployment. TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 2 INTRODUCTION Efforts to apply adaptive traffic signal control systems in urban environments face two typical challenges. The first challenge, which must be met in all application contexts, stems from their dependence on real-time information. To be effective in practice, a traffic control method requires quite accurate knowledge of traffic flows (1, 2). However, given that prediction of local traffic flows must be accomplished with a limited number of sensors (1, 3, 4), various sources of uncertainty can degrade the accuracy of flow prediction. The quality of detectors can be highly dependent on proper installation, and for some sensing technologies, performance is also influenced by dynamic environmental factors, e.g., weather and traffic conditions (5, 6). Mis-counting or over-counting can be caused by arbitrary lane-changing behavior of human drivers. In urban settings, there are additional uncertainties and disruptions. Road closures might trigger significant changes in traffic flow patterns. Temporary lane blockages might be caused by turning trucks, stopping buses, or onstreet parking in progress. Unsignalized intersections with side streets and alleys can contribute additional undetected flows. These nontrivial perturbations are often ignored in existing work. A second challenge in applying real-time adaptive signal control systems, which is more specific to urban road networks, is that passenger vehicles, transit, and pedestrians must share the right of way, and traffic signal control must give active attention to and be compatible with other travel modes. Improving performance of transit operations has attracted growing interest (7, 8, 9, 10, 11). Likewise, pedestrian flow is being given increasing priority in many urban environments (e.g., (12)) and there is increasing emphasis on multi-modal urban traffic environments (13, 14, 15), as compared to the vehicle-centric focus in conventional designs. In this paper, we consider these challenges in the context of the SURTRAC adaptive traffic signal control system (16, 17, 18), which was deployed on a nine-intersection network in the East Liberty region of Pittsburgh, PA in June 2012. SURTRAC (Scalable Urban Traffic Control) implements a totally decentralized approach to real-time adaptive signal control, adopting much the same conceptual framework as promoted by earlier online-planning approaches to adaptive signal control (2, 19, 20). SURTRAC is distinguished by its use of a novel formulation of the intersection scheduling problem (16), which enables intersections to compute near-optimal schedules over an extended planning horizon in real-time. Local schedulers at each intersection in a controlled road network operate asynchronously in rolling horizon fashion, communicating outflows to their neighbors to increase visibility of future incoming traffic and achieve coordinated behavior. When necessary, additional coordination mechanisms are applied to adjust local schedules to compensate for mis-coordination with its neighbors (17). An evaluation of the East Liberty deployment of SURTRAC performed shortly after installation indicated that significant performance improvement was achieved relative to the coordinatedactuated signal plans that were previously in place (18). Nonetheless, limitations due to various uncertainties, disruptions and multi-modal demands have become apparent as we have observed its continued operation, and over the past year a number of extensions have been investigated and incorporated to achieve more effective and stable operations, based on analysis and/or simulations using field data. Before discussing these strengthening strategies, we first summarize the SURTRAC approach to real-time adaptive traffic signal control and the initial pilot deployment. SURTRAC SCALABLE URBAN TRAFFIC CONTROL In SURTRAC, the traffic signal control problem is formulated as a decentralized, schedule-driven process (16, 17). Each intersection is controlled independently by a local scheduler, which mainTRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 3 tains a phase schedule that minimizes the total delay for vehicles traveling through the intersection and continually makes decisions to update the schedule according to a rolling horizon. The intersection scheduler also communicates outflow information implied by its current schedule to its neighbors, to extend visibility of incoming traffic and achieve network level coordination. At the individual intersection level, the ability to consider real-time (second-by-second) variability of traffic flows is made tractable by a novel formulation of online planning as a single machine scheduling problem (16). Key to this formulation is an aggregate representation of traffic flows as inflows. Each inflow includes a sequence of jobs, where a job contains vehicles in close proximity that have the right of way in a given phase, over a limited prediction horizon. Each job can be represented as a triple, i.e., . These job sequences preserve the non-uniform nature of real-time flows while providing a more efficient scheduling search space than traditional time-tick based search space formulations. The scheduling problem is to construct an optimal sequence of all jobs that preserves the ordering of jobs along each inflow. A given sequence dictates the order in which jobs will pass through the intersection and can be associated with an expected phase schedule that fully clears the jobs in the shortest possible time, subject to basic timing and safety constraints. The optimal sequence (schedule) is the one that incurs minimal delay for all vehicles. This scheduling problem is solved using a dynamic programming process. When operating within an urban road network, any local intersection control strategy without sufficiently long prediction horizon is susceptible to myopic decisions that look good locally but not globally. To reduce this possibility, network level coordination mechanisms are layered over SURTRAC’s basic schedule-driven intersection control strategy. As a basic protocol, each intersection sends its projected outflows to its direct neighbors (17). Given an intersection schedule, projected outflows to all exit roads are derived from models of current inflows and recent turning proportions at the intersection. Intuitively, the outflows of an intersection’s upstream neighbors become its predicted non-local inflows. The joint local and nonlocal inflows essentially increase the look-ahead horizon of an intersection, and due to a chaining effect, a sufficiently long horizon extension can incorporate non-local impacts from indirect upstream neighbors. This basic coordination protocol is quite similar to that previously utilized in (19). One difference is that we assume asynchronous coordination, so that temporary communication failures can be mostly ignored. Furthermore, the optimistic assumption that is made is that direct and indirect neighbors are trying to follow their schedules. Normally, the optimization capability of the base intersection control approach results in schedules that are quite stable, given enough jobs in the local observation and large jobs (platoons) in the local and non-local observation. It is also the case that minor changes in the schedules of neighbors can often be absorbed, if there is sufficient slack time between successive jobs. In practice, circumstances can cause schedules to change, in which cases mis-coordination can occur, especially for neighbor intersections that are very close together. To this end, additional coordination mechanisms are incorporated into SURTRAC for handling specific nontrivial miscoordination situations. One common inefficiency is caused by spillback that blocks the progress of traffic flow from an upstream intersection. The basic coordination protocol is augmented with a mechanism that acts to detect and prevent unnecessary spillback in advance of its occurrence by accelerating phase changes. Another source of mis-coordination is the tendency for the schedules of coordinating neighbors to oscillate due to small inconsistencies, which is handled by a second mechanism. Further description of these additional coordination mechanisms can be found in (17). TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 4 THE EAST LIBERTY DEPLOYMENT SITE The SURTRAC system was installed on a nine-intersection road network in the East Liberty neighborhood of Pittsburgh, PA, (intersections A-I shown in Fig. 1) for initial field testing in June 2012. Although the total network size is not large, this road network has several interesting characteristics. First, in contrast to the arterial settings that are typically studied in most traditional systems, this network has more of a grid-like character. It contains a triangle where three major streets (Penn Circle, Penn Avenue, and Highland Avenue) cross, with changing traffic flows throughout the day. The network also consists of compact road segments. The lengths of roads between intersections range from 90 to 500 feet with an average of 272 feet, imposing a nontrivial challenge for achieving effective coordination in a decentralized traffic control system. Finally, like most urban environments, there are a range of uncertainties and disruptions stemming from bus movements, commercial deliveries, construction projects and pedestrian traffic. Detection Capabilities Several information sources provide inputs to SURTRAC. Both phase and pedestrian call status are extracted continually from each intersection controller. For any given intersection, phase start and end time points are used to ensure synchronization of the local scheduler, and are also communicated to neighboring intersections. The pedestrian call status is only used locally. A given pedestrian call pedi corresponds to phase i in the controller, and remains pending until the pedestrian walk time is serviced in phase i. Detector groups (realized as video camera sensing zones) are placed on both entry and exit road segments for reporting two fundamental measures: traffic counts and occupancy of vehicles. For each entry link, a group of stop-bar detectors is placed near the intersection, and a group of advance detectors is placed sufficiently far away from the intersection; For each exit link, a group of exit detectors is placed near the intersection. The advance detectors associated with a given intersection are typically the exit detectors of its upstream neighbors, if the intersection is not a boundary node in the overall adaptive network. Further details can be found in (18). Since video detection is employed throughout the pilot test site, all of these detectors can be defined with no additional cost than would be required to support the detection requirement for vehicle-actuated logic and other traffic-responsive control techniques. Modeling Inflows Effective operation of the SURTRAC scheduler at a given intersection requires an accurate view of local vehicle inflows. To obtain inflows, we rely on a set of local link flow profiles. Each link flow profile describes the state of queuing and arriving vehicles along a particular link in a high-resolution prediction horizon. The main challenge is to achieve accurate short-term estimation through the limited detection available in real-world settings (1, 4). Link flow profiles are computed via an input-output technique (3), using advance and stopbar detectors on each entry link. On each road, all vehicles recorded by advance detectors are assumed to be moving at constant free-flow speed (v f ) when they are not stopped, and the predicted arrival time of each vehicle is shifted by the horizon L/v f , where L is the location of the advance detectors. Actual vehicle departures are recorded by stop-bar detectors. From time t0 to t1, each arriving vehicle is shifted by (t1− t0), and the queue q is adjusted by the difference of the number of predicted arrivals and the number of departures (20). For a given lane, a queue is discharged in TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 5 FIGURE 1 : The East Liberty pilot test site with nine signalized intersections (A-I). In urban environments, there are various issues faced by a real-time adaptive traffic signal control system. As a basic requirement, the system should be robust to significant changes in flow patterns over time (e.g., road closures, such as the closing of the bridge on South Highland Avenue from March to October 2013). In this paper, several issues are considered (examples noted on the map): (1) queue management for unaccounted traffic due to detection errors and mid-block traffic flows that are not covered by limited detection (detectors are placed near the intersections); (2) disruption management when the flow is temporarily blocked during green periods (e.g., bus stops with long dwell time periods near the stop bar); (3) flexible minor flowmanagement to achieve more effective and stable operations; and (4) more active pedestrian flow management to limit the average wait time for pedestrians, without interrupting the progressing of major vehicular flows. TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 6 a green phase at the saturation headway (shw) after the start-up lost time. These model parameters are estimated from historical flow data. Once equipped with link flow profiles, local inflows can be obtained by applying a road-tophase mapping (17), based on turning movement proportions at the intersection. This technique cannot detect queues beyond the advance detectors (1, 21). This does not cause a problem in our case, however, since non-local inflows are communicated from upstream neighbors in the road network. System Performance and Robustness A formal evaluation of the initial SURTRAC deployment, consisting of comparison of a series of “before” and “after” drive through runs, was carried out shortly after installation, and shown to yield significant performance improvement over the coordinated-actuated timing plans that preexisted at the pilot site (e.g., 25% reduction in average travel times, 40% reduction in average wait times) (18). SURTRAC has been running continuously at the test site since its initial installment, and continues to provide a live testbed for further development (see below). In addition, highresolution field data provides significant support for performance analysis (22). Over the course of SURTRAC’s ongoing deployment in East Liberty, traffic flow patterns have changed significantly. Between March 4 and October 23, 2013, the bridge on South Highland Avenue (location shown in Fig. 1) was closed for replacement, essentially cutting off one of the network’s three major flows for more than six months. A large portion of traffic (including all bus lines) on Highland Avenue was then forced to pass through intersection D, which is a pivotal intersection that services for most vehicles in this road network (even before the event). As evidence of the inherent robustness of the adaptive approach, Table 1 shows the flow difference for movements at intersection D before and after the closing of Highland bridge, both averaged over six weeks of collected data. Each movement is defined by the names of three intersections. As shown in the table, the total throughput in each day increased by 17.8% to 20.7% after the bridge closing. Individual flows CDE, EDC, EDA, ADE, EDF, FDE increased heavily (ranging from 14.9% to 48.0%), whereas CDF, ADF, and FDC actually decreased proportionally. After the change, CDE became the biggest movement, although EDF and EDA (which previously dominated) remained heavy. Although we have not been able to measure the effect that this change in flow patterns would have had on the pre-existing coordinated-actuated timing plans, we suspect that it would have resulted in a significant “aging” problem. In addition, Table 1 provides basic flow information for discussions in the following sections, since intersection D is a pivotal intersection for most vehicles, in this urban road network. STRENGTHENING STRATEGIES As we have observed operations of the SURTRAC deployment over the past year, we have detected specific sub-optimal behaviors due to various uncertainties, disruptive events and multi-modal demands. This has led us to investigate a set of strengthening strategies and incorporate them into the baseline system. These strategies are summarized and analyzed in the subsections below. Queue Management For any intersection entry link, the queue is a hidden state that changes over time. Given various forms of sensing uncertainty (e.g., detection errors or hidden flows contributed by mid-block roads), the estimated queue at any point can deviate quite significantly from the actual queue. To TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 7 TABLE 1 : Flow information at intersection D before and after the closing of Highland bridge. Most vehicles traveling in this road network pass through this intersection. Between March 4 and October 23, 2013, the bridge on South Highland Avenue (location shown in Fig. 1) was closed for replacement, essentially cutting off one of the network’s three major flows for more than six months. The road closure forced a large portion of traffic (including all bus lines) on Highland Avenue to pass through intersection D. Such significant changes in flow patterns often impose a challenge for the robustness of traffic signal control systems. (a) Vehicle counts of all movements before the bridge closing (1/21/2013 — 3/3/2013) EDF EDA EDC FDC FDE ADE ADF CDE CDF Total Mon 3,814 4,003 1,627 2,191 868 3,936 799 3,638 3,294 24,171 Tus 4,289 4,232 1,731 2,337 888 4,163 795 3,721 3,376 25,530 Wed 4,191 4,190 1,758 2,316 940 4,335 782 3,839 3,522 25,872 Thu 4,359 4,556 1,871 2,351 920 4,513 816 3,947 3,436 26,769 Fri 4,405 4,740 1,883 2,577 996 4,640 869 4,313 3,696 28,119 Sat 4,102 4,226 1,836 2,398 965 4,110 902 3,971 3,273 25,782 Sun 3,281 3,259 1,486 1,939 827 3,236 690 3,031 2,636 20,385 avg 4,063 4,172 1,742 2,301 915 4,133 807 3,780 3,319 25,232 (b) Vehicle counts of all movements after the bridge closing (3/10/2013 — 4/13/2013) EDF EDA EDC FDC FDE ADE ADF CDE CDF Total Mon 4,715 4,922 2,112 2,223 995 4,889 762 5,294 3,099 29,011 Tus 4,907 5,265 2,285 2,204 1,045 5,129 739 5,563 3,221 30,357 Wed 4,982 5,327 2,287 2,258 1,079 5,129 750 5,636 3,233 30,681 Thu 5,041 5,575 2,381 2,336 1,085 5,476 773 5,926 3,280 31,873 Fri 5,386 5,931 2,501 2,457 1,134 5,528 779 6,161 3,260 33,137 Sat 4,952 5,236 2,298 2,348 1,106 5,184 848 5,943 3,115 31,029 Sun 3,945 3,995 1,975 1,854 912 4,134 676 4,631 2,490 24,612 avg 4,847 5,179 2,263 2,240 1,051 5,067 761 5,593 3,100 30,100 (c) Percentage increase in vehicle counts after the bridge closing EDF EDA EDC FDC FDE ADE ADF CDE CDF Total Mon 23.6% 23.0% 29.8% 1.4% 14.7% 24.2% -4.6% 45.5% -5.9% 20.0% Tus 14.4% 24.4% 32.0% -5.7% 17.6% 23.2% -6.9% 49.5% -4.6% 18.9% Wed 18.9% 27.1% 30.1% -2.5% 14.8% 18.3% -4.1% 46.8% -8.2% 18.6% Thu 15.7% 22.4% 27.3% -0.6% 17.9% 21.4% -5.3% 50.1% -4.5% 19.1% Fri 22.3% 25.1% 32.8% -4.7% 13.9% 19.1% -10.3% 42.9% -11.8% 17.8% Sat 20.7% 23.9% 25.1% -2.1% 14.6% 26.2% -6.0% 49.6% -4.8% 20.4% Sun 20.2% 22.6% 32.9% -4.4% 10.3% 27.8% -2.1% 52.8% -5.5% 20.7% avg 19.3% 24.1% 29.9% -2.7% 14.9% 22.6% -5.7% 48.0% -6.6% 19.3% TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 8 cope with this circumstance and improve the accuracy of queue estimation, we have developed and integrated a more sophisticated queue management strategy. We first define some basic states that can be obtained from stop-bar and advance detectors (3, 20, 21, 23). The queue spillback state QS at each link n is recognized if the occupancy ratio on any advance detectors is nearly 100% for tQS seconds. Let ln,i be the lanes of each entry link n serviced in phase i. The queue blocking and queue clearance statesQBn,i andQCn,i are respectively obtained if the occupancy ratios of all stop-bar detectors on ln,i are nearly 100% for tQB seconds and 0% for tQC seconds. By default, tQS = tQB = 3 · shw, and tQC = 2 · shw. With these state definitions in mind, the basic queue length calibration is realized as follows. For each entry link, the predicted queue size q is a typical hidden state. A basic adjustment is q = max(0,q), and q = min(NSC,q), where NSC is the link storage capacity. The queue size is also adjusted from the measured states: q = NSC if QS and QB flags are identified, as the link is not serviced; and q = 0 if QC is obtained, and the link is serviced. The adjustment by spillback detection is especially important for video detection under high traffic volumes, where the count accuracy may deteriorate substantially due to the difficulty in sensing gaps between vehicles. Some properties can be obtained from these basic adjustments. If q is over-estimated, an ideal full clearance will adjust q to 0. For each intersection, it costs an extra tQC seconds. For its neighbors, the estimation error imposes some uncertainty on weights and execution durations of jobs in non-local inflows, and might cause some disturbance of the coordination between intersections. From the viewpoint of robust scheduling, any over-estimation might be considered as a buffer insertion (24), and since SURTRAC with the rolling horizon scheme is essentially doing on-line rescheduling continuously, it can effectively respond to such unexpected dynamics and provide good stability guarantees. If q is under-estimated, however, significant delay can occur from residual queues, and these residual queues will not be seen in subsequent cycles. The situation can become significantly worse if the queue starts to spill back to upstream intersections. Thus underestimation should normally be avoided, although arriving vehicles in the look-ahead horizon will sometimes alleviate most negative impacts on pure queue clearance if the current phase is further extended to accommodate arriving platoons. For each entry link n, we have the link flow rates f rarr n and f r dep n respectively from the groups of advance and stop-bar detectors. Each average flow rate is updated every 300 seconds. The arrival/departure ratio is defined as ADRation = f rarr n / f r dep n . The arrival-adjusting strategy is used to account for general arrival inaccuracy. As in (3), we assume that the group of stop-bar detectors can yield a reasonably accurate estimation of departing vehicles. If ADRatio < 1, some arriving vehicles are missed, and the numbers of queuing and arriving vehicles are under-estimated. Thus, when vehicles are detected at the advance detectors, the count is divided by ADRatio to reclaim those missing vehicles in the link arrival profile. In the East Liberty pilot test network, link arrival/depart ratios can be influenced by different uncertainties, e.g., on-street parking (on all three major streets), hidden flows from/to mid-block side streets (Baum Boulevard for Highland Avenue, Sheridan Avenue for Penn Avenue and Penn Circle), and detection errors. Fig. 2a gives the actual hourly ADRatio data for one day (from 6am to 21pm, July 2, 2012) for all entry links of the intersection of Penn Avenue and Penn Circle (the central intersection D of the pilot test site). This data shows that link arrival/depart ratios varied greatly, where the fluctuations indicate nontrivial uncertainty. TRB 2014 Annual Meeting Paper revised from original submittal. Xie, Smith, Barlow, and Chen 9 Link AD Link ED Link CD Link FD Hour L in k A rr iv a l/ D ep a tu re R a ti o