Operation of Traffic Signal Systems in Oversaturated Conditions, Volume 2 – Final Report (2014)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Appendix A: Literature Review Summary of the Literature Review Development of strategies to handle oversaturated conditions is not a new topic of consideration. The research team found a wide variety of work in both diagnosis and estimation of oversaturation and control strategies and scenarios. In diagnosis and estimation we focused on reviewing techniques for measurement of queues and techniques for measuring the degree of saturation and surrogates for the degree of saturation. There has been quite a bit of work in the past on estimation of delay during oversaturated conditions and approaches for modeling oversaturated conditions. These efforts focused primarily on Highway Capacity Manual-type analysis and are thus not directly applicable to this project. Research on queue estimation is dominated by input-output modeling approaches. Input-Output methods are limited to estimating queues up to the point of the input detector, but not beyond this point. For arterial streets this requires installation of exit-side detection in order to measure a queue that is the full length of the link. Such detector installation can be cost-prohibitive. Methods for measuring the degree of saturation can identify the saturation level up to the point of saturation, but estimates of saturation above 1.0 have not been shown to be reliable, except for those estimates used by SCOOT and SCATS. The methods researched in this project provide some insight into the level of severity of the oversaturation (and queue length), which we believe is critical in identifying appropriate mitigation strategies. In the review of strategies and scenarios we looked at previous research on adaptive control systems, optimal control formulations, and various other approaches. Features of adaptive control systems and most other strategies are described in the literature in a qualitative manner. Concepts can be leveraged, but specific algorithms are not typically described quantitatively. Notably the features of SCOOT and SCATS that handle oversaturated conditions are mostly if…then type rules with thresholds that change some parameters or impose additional constraints on certain decision variables. Descriptions of these features are not accompanied by research indicating their ability to be effective in the real-world. Simulation studies evaluating the performance of these features could not be found either. Quantitative approaches (i.e. optimal control formulations) that were found in the literature all require information on traffic volumes, queue lengths, or both. Volume information is the most difficult information to obtain during oversaturation using state-of-the-practice detection systems, which makes most of the optimal formulations rendered ineffective. The literature review is divided into two focus areas. The first focus area is on the diagnosis of oversaturation and the second focus area is on strategies for mitigation of oversaturation. In several cases, the same reference material appears in both sections but the focus of discussion is on either the diagnosis or the strategy component of the particular research literature. Operation of traffic signal systems in oversaturated conditions Page A-1

Literature Review on Diagnosis of Oversaturated Conditions Although there has not been a significant amount of literature devoted to how to manage oversaturated traffic signal systems, there has been even less effort devoted to the identification of oversaturated conditions, with both spatial and temporal extent. Many management strategies as detailed in Part 2 of the literature review assume that arrival volumes can be somehow known and the oversaturated conditions can be accurately predicted based on this information. In the real world, this is simply not the case. Existing detection systems provide observations of flows at a fixed point on a link, which, during saturation, fail to provide the same accuracy of measurements. The techniques identified in this section of the literature review cover methods for estimating oversaturated conditions or measures of saturation at traffic signals. Using traffic data from signal systems to diagnose and identify oversaturation is sporadic and inconsistent in the literature, as we summarize in the following section. An oversaturated intersection movement (or lane group) can be defined as one in which the traffic demand exceeds the capacity. Based on this definition of oversaturation, the v/c ratio (v is demand and c capacity) has been used in theoretical formulae to identify whether an approach or movement is oversaturated. Analytically, the v/c ratio for each lane group can be estimated directly by dividing the demand flow rate by the capacity, by using the following equation: Eq. A-1 where, is the degree of saturation (v/c ratio) for lane group ; and are demand flow rate and capacity for lane group respectively; and are the flow ratio and the green ratio for lane group respectively. A lane group with is identified as oversaturated. For a single intersection with two competing demands, Gazis (1964) expanded this concept to diagnose oversaturation by testing the following inequality: Eq. A-2 where, and are arrival rates for two directions; and are saturation flow rates for two directions; L is the total lost time and C is the cycle length. Since Eq. A-2 only fits for intersections with fixed cycle length and lost time, Green (1967) modified it to Eq. A-3, which was called “absolute” oversaturation to deal with the situation when C is not a fixed value. ( ) i i i i i C g s v c vX      == iX i iv ic i ( )isv iC g      i 1>c v )(1 2 2 1 1 C L s q s q −>+ 1q 2q 1s 2s Operation of traffic signal systems in oversaturated conditions Page A-2

Eq. A-3 Direct application of the above models, however, is difficult. This concept might be extended from an intersection definition to a network definition by adding all of the demand flows in a network, but it is not clear how such as definition could be used in diagnosis and treatment of specific operational issues. Further difficulties in using such a definition arise because of the uncertainty of the capacity and saturation flow and due to the difficulty in measuring the arrival flow using current data collection systems (especially under congested situations – the very conditions that we are trying to identify). Because of this, other researchers have pursued alternative characterizations of oversaturation. Definitions of Congestion and Level of Service Longley (1968) identified two types of urban traffic congestion: a primary congestion that is caused by the development of queues at signalized intersections, and a secondary form of congestion that is caused by the blockage of unsignalized intersections by primary congested traffic. Longley presented a procedure for controlling congested controlled networks, when primary congestion is unavoidable. The basic premise of Longley’s procedure is to manage queues so that a minimum number of secondary intersections are blocked. This distinction between primary and secondary congestion and the relationship to signalized and unsignalized intersections is not useful in our work here as we are primarily concerned about signalized intersections. Pignataro et al. (1978) defined traffic operations in controlled network based on congestion levels as: a) Uncongested Operations: The situation where there is no significant queue formation. Traffic performance may range from very low demand per cycle to conditions where the demand is a significant fraction of the capacity value. Short queues may occasionally occur but do not last for any length of time (we presume “any length of time” was meant to indicate “not more than a few cycles”). b) Congested Operations: Refers to the entire range of traffic operations which may be experienced when traffic demand approaches or exceeds the road and/or intersection capacity. Furthermore, congested operations can be divided into two subcategories: saturated and over-saturated operations. i) Saturated Operations: a term that describes that range of congestion where queues form but their adverse effects on the traffic in terms of delay and/or stops are local. 1 2 2 1 1 >+ s q s q Operation of traffic signal systems in oversaturated conditions Page A-3

ii) Oversaturated Operations: is characterized as a situation where a queue exists and it has grown to the point where upstream traffic operations are adversely affected. In Pignataro’s taxonomy, the distinction between saturated and oversaturated operations is thus the effect of the queuing on upstream operations. It is an interesting proposal for the distinction of the terms of “saturation” and “oversaturation”, but no quantitative measures are provided in Pignataro et al. (1978) on how to identify such situations. NCHRP 3-38 defined congested (saturated) traffic conditions as follows: Local congestion: occurs when more vehicles face a cycle failure that does not result in damaging or excessive queues. Extended congestion: cycle failures repeat such that queues extend damagingly through upstream intersections causing the capacity of the upstream intersections to be reduced. Regional congestion: occurs when the queue from a critical intersection joins or influences the queue at upstream critical intersections. Intermittent congestion: occurs as a natural result of stochastic traffic arrivals. Even at light volumes timed using the Poisson method, one might expect cycle failures 5% of the time. Recurrent (cyclical) congestion: occurs at an intersection with insufficient capacity that occurs predictably as a result of foreseeable demand patterns. Prolonged congestion: congestion at intersections creates such inefficiencies that demand must fall below the reduced capacity for extended periods to permit overflow queues to clear. These different descriptions for congestion are qualitative. This taxonomy for congested conditions does little to help us map palliative strategies to saturated conditions and thus we explore more quantitative measures of previous research, such as queue measurement, travel delay/time, travel stops/speed, the flow-density/occupancy relationship, and green time utilization for diagnosis of oversaturation. Definitions Based on Queue Length Queuing is one of the most important indicators of oversaturation. Due to heavy demand or insufficient capacity, oversaturation has been characterized as the growth of queues. Gazis (1964) first characterized oversaturation that “a stopped queue cannot be completely dissipated during a green cycle”. This is also commonly known as a cycle failure or a phase failure. Later, many other researchers explored the relation between queue length and traffic conditions at intersections. For example, Kimber and Hollis (1979) proposed a model to describe the relationship between queue length and intensity, which is described as the demand-to-capacity ratio. Akcelik (1981, Operation of traffic signal systems in oversaturated conditions Page A-4

1988) investigated the queue-based intersection delay under oversaturated conditions. Abu-Lebdeh and Benekohal (2003) defined oversaturation such that “traffic queues persist from cycle to cycle either due to insufficient green splits or because of blockage”. These efforts contribute to the understanding of oversaturation which have been summarized by Roess et al. (2004) in his textbook Traffic Engineering that “the oversaturated environment is characterized by unstable queues that tend to expand over time with potential of physically blocking intersections (blockage, spillback), thus slowing queue discharge rates…”. Thus it is evident that queues (longer than a link length) indicate the oversaturated situation. However, for some cases, even when the maximum queue length is shorter than the link length, frequent residual queues could be considered as oversaturation. Such queuing could be due to insufficient green splits, downstream blockages, and/or heavy pedestrian flows. The FHWA study on Signal Timing Under Saturated Conditions (Denney et al., 2008) defined different types of traffic conditions based on queue information (associated with an individual intersection). 1) Light traffic. Characterized by the expectation of minimized cycle failure. In such conditions, the signal can fully serve arrival queues and cycle failures are expected to be infrequent at less than 25% of all cycles. 2) Moderate traffic. Characterized by the expectation of fair operation. At these intersections, drivers expect the operation to be fair, which means that the degree of saturation on each approach is approximately the same. Some cycle failures are to be expected but do not necessarily violate the expectations of motorists. No approaches have queues that are growing disproportionately to other approaches. 3) Heavy traffic. Characterized by frequent cycle failures, but with a residual queue that ebbs and flows without growing uncontrollably. The flows exceed capacity nearly half of the time due to stochastic arrivals. 4) Oversaturated operation. Characterized either by excessive residual queues that grow without control, therefore causing more widespread damage to the operation of a network. Operation of traffic signal systems in oversaturated conditions Page A-5

Figure A-1. Saturation and residual queues (Source: Denney et al., 2008) Figure A-1 graphically illustrates the relationship between such traffic conditions (including light, moderate, heavy and oversaturated) and the residual queue length. When average demand equals capacity, the percent of cycles with residual queues will be 100%; while with demand exceeding capacity, a growing residual queue is inevitable. Similarly to the NCHRP 3-38 taxonomy, these qualitative descriptors are useful to guide thinking about oversaturation, but cannot be used directly to map palliative strategies to oversaturated conditions. Queue length as well as residual queuing (queues that persist after switching from green to red) for a given phase are the preferred indicators for diagnosis of oversaturation. However, accurate queue estimation typically depends on the arrival flow information, (i.e. input demand), which requires installation of upstream detectors. For most agency-standard detection technology deployments at intersections, this requirement cannot be satisfied. Typically, advance detectors are installed within a few hundred feet from the stop line, depending on the speed of approaching traffic, for vehicle actuation purposes on high-speed approaches (dilemma zone protection). Traditional input-output approaches to estimating queue length will not work appropriately if the queue length spills over the advance detector location. Therefore, either additional upstream detectors need to be installed or alternative methods to estimate queue length based on a typical layout of detector placements needs to be developed (or use of alternative measures to identify queue length). Measures of Oversaturation Based on Delay/Stops/Speed/Travel Time Along with the presence of long queues during oversaturated time periods, overflow delay (as part of the queuing delay), becomes one of the major components of total delay at an intersection. Figure A-2 (Dion et al., 2004) compares typical delay accumulation curves for both under-saturated and over-saturated conditions. Figure A-2 also shows that the overflow delay Operation of traffic signal systems in oversaturated conditions Page A-6

becomes over time, a more and more significant component of over-all intersection delay. How this delay accumulates is extremely difficult to measure. Figure A-2. Idealized cumulative arrivals and departures for under and oversaturated conditions (Source: Dion, F. et al., 2004) Many different delay models for oversaturated intersections have been proposed and compared with traditional delay models which were designed to measure delay during undersaturated conditions (Fambro & Rouphail, 1997; Kang, 2000, Dion et al., 2004; Benekohal & Kim, 2005; Kim & Benekohal, 2005;). Figure A-3 compares Webster’s random delay model, which only applies for saturation values below 1, and theoretical overflow delay models which can extend to v/c ratios above 1.0 (Roess et al., 2004). As illustrated, the estimated delays at oversaturated intersections significantly increase compared with random delays of undersaturated intersections. Most research has hypothesized that this growth rate in the average delay per vehicle varies linearly as the v/c ratio increases. These methods provide models for estimating delays for v/c ratios higher than 1, but they have little value in practice because of the difficulties in the measurement of the input components of the model – particularly the demand volumes. Any placement of a fixed detection location, at some point, cannot capture the demand volume. This is true because as the demand grows, the point of arrival to the back of the queue continues to grow further upstream. Thus, to meet our goals to estimate oversaturated conditions with real-world detection systems, we cannot utilize delay as the primary estimator of oversaturation. Operation of traffic signal systems in oversaturated conditions Page A-7

Figure A-3. Random and overflow delay models compared (ource: Hurdle, Roess, et al., 2004) Other performance measures, such as speed (space mean and other various forms), number of stops, and travel time also change significantly during oversaturated conditions. For example, vehicles repetitively accelerate and decelerate in congested conditions leading to a significant increase of number of stops. Travel times also increase significantly as speed decreases. Speed in and of itself is not sufficient to identify oversaturation in arterial traffic systems because detected speed on arterials drops to zero when any queue forms over a detector when the traffic signal indication is red. Most models which are designed to estimate speed, stops, and travel time performance during oversaturation are based on measuring queues at intersections. For example, the model developed by Cronje (1983a, 1983b, 1986) is based on the expected queue size at the beginning of each cycle. Therefore, we conclude that queue estimation is the crucial component in the diagnosis of oversaturated conditions. Flow-Occupancy Diagram/Fundamental Diagram Most of the performance measurements mentioned above (queue, delay, speed, travel time, etc.) can be estimated based on two basic measures, i.e. flow and occupancy of fixed-location detectors. Density of traffic on an arterial can be estimated by occupancy at a fixed location. The relationship between these two measures, i.e. the flow-occupancy diagram or the fundamental diagram, can be used to indicate traffic conditions. Different conditions (undersaturated, saturated and oversaturated) are located at different areas of the fundamental diagram. Recently, some related studies have focused on how to use such information to estimate arterial traffic system performance. For example, Perrin et al. (2002) used occupancy data to estimate the v/c ratio and level of service (LOS); Sharma et al. (2007) combined detector and signal phase information to estimate vehicle delay and queue measurements; and Hallenbeck et al. (2008) used stop-bar detector data combined with signal state data to estimate arterial traffic conditions (congestion) from the perspective of the fundamental flow-occupancy diagram. Figure A-4 (Hallenbeck et al., 2008) indicates that oversaturated conditions can correspond to the right-hand side of the Operation of traffic signal systems in oversaturated conditions Page A-8

volume-occupancy diagram. Although most of the data in these studies are generated from simulations, these general ideas will have a contribution on our research on the identification of oversaturation. Figure A-4. A comparison of occupancy percentage and corresponding arterial congestion. Different colors indicate different congestion levels. (Source: Hallenbeck et al., 2008) Utilization of Green Time Another measure that has been used to identify saturation is green utilization. If a traffic signal phase is oversaturated, vehicles will continue to discharge at the saturation flow rate until the end of the effective green. When the traffic phase is undersaturated, there is some proportion of the green time (or green split) that is not used by traffic and is thus available to other phases when the intersection is operated in actuated mode. Under fixed-time control, this extra time is identified as green time with no associated occupancy on traffic detectors that serve that phase. The proportion between the used green time and the green phase, i.e. green utilization, can be treated as an indicator of the saturation level of a particular traffic phase (Gettman, et al, 2007). This method, however, does not serve to estimate the degree to which a certain traffic phase is oversaturated (i.e. how long the queue actually is). The indicator simply identifies that there is not enough green time to service all of the traffic demand, but it does not indicate how much green time would be necessary to clear the queue. A similar concept is also proposed by Smaglik et al. (2007). This research uses the proportion of vehicles arriving on green to identify the traffic arrival type. As presented in Figure A-5, by using setback detectors, vehicle arrivals are collected and tabulated into cycle-by-cycle bins; and Eq. A-4 is used to calculate the proportion of vehicles arriving on green, which is the proportion between the number of vehicles arriving on green and the total number of vehicles arriving within a cycle. Applying the formula provided by HCM 2000 (Eq. A-5), the platoon ratio can be estimated as well as the arrival type and the quality of progression (Figure A-6). Operation of traffic signal systems in oversaturated conditions Page A-9

Eq. A-4 where, is the proportion of vehicle arriving on green; is the number of vehicles arriving on green; and is the number of vehicles arriving on red. Eq. A-5 where, is platoon ratio; is cycle length; and is green interval. Figure A-5. Calculation of proportion of vehicles arriving on green (P): (a) binning of vehicle arrivals during different indications; (b) cycle by cycle binning (Source: Smaglik et al., 2007) Figure A-6. Relationship between arrival type and platoon ratio (Source: HCM 2000)         + = gr g NN N P P gN rN       = g CPPp pP C g Operation of traffic signal systems in oversaturated conditions Page A-10

It is necessary to point out that the high green utilization of a given phase is not always necessarily due to oversaturation. In some cases, good coordination design can generate high utilization on a phase by synchronizing the arrival of the incoming platoon with the start of green. Green utilization is not a complete picture of latent demand in much the same way that input-output methods for queue estimation fail to measure the stored demand. Green utilization can only indicate that there is more demand for a given phase, but not by how much. The State of the Practice in Diagnosis of Oversaturated Conditions All the measures described above can be used as indicators to diagnose oversaturation with varying levels of effectiveness. Practical applications, however, are restricted by the existence of data collection systems, appropriate field hardware and communications systems, and operational support. At best, the use of data from signal systems to diagnose and identify oversaturation is sporadic and inconsistent. In this section we summarize some capabilities of existing traffic signal systems for identification of oversaturated conditions. Flow Profile Estimation in the SCOOT Adaptive Traffic Control System The SCOOT (The Split, Cycle and Offset Optimization Technique) (Hunt et al., 1981) adaptive traffic control system has capabilities and features designed to deal with oversaturation (Martin, 2006). As an adaptive system, SCOOT depends on good detection data so that it can respond to changes in flow. Detectors are normally required on every link. Their locations are important and they are usually positioned at the upstream end of the approach link (Other types of detectors such as stop-bar detectors can also be used to improve the traffic state estimation). These detectors (referred to as exit detectors) provide arrival flow information (i.e. input) to the SCOOT traffic state estimation module so that measures such as queue length can be determined (Figure A-7). To diagnose oversaturation, SCOOT estimates the degree of saturation by measuring the flow upstream of the stop-line and the “online saturation occupancy” measurement in lieu of a fixed saturation flow rate. The online saturation occupancy of a link is “the rate at which the queue modeled by SCOOT discharges from stop-line” (Bretherton and Bowen, 1990). The saturation percentage (%SAT) in SCOOT is then estimated by following equation (Martin, 2006): Eq. A-6 where: g is the green time, q is the combination of flow and occupancy; STOC is maximum outflow rate of a queue over the stop line. The unit of q and STOC are LPU (link profile unit) and LPUs/sec, respectively. LPU is an internal unit defined in SCOOT, where one vehicle is equivalent to approximately 17 LPUs, and a saturation flow of 2000 vehicles/hour is equivalent to a SCOOT saturation occupancy (STOC) of 10 LPUs/Second. Oversaturated conditions are indicated by high %SAT (larger than 100%). )*(% gSTOC qSAT = Operation of traffic signal systems in oversaturated conditions Page A-11

Figure A-7. Queue estimation in SCOOT (Source: Martin, 2006) Degree of Saturation Measurements from the SCATS Adaptive Traffic Control System Unlike SCOOT, the SCATS (Sydney Coordinated Adaptive Traffic System) (Sims & Dobinson, 1980) adaptive traffic control system only uses stop-bar detectors as “its primary source of data and so has no information about near-future arrivals” (Dineen, M., 2000). Compared with SCOOT, Dion and Yagar (1996) point out that “whereas SCATS is reactive to short-term traffic fluctuations it does not have SCOOT’s predictive capability and is therefore less proactive”. SCATS uses the Degree of Saturation (DS), i.e. the “ratio of the effectively used green time to the total available green time”, to diagnose the level of congestion for a given traffic phase. DS is estimated by: Eq. A-7 where: NF is a bias factor (weighting factor); g is green time; T is Total non-occupancy (space) time; t is space time which is unavoidably associated with each vehicle; r = remaining (or unused) phase time; and g’ is effectively used green time (Dineen, M., 2000). This approach requires a short detector and very accurate measurement of occupancy since during oversaturation, the gap times between vehicles can be very small. The ratio can be > 1 only when T < t*n, which indicates that there is less actual “space time” remaining than what would be available if the phase served the saturation flow rate for the entire green time. This is not possible to occur unless r=0 (there is no unused phase time since the phase is forced off (maxed-out) and did not gap-out). Figure A-8 shows examples of degree of saturation profile from the SCATS (Martin, 2006). In the figure, the measurements of degree of saturation are indicated as the vertical lines (green) for each traffic cycle. A traffic phase is estimated to be oversaturated when the green vertical line exceeds the horizontal (red) line for each phase. Notice in the bottom right corner of the figure rg gNF rg ntTgNFDS + ′ = + −− = ][)]*([ Operation of traffic signal systems in oversaturated conditions Page A-12

that the cycle time of this intersection is gradually increased (and decreased at times) in conjunction with the measurement(s) of oversaturation on phases at this intersection. Notice also how noisy the cycle-by-cycle estimates of DS are over time, particularly for left turn phases (1, 3, 5, and 7). Figure A-8. Degree of saturation estimates over time from the SCATS adaptive traffic control system (Source: Martin, 2006) Flow Profile Estimation in the OPAC Adaptive Traffic Control System The OPAC (Optimized Policies for Adaptive Control) (Gartner, 1983) adaptive control system implements a rolling horizon strategy to optimize signal timing by predicting flow profiles as well as predicting queues that will result due to various potential changes to the signal timings (Gartner et al., 2002). Based on loop detectors placed upstream of each approach, a flow profile is developed for each phase (Figure A-9). The head of the profile is actual counts from upstream link detectors; and the tail of the profile is projected for the near future using a simple model consisting of a moving average of all past arrivals on the approach. A simple input-output method is applied by OPAC to predict queue information, i.e. the estimated queue length is the sum of the initial queue length plus the difference of arrival and departure of each interval in the stage. Clearly, the accuracy of prediction highly depends on the location of upstream detectors which provide the arrival information. Placing the detectors well upstream of the intersection (10 to 15s travel time) will allow actual arrival information to be used for the head period; thereby increasing the accuracy of following predictions. However, such method cannot deal with oversaturated conditions as arrival information is missing once queue spillovers upstream detector. There is little discussion in available literature of any particular features of Operation of traffic signal systems in oversaturated conditions Page A-13

OPAC or algorithm enhancements that allow OPAC to estimate oversaturation or adjust its flow profile estimates during congested conditions. Figure A-9. Flow profile of OPAC (Source: Presented by Nathan H. Gartner on the workshop of Adaptive Traffic Signal Control Systems, TRB, Jan. 7, 2001, Washington D.C.) Oversaturated Conditions Measurement in the RHODES Adaptive Traffic Control System The RHODES (Real-time Hierarchical Optimized Distributed Effective System) (Mirchandani & Head, 2001) adaptive traffic control system uses the output of the detectors on the approach of each upstream intersection, together with the traffic state and planned phase timings from the upstream signals to predict future arrivals at the intersection. These inputs are used in the “QUEUE” algorithm (Head, 1995), a simple input-output estimation procedure to determine the queue length of traffic for each control phase. In detail, the queue at cycle , , is equal to the residual queue at , , plus the predicted arrivals, , and minus the estimated departures, , using a given queue discharge rate (Eq. A-8). Eq. A-8 To keep biases from entering into the estimates, RHODES identifies some certain epochs that the queue length is zero on the basis of the information from stop-bar presence detector (Head, 1995). RHODES handles queue estimates that extend beyond the location of upstream detectors by allowing a traffic agency operator to tune the saturation flow rates of the queue discharge process 1t )( 1tq 0t )( 0tq ),( 01 tta ),( 01 ttd ( ) ( ) ),(),( 010101 ttdttatqtq −+= Operation of traffic signal systems in oversaturated conditions Page A-14

by time of day. As such, during rush hours the estimated queues can be made to dissipate more slowly on oversaturated approaches allowing the baseline adaptive algorithms to operate as if the intersection were still under-saturated on all approaches. RHODES also includes the ability to weight the delay estimates on one traffic phase higher than the delay on other traffic phases at the same intersection (and adjust the weights by time of day). This feature also helps to overcome the limitations of the queue estimation algorithm during oversaturated conditions by allowing certain phases that are known to have more traffic to be serviced with longer green times. In Pinellas County, FL, the RHODES installation on SR-60 has shown vast improvements during heavy arterial flow scenarios utilizing these key enhancements. Although these adjustments can improve the effectiveness of the adaptive algorithms during oversaturation, RHODES, along with most other model-based adaptive systems, recommends placement of additional detectors at the upstream ends of links (i.e. exit detection) similar to the detection requirements for SCOOT. Extending detection placement in this way allows estimation of queues that extend to the entire link length. Of course, such additional detection comes with an associated increase in total system costs. Green Utilization Measures in the ACS-Lite Adaptive Traffic Control System Similar to the SCATS adaptive signal control system, ACS-Lite (Adaptive Control Software Lite) (Luyanda et al., 2003) uses phase utilization as it’s indicator of degree of saturation for each traffic phase. Phase utilization is calculated based on second-by-second occupancy data from stop-bar detectors. Combining phase timing information with volume/occupancy data, the average used green is calculated (Figure A-10). Phase utilization is then used to compare the need for additional green time and availability of excess green time of each phase on the controller. In Figure A-11, Phase 1 is shown to be saturated (red color) since the entire phase split is used over the last seven cycles and occupied with traffic the entire time. Operation of traffic signal systems in oversaturated conditions Page A-15

Figure A-10. Phase utilization measures from the ACS-Lite adaptive control system (Source: Gettman, 2005) Figure A-11. Phase utilization measures from the ACS-Lite adaptive control system (Source: Gettman, 2005) ACS-Lite currently contains no explicit handling of oversaturation or adjustment factors to approximate oversaturation on a particular traffic phase. Green utilization is scaled from 0-100%. Like SCOOT, and to some extent OPAC, ACS-Lite includes the measurement of a cyclic flow profile but it does not use this information to estimate oversaturation. Occupancy data from upstream advance detectors (~300ft from the stop bar) are used to measure when the traffic on an approach arrives to the signal during red or green. ACS-Lite includes no explicit handling of oversaturated conditions, but visualization screens can clearly indicate when an approach is saturated as the occupancy percentage typically is not reduced to less than 100% during the cycle. Operation of traffic signal systems in oversaturated conditions Page A-16

Queue Estimation in the German ACS/BALANCE/MOTION Adaptive Control Systems ACS, BALANCE, and MOTION are all adaptive control systems developed in Germany that are based on the estimation of queues to recognize the traffic state (Mueck, 2005). Some progress has been made in recent years in these systems for estimating queues that grow past the detector location. A typical European traffic management agency places detectors 50-100ft upstream of the stop-line (Mueck, 2002). Mueck’s algorithm for estimating queue length is based on the measurement of the so-called fill-up time. The fill-up time is “from the beginning of the red time of a signal until continuous occupancy of a detector” (Mueck, 2002). The fill-up time is found to be correlated to the level of congestion on an approach. Essentially, this algorithm identifies that when there is a residual queue that cannot be discharged fully during the green time, the fill-up time to 100% occupancy on the detector is much faster than the fill-up time when the queue is fully discharged and another platoon (or turning traffic from side streets) arrives at the stop bar. When the fill-up time “falls short depending on the distance between detector and stop-line” (Mueck, 2002), it is determined that a residual queue is forming and that the approach is now “congested”, such that: Eq. A-9 where is the measured fill-up time; and is the reference value. Then, this method estimates the residual queue length by estimating the growth of the surrogate “congestion characteristic” (for cycle n) using an exponential smoothing method: Eq. A-10 Empirical data collected in Germany shows a roughly linear relationship between the maximum back-up length and the congestion characteristic (Figure A-12).    > ≤ = 0 0 0 1 dtdt dtdt δ dt 0dt nδ 1)1( −−+= nnn aa δδδ δδ Operation of traffic signal systems in oversaturated conditions Page A-17

Figure A-12. Maximum back-up length (veh) over smoothed congestion characteristic (Source: Mueck, 2002) Then, the back of queue length can be estimated by Eq. A-11 based on the congestion characteristic estimate. Eq. A-11 where m is a multiplicative gradient factor, which should be determined for each detector individually based on its location from the stop bar and its level of sensitivity. Figure A-13. Evaluation of fill-up time queue estimation (Source: Mueck, 2002) Figure A-13 shows remarkable results for estimating queues that extend up to 5-10 times further upstream from the actual detector location (Mueck, 2002). However, this heuristic approach is based upon the premise that arrival rate of traffic flow is constant within a cycle; therefore the measured fill-up time can be proportionally inverse to the queue build-up time. Such an approach nn mL δ= Operation of traffic signal systems in oversaturated conditions Page A-18

could be problematic when the arrival rate of traffic flow fluctuates greatly, which is not unusual with the gating effect of an upstream traffic signal. Link Load in the TUC Adaptive Control System The TUC (Traffic-responsive Urban Control) adaptive traffic control system (Diakaki et al., 2003; Dinopoulou et al., 2005; Kosmatopoulos, et al., 2006) was developed to provide coordinated, traffic-responsive control in large-scale urban networks, even in cases of saturated traffic conditions. TUC includes four distinct control modules: split control, cycle control, offset control and public transport priority. The system is dependent upon real-time measurement, i.e. average number of vehicles within each network link z over a cycle; and occupancy measurement measured by a traditional detector in each link z is utilized to estimate by a suitable function: Eq. A-12 where is the maximum number of vehicles that can be in link z; is the proportion between the detector distance from the stop line and the whole link length (an indicator of detector location); and f is defined as a suitable function. The available literature on TUC does not describe such an important function in detail. The proportion of and , is defined as the link load in TUC and actually indicates the degree of saturation of link z. NCHRP 3-79 NCHRP 3-79 provided a comprehensive summary of the current state of the practice on urban street performance measurements including delay, queue, travel speed, and travel time. Most of their results are concluded in their interim reports (Bonneson, 2005); and some are published in journals (Sharma et al., 2007; Sharma & Bullock, 2008). The 3-79 project proposed two techniques for measuring queue lengths and corresponding saturation level of an arterial traffic link. Input-Output and Hybrid Techniques for Delay and Queue Length Measurement Input-output (I-O) and Hybrid input-output (Hybrid), are proposed for measuring queue lengths and delays. The basic concepts are similar; i.e., by utilizing the information of the arrival and discharge flow rates, delay and queue length at a signalized intersection can be quantified using input-output analysis (Figure A-14). The I-O technique only requires one upstream detector, which is used to measure the arrival flow rate. Combining this information with the signal phase status, estimated turn proportions, and saturation flow rate, this technique can develop a queue polygon and estimate delay for the cycle. The hybrid technique extends this approach to utilize two sensors located upstream and at the stop-line to directly measure the arrival and departure rates, respectively. Figure A-15 describes the process of both techniques (Sharma et al., 2007). zx zo zx ),( max, zz z z ofx x λ= max,zx zλ zx max,zx Operation of traffic signal systems in oversaturated conditions Page A-19

The accuracy of both techniques highly depends on the input flow measured by the upstream sensor and thus the location of such a sensor is crucial. Generally, this sensor should be located sufficiently far back in order to avoid frequent queue spillback, but at the same time, the sensor should not be located so distant that “driveway activity between the sensor and the stop line seriously degrades the accuracy of the predicted arrival flow profile” (Bonneson, 2005). Figure A-14. Delay & Queue Polygon using Input-output Process (Source: Bonneson, 2005) Operation of traffic signal systems in oversaturated conditions Page A-20

Figure A-15. Process (a) I-O; (b) Hybrid (Source: Sharma et al., 2007) Non-Intrusive Detection for Queue Length Measurement The second technology proposed by the NCHRP 3-79 research team to measure queue length is video detection. This approach is based on the use of a series of video detection zones, which monitor a segment of an intersection link, with coverage of all approach lanes. Queue length is indicated by the number of detection zones that are fully occupied during the red interval. Delay can then be estimated from this information on the length of the queue (Bonneson, 2005). The advantage of such a technique is that “they can monitor a length of roadway”; but it is acknowledged that the accuracy of the video detection technology degrades with increasing distance from the sensor (Bonneson, 2005). The placement of most video detectors that are used for phase actuation is such that queues could not be sensed to extend more than 300-400ft from the stop bar. Bus Probe for Delay and Running Time Measurement This technique is based on “the use of a transit-based automatic vehicle location (AVL) system for running time and overall travel time measurement” (Bonneson, 2005). The AVL data can provide the arrival time of buses at key points along major urban streets which can then be Operation of traffic signal systems in oversaturated conditions Page A-21

converted into segment-based travel time. An estimate of delay, and perhaps the length of standing queues could then be made by comparing the running time with the free-flow running time. One major issue related to this technique is the frequency of bus stops since this frequency will affect the accuracy of travel-time estimation. Traffic Flow Characteristics for Running Time Measurement This proposed technique uses measured traffic volume to estimate average running speed according to a calibrated relationship of speed-volume. The running speed is then converted to running time by dividing speed into segment length. This technique requires separate detectors for each traffic lane to accurately measure the traffic volume as shown below in Figure A-16. Figure A-16. Framework for estimating running speed using measured traffic flow characteristics (Source: Bonneson, 2005) Operation of traffic signal systems in oversaturated conditions Page A-22

Control Strategies for Oversaturated Conditions Control strategies for managing traffic congestion in oversaturated conditions seek to improve facility capacity or reduce facility demand. Enhancements to the roadway capacity can be achieved by increasing the physical capacity of the roadway system or by maximizing the operational capacity by improving the traffic signal controls. On the other hand, demand reduction might be achieved through restrictive control procedures such as metering or by using macro-level measures aimed at reducing vehicle use or influencing travelers to make significant modifications to their travel mode, departure time, route or destination (Quinn, 1992). This project does not address demand side management strategies such as influencing mode choice or providing ATIS routing messages and, as such, we will not review such strategies. Several signal control strategies specifically designed for management of oversaturated conditions have been described in the literature. These strategies span a range of characteristics from being static or dynamic, reactive and proactive, having single objectives or multiple objectives, and applied for isolated intersections, closely-spaced intersections, diamond interchanges, arterial networks, and grids. This section focuses on review and critique of the algorithms and strategies that have been previously developed and described in openly-available literature. Strategy Taxonomies for Managing Urban Congestion Taxonomies for congestion management strategies have not been explored in much detail in the past. Most of the research that was reviewed focused on specific types of applications without a detailed effort devoted to framing the context of the problem. Pignataro et al. (1978) classified strategies designed to address urban traffic congestion into: • Signalized approaches: minimally responsive policies • Signalized approaches: highly responsive policies • Non-signal-related approaches: demand management, transit operations, etc. Huddart and Wright (1989) classified congestion prevention treatments into: • Static protection measures • Dynamic protection measures • Traffic input control (otherwise known as demand management) measures. Pignataro et al. (1978) classified measures designed to tackle urban traffic congestion into minimal response, highly responsive signalized policies, and non-signal treatments in a signalized network. These methods are explained as follows: 1. Signalized control measures - Minimal response methods: (a) Intersection: Cycle length, Block length, splits, and extra phases Operation of traffic signal systems in oversaturated conditions Page A-23

(b) System: Phases adjusted for progression, equity offsets, splits allocated to available queue storage. - Highly responsive methods: (a) Intersection: Maximum queue policy (b) System: Accommodate queue spreading from intersection 2. Non-signalized treatment - Regulatory measures: (a) Enforcement of exacerbating conditions (e.g., double parking) (b) Prohibition of exacerbating traffic movements (e.g., parking, turning) - Operations measures: (a) Enhancement of turning facilities (e.g., left bays, right bays, dual turning lanes, right-turn-on-red ) (b) Enhancement of lane arrangements (e.g., one-way street, reversible lane) (c) Strategies for handling disruption (e.g., pedestrian, bus stop, parking/ no-parking, mid block ) No other attempts to categorize types of strategies were found in the literature. Neither of these taxonomies is exhaustive or particularly systematic in its approach to categorizing strategies. This lack of a good taxonomy of types of strategies mirrors the lack of specificity in the qualitative approaches applied to the definition of types of congestion or oversaturation. In the following sections, we detail strategies for managing oversaturated conditions found in the literature in order from simple to more complex. This organization of strategies generally proceeds from approaches that deal with individual approaches or intersections to strategies that address network conditions. Switching of Green Time Dunne and Potts (1964) developed an algorithm for controlling an isolated intersection based on a closed-loop feedback control concept. Green phases were changed according to the variation in traffic demand. The concept was further improved by Gazis and Potts (1965) for traffic signal control in oversaturated conditions. Gazis and Potts’ control philosophy was to minimize total system delay by maximizing intersection productivity. In this approach, a maximum green time is provided to the approach with the higher saturation flow until the queue on that approach is dissipated, while other approaches receive a minimum green time. The control decisions fluctuate only between these two values, minimum green or maximum green, and the cycle time remains constant. This approach assumes that the traffic demand can be known, and that the demand does not change due to the control decisions (i.e. drivers do not re-route). Michalopoulos (1975) used a reservoir analogy to improve the policy developed by Gazis by developing an algorithm that determines what the “switch-over” points should be. The “switch-over” concept in this case refers to the points in time at which the green duration of a certain approach should be changed from maximum green to minimum green allowing other Operation of traffic signal systems in oversaturated conditions Page A-24

approaches’ green durations to change from minimum to maximum and vice versa (Gazis et al., 1968). Michalopoulos and Stephanopoulos then later (1977, 1978) proposed an efficient two-stage timing method (known as “bang-bang” control) to find the “optimal” switch-over point during the oversaturated period. Later, Chang et al. (2000) showed that the continuous delay model used by Michalopoulos and Stephanopoulos is inadequate in identification of the optimal cycle length since a penalty for stops is not considered in the model’s formulation. Khakzadi (1980) demonstrated that the optimal control of an intersection cannot be achieved by only switching green times from a maximum to a minimum green splits while the cycle remained constant. By allowing the cycle length to vary within specified upper and lower limits, Khakzadi’s improved algorithm identifies the green allocation policy that minimizes delay and limits the queue growth rates. This formulation again assumes exact knowledge of the arrival rates over time. Chang and Lin (2000) further improved the Gazis model by developing a discrete optimization algorithm that uses the bang-bang-like control method to determine the switch-over point. Unlike in Michalopoulos and Stephanopoulos’ continuous model, the Chang and Lin model can allocate the switch-over points exactly at the end of a cycle. According to the authors, the discrete operation provides a smooth, regular, and ordered transfer of control and reliable calculation of delays. The bang-bang-like control operates alternatively and sequentially with given minimum and maximum green times. According to Chang and Lin’s study, such a control strategy can significantly outperform a policy that only increases the cycle time and gives each phase the same amount of additional green time. Chang and Lin indicate that the performance of their algorithm is robust even when the input data have some measurement error, but the approach still requires a trajectory of given, known input volumes. Potential Application to Practice These approaches that switch green times between minimum and maximum values must also take into account that these times are typically regulated by laws in many agencies (e.g., pedestrian crossing time). Such regulations could be considered as constraints in the mathematical formulations. In general though, the main limitation of such policies is that the optimization algorithm is driven by knowledge of the exact demands during the oversaturated period. However, the policy where the maximum possible green is allocated to the most congested approach is an intuitive one. Such a determination could easily be made with a queue estimation algorithm for each approach or movement at an intersection. Maintaining Queue Ratio Longley (1968) introduced the concept of a real-time “queues proportionality” strategy. Longley suggested that green times should be adjusted to balance the queues at the approaches to an intersection during oversaturated conditions. The objective is to minimize the number of intersections that are blocked by queues. Longley first considered the dynamic behavior of an Operation of traffic signal systems in oversaturated conditions Page A-25

isolated intersection and then extended the analysis to a network of multiple intersections. Longley suggested that the stability of the network may be improved by the use of coordinated signals, and by establishing stability criteria for the parameters of each controller. The policy, however, requires exact measurements of queue lengths. Queue-Actuated Control (Extended Green Time) Miller (1965) suggested the use of actuated control to minimize delays in critical intersections. He suggested that the green time of each approach be extended to a maximum value, to accommodate the additional traffic demand. The proposed policy does not depend on queue measurement (i.e., number of queued vehicles), but rather on attaining a certain threshold value that triggers the actuation. This concept appears to have motivated much of how existing modern actuated-coordinated systems operate today in undersaturated conditions. The queue-actuated control was claimed to be effective in high volume levels in general and, in particular, intersections with high turn-in volumes. Miller’s concept was later expanded by Weinberg (1966) to include downstream delays in the computation of total delay. Including the downstream delay in the formulation provides coordination between the critical intersection and its neighboring intersections to ensure that the reduction in critical intersection delay would not come as a cost of increased delay at other intersections. Weinberg’s model was modified and evaluated later by Ross et al. (1971) via simulation. The simulation results indicated that the policy is effective to reduce total delay during saturated conditions. The policy implementation required special treatment regarding the allocation of detectors. Strategies for Coordinated Intersections For under-saturated traffic conditions the offsets between green phases can be determined based on the objective of maximizing the progression band. As congestion increases and intersections become more saturated, residual queues (and queues from side-street turning traffic) start to disrupt movement at upstream intersections and the progression band concept does not apply any longer. If oversaturation continues for a considerable time period, standard fixed-parameter timing plans are likely to aggravate movement disturbance caused by the "spillback" of queues into upstream intersections since they do not take into consideration the actual traffic state (May, Montgomery and Quinn, 1988). This means that conventional procedures for optimizing fixed-time signal control, such as TRANSYT (Robertson, 1969), deteriorate rapidly when severe congestion persists. Queue Length Control Gordon (1969) was the first to introduce the concept of optimizing the green times to utilize the storage capacities of approaches to store queues at saturated intersections instead of minimizing the total delay. Gordon’s idea was to make efficient use of the upstream link’s storage capacity by delaying the service of a particular approach as long as possible. Operation of traffic signal systems in oversaturated conditions Page A-26

Offset Design Schemes for One-Way Progression Pignataro et al. (1978) suggested that in oversaturated conditions a more effective one-way progression scheme is to use “reverse progression” offsets rather than planning for a progression band of unimpeded flow. In a reverse (or “negative”) progression, the offsets are set to allow the green of a downstream intersection to starts before the upstream green in order to flush the residual queue. In addition, Pignataro et al. recommended using the following principles to determine the signal split at the upstream intersection on an oversaturated arterial where a "reverse progression" offset is being applied: 1. Reduce minor-street green time to ensure fewer turn-in vehicles will take a disproportionate amount of the storage in the oversaturated link. When restricting the turning phase time, care should be taken that it should not limit the cross-street through movement. Pignataro et al. recommended physical capacity improvements (e.g. turn lane bays and signalization) in this case. 2. In allowing such cross-street movements during oversaturation, it should be recognized that the arterial movement needs requires only as much green as it can use effectively use at the downstream intersection. Any additional green is, in fact, wasted but could be allocated to the arterial to restrict the cross turning movement (if signalization by movement does not exist). 3. If possible, turn prohibitions should be considered on the cross street so that the crossing through movement is not impacted by turning vehicles that are queued on the side street. Pignataro et al. (1978) also suggested a queue management strategy that was termed "equity offsets." This strategy, based on the principle of reverse progression, seeks to provide equitable treatment of competing flows at a congested intersection located upstream of an oversaturated link. Pignataro et al. identified two main cases for determining the intersection splits: Where (1) there are negligible turn-ins from the cross (minor) streets at the upstream intersection, and (2) where the split at the intersection can be as commonly determined (e.g. Webster's delay minimization formula). Where there are substantial turn-ins from the cross streets at the upstream intersection, the cross street traffic should be allowed to have enough green to put its "fair share" of vehicles into the oversaturated link. Pignataro et al. did not specify in detail what a “fair share” of green time would be. Pignataro et al. (1978) did not provide mathematical equations to cover all combination of approaches and volume/turn combinations. However, they offered a set of principles that can be used to govern the formulations of the most appropriate strategy. Rathi (1989) described a procedure that uses simultaneous and negative offsets along the arterial to control signals and manage queues and negative offsets and flaring of green along cross roads. Operation of traffic signal systems in oversaturated conditions Page A-27

Recent research by Baird, et al (2007) and Smaglik, et al (2007) describe simple controller logic for truncating a phase early when faced with a downstream restriction of flow. While not explicitly known (from downstream detection) that the downstream point is restricted, the logic assumes that no flow on a detector when the light is green indicates that there is nowhere for the vehicle to go. Thus it is better to truncate the green time for the current phase and move on to another phase. Models for Queue Interactions between Closely-Spaced Intersections Rouphail and Akcelik (1992) presented an analytical model for predicting the effects of queue interaction on delays and queue length at closely-spaced signalized intersections. Queue interaction refers to the model’s ability to predict a reduction in upstream saturation flow rate due to interference from a downstream queue. It was shown that the presence of downstream queues has a strong influence on the performance of the system with limited space for queuing. The queue interaction effect may alter the location of a critical intersection. Prosser and Dunne (1994) analyzed the paired-intersection problem by presenting a procedure that explicitly considers queue-blocking effects for determining the capacities of movements at closely-spaced intersections. They employed a graphical technique to estimate reduction in the effective green time. The model assumes no vehicles at an upstream intersection will discharge when queue spillback occurs downstream. Messer (1998) extended the Prosser-Dunne model to incorporate a wider range of operating conditions and developed an algorithm based on this extended model to determine the effective green time, phase capacity for the two intersections, and the relative offset between the two signals. The extended model does not stipulate the downstream intersection to be oversaturated, as the blockage due to spillback may occur during under-saturated conditions. Strategies for Diamond Interchanges Diamond interchanges operate as critical links between freeway and surface-street roadway facilities. During peak periods, inefficient operation of a diamond interchange and neighboring traffic signals may cause the system to become a bottleneck, degrading not only the capacity of the interchange but also that of the arterial and, in some cases, even the capacity of the freeway due to spillback on the exit ramps (Kovvali et al., 2002). Kovvali, et al developed an extension for PASSER III to consider oversaturated conditions in optimization of diamond interchange timings. By modeling the effects of spillback, the authors claim to outperform the signal timing approach in PASSER III. The most important factor to be considered in the oversaturated diamond interchange is the management of queue formation on external approaches. Local agency engineers tend to favor strategies that cause queue formation on the cross streets in order to enhance progression. In the case of diamond interchanges, these cross-streets are the freeway off ramps or frontage roads. Meanwhile, state highway officials prefer strategies that control queue length on the off-ramp so vehicles do not back onto the mainline freeway facilities and impede Operation of traffic signal systems in oversaturated conditions Page A-28

their operation, possibly at the expense of the arterial street network (Herrick and Messer, 1992). Queue spillback and oversaturated conditions at interchanges can have significant detrimental effects. Signal control at diamond interchanges has traditionally has been provided by either a three-phase signal sequence or the four-phase, two-overlap signal phase sequence (TTI-4-phase). The three phase sequence favors progression for the arterial through movement. Three-phase control requires sufficient queue storage space for the left-turn and cross-street movements to avoid queue spillback. Four-phase with overlaps (TTI-4-phase) is commonly seen when spacing between the two intersections of the interchange is very limited and thus queue spillback in between the two intersections would be a major concern (Tian, 2006). Texas Urban Diamond Signal Control was developed by Texas Transportation Institute (Herrick and Messer, 1992). This strategy incorporates both three-phase and four-phase sequences with overlaps. This control strategy seeks to maximize the benefit of both control strategies as traffic demand fluctuates. The three-phase control tends to favor the progression of the through movements on the arterial, typically preferred when the ramp traffic is balanced and the arterial left-turn traffic volumes are low. The four-phase control, on the other hand, is commonly used when the storage spacing between ramps is limited and queue spillback affects the upstream intersections. The Texas Urban Diamond Signal Control strategy switches between the two control operations based on a time of day schedule. The Arlington method was developed by the City of Arlington, Texas, for diamond interchange control under saturated conditions. The Arlington approach uses a dynamic phase selection process based on detection of the critical movements (i.e. interior left-turns) to activate two-phase, three-phase, or four-phase control based on cycle-by-cycle detector information. The Arlington method offers flexible operation to the interior left-turning movements by incorporating protective/ permissive left-turn operation for arterial traffic. During low traffic volume, two-phase operation is applied. As the traffic demand increases, either the three or four-phase control is activated based on the demand level. Messer and Chaudhary (1992) indicate that the additional phases initiated by activation of queue detectors may quickly result in “explosive” cycle lengths. The increase in cycle length and increase in the number of phases both yield increases in delay. As a result, the control strategy becomes ineffective if traffic demand simultaneously increases to the point of saturation on several approaches to the interchange. Kim and Messer (1992) developed a dynamic model as a mixed integer linear programming problem to provide an optimal signal timing plan for diamond interchanges during oversaturated conditions. The model maximizes system productivity (throughput), minimizes system delay, and controls queue lengths on external approaches to a diamond interchange. This technique is sensitive to knowledge of the traffic demand profile. Therefore, any significant variation in the Operation of traffic signal systems in oversaturated conditions Page A-29

actual traffic demand may render the signal timing plans established by this optimization approach ineffective. Real-Time Adaptive Control Algorithms Most of the techniques reviewed so far were originally designed to be established “off-line” and operated when oversaturated conditions were anticipated during a pre-set time. The challenge of most “off-line” methods is that the traffic volumes that are used to establish the control parameters of a given strategy are seldom realized in the real-world exactly. Particularly for oversaturated conditions, if the traffic demand increases to the saturation level before the capacity-maximizing strategies are scheduled to be applied, queues can quickly build up, leading to a serious delays and ineffectiveness of the algorithms. To avoid such scenarios, some adaptive traffic control systems can detect changes in traffic patterns before the peak period and begin capacity maximization algorithms or strategies earlier than a pre-scheduled approach. The main impediment to adaptive traffic control systems, however, is the significant cost in deploying detection systems that can supply the necessary traffic data for online decision-making. The main benefit of applying adaptive systems to oversaturated conditions is that they delay the onset of the oversaturation and can more effectively remove the conditions during the recovery regime. SCOOT The SCOOT (Bretherton, R.D., 1989) control system includes algorithms for dynamic control of individual intersections, arterials, and grids/networks. The core algorithms of SCOOT use a link flow profile (a composite representation of volume and occupancy) to tune the cycle time, splits, and offset values of each intersection. These algorithms have been proven to reduce delay in light to medium traffic conditions. However, if queuing occurred right up to the exit detector, SCOOT was not able to model this condition, and would could not detect the stationary vehicles (i.e. no demand) and reduce the green time, in turn increasing the congestion. Since SCOOT Version 2.4 (Bretherton and Bowen, 1990) many features have been added to SCOOT to tackle the problem of address severe congestion (Martin, 2006): • Normal/typical cycle time tuning • “Trend saturation” to schedule rapid increases to cycle time • Gating and action at a distance • Congestion offsets and congestion links Initially, both SCOOT and SCATS have algorithms that allow trading of split time from one phase to another based on the degree of saturation of each phase. Conventional actuated traffic control handles this condition with fixed or floating force offs. When force-offs “float” the extra time not used by a phase that gaps-out is provided to the coordinated phase on that ring. With fixed force-offs, the next phase in the sequence can use the extra time if it is oversaturated. In SCOOT, if the %SAT is >1 for a phase, if there are any skipped phases in that ring, the extra time can be used by the oversaturated phase if it follows in the sequence or can be rotated in the sequence. In Operation of traffic signal systems in oversaturated conditions Page A-30

addition, each cycle in SCOOT the split optimizer can increase or decrease split times by a small amount (+/- 4 seconds) to keep all of the splits below 90% saturation (which could be a user-defined value, but not typically modified). If re-allocation of the split time cannot accomplish this, then an increase to the cycle time is recommended. One criticism of this policy is that it tends to increase the cycle time of many intersections for oversaturation on approaches that may have only minor flows (i.e. side-streets). Similarly, balancing saturation levels favors side street delays at the expense of progression on arterials and coordinated operation. To accommodate this, SCOOT includes the concept of split “weighting” that allows the split optimizer to consider certain phases as more important than others when considering adding additional split time. A similar concept was implemented in the ACS-Lite adaptive system (Gettman, et al 2007) as well as RHODES. Split weighting is a preventative measure to delay the onset of congestion on key links. SCOOT also includes the concept of trend saturation to provide more rapid increases to cycle times during congestion. Apparently a lower threshold value (say, 80% saturation) can be determined to start rising the cycle time earlier at certain times of day when it is known that the traffic will rise rapidly, such as right before the peak periods. SCOOT also includes a number of features to improve progression performance during heavy flows. Initially, SCOOT (and other adaptive approaches including ACS-Lite, RHODES/OPAC, and SCATS) tunes offsets to accommodate changes to directional flows in their base algorithms. SCOOT tunes offsets to improve the total delay and stops on each approach to the intersection based on how the link flow profiles will change when small modifications to the offsets are made. If the modification is determined to be beneficial (total delay is reduced), SCOOT will adjust the offset a small amount (e.g. +/- 2 seconds) in the next cycle. When congestion rises further, it becomes more difficult to determine which direction of change will be beneficial, so SCOOT allows the user to provide biasing weights to directional flows by time of day. This pre-sets SCOOT to favor certain directions of travel that the user knows to have heavier flow. While this can be effective, it could make situations worse if there are incident conditions or a special event and the flow patterns change dramatically. In addition to the ability to bias the performance of offsets on certain links, the user can also fix or set an offset value in SCOOT. This is an important concept for intersections that are spaced very closely to each other. If the offset is fixed, then SCOOT will not attempt to tune that offset value to accommodate changes in the traffic flow. Finally, SCOOT includes the concept of a “congestion offset”. A congestion offset is one that is pre-determined to be used when a certain link is congested. Congestion is defined as occupancy over the upstream detector on the link (at the exit point) continuously for several seconds (e.g. 15s). When this condition is satisfied, SCOOT will adjust the offset to the pre-set “congestion offset” value immediately without going through a series of small adjustments as would be experienced with its normal offset tuning algorithm. This concept is similar to one that is used in many of the situations reported by practitioners that they have implemented using traditional actuated controllers with detector logic capabilities. The practitioners typically did not modify Operation of traffic signal systems in oversaturated conditions Page A-31

offsets in this situation but rather increase the green time of the phase that the queue is served by. The combination of biasing, congestion offsets, fixed offsets, and the normal tuning operation allows significant flexibility in the operation of SCOOT for offsets, although these features must be used carefully together. Similar to the concept of a congestion offset, SCOOT also includes the capability to use the congestion conditions on one link to affect the splits and offsets at other links. This provides the general ability by using logical conditions (if…then rules) to implement gating (holding more traffic on links that have higher storage capacity) or flared green, where downstream link green times are increased to move traffic away from a congested area more rapidly. (Martin, 2006) describes capabilities in SCOOT to define both “congestion links” and “trigger links” and it appears that a “congestion link” refers to the continuous occupancy over the upstream detector for a specific number of seconds, and a “trigger link”, for the purpose of beginning gating, is defined by the degree of saturation %SAT measure. For the purpose of “congestion” response, it appears that a single if…then condition is used to start the special operation. For the purpose of gating response, it appears that multiple trigger links are used in a “cluster” to begin a gating operation in a region where multiple gated links have their green times reduced to restrict in-flow and others may have their green times increased to improve out-flow from the area. As a natural extension to all of these other capabilities, SCOOT includes the ability to weight the importance (i.e. the “congestion importance factor”) of these if…then conditions, so that triggering special handling (gating, biasing split optimization, or offset modifications) of congestion at one location can be prioritized versus special handling for congestion at other locations. Hysteresis is also invoked for the gating operation to dampen the switching from gating operation to normal operation. How all of these features are combined together is not well described in available literature, but it seems reasonable that under certain recurrent congestion scenarios the combination of tools in SCOOT could be configured to improve throughput substantially. No real-world examples of these operations could be found in open literature, but we surmise that these features were implemented over time in reaction to inefficiencies in the operation of the base algorithms of SCOOT in real-world deployments. SCATS The SCATS control system operates similarly to SCOOT as it tunes the cycle, splits, and offsets of intersections in a group or section. Open literature (Martin, 2006) identifies that SCATS has far fewer features than SCOOT that are specifically designed for oversaturation management. SCATS allows re-allocation of split time from one phase to another at an intersection to balance the degree of saturation on all phases. In addition to split re-allocation, if all stages are maxed-out, then the cycle time is allowed to be increased to keep the most saturated phase at or below a threshold value (typically 90%). In each split plan one or more phases can be designed as the “stretch” phase (“stage”, in SCATS terminology). As the cycle time is increased up to a specific pre-specified value, any additional increase in the cycle time will only increase the Operation of traffic signal systems in oversaturated conditions Page A-32

available green time for the stretch phase. This approach uses existing engineering knowledge of field conditions to allow biasing of the additional split time to the phases that are likely to become oversaturated and cause more damaging results than oversaturation at other stages. In addition to incremental changes to the cycle time (similarly to SCOOT), SCATS includes the capability to quickly jump to a higher or lower cycle time using thresholds on the degree of saturation a given key phase. Offset selection in SCATS is much simpler than in SCOOT. SCATS provides four options for offsets in each traffic section, two-way progression for low traffic, two-way progression for high traffic, and both inbound and outbound offsets for directional flows, such as are typical during AM and P.M. peak periods on arterials. Pattern matching is used to select which set of offsets is used under any given pattern. UTC Operational Strategies The UTC strategy has been developed to address traffic-responsive network-wide signal control, particularly under saturated traffic conditions. The aim of the UTC strategy was to provide, at each cycle, traffic-responsive signal settings, taking into account the overall traffic conditions within an urban network (Quinn, 1992). A list of area strategies was compiled from a literature review and from a pilot study interview at Leeds UTC (Gray and Ibbetson, 1991). Operators at each of the ten cities were asked whether they used any of the following strategies: 1- Forced and hold green 2- Gating and metering 3- Maximum capacity flow 4- Negative offset- reverse green waves 5- Green waves with cross streets 6- Flared green with cross street (increases the green time windows at downstream intersections) 7- Diversion away from congestion 8- Shorter cycle length 9- Longer cycle length Not all strategies were commonly used by the operators that were interviewed. Force and hold greens were only used in special occasions or events (e.g., emergency vehicles). Metering and extended green were identified as the most effective techniques. Metering was predominantly used to reduce the effect of congestion on the network, while extended green was used to recover from congestion. Negative and simultaneous offset were used within predetermined plans for Operation of traffic signal systems in oversaturated conditions Page A-33

long queue formations. Quinn (1992) suggested that the two techniques could be used for improving operations on arterials. Traffic Metering/Gating Metering or gating strategies impede traffic input at a suitable point upstream to prevent the volume from reaching critical levels at downstream locations. These strategies can be applied either locally, to protect a particular intersection, or on an area-wide level. The practitioner must determine suitable links with enough spacing to store the metered traffic. The Gated links are those links which have been designated as links that store the queues which would otherwise block the bottleneck link(s). Rathi (1991) discussed queue management control strategies that meter the rate of flow into and within high-traffic density networks (control areas). He broadly categorized metering control strategies as “internal” metering and “external” metering approaches. Internal Metering strategies include Critical Intersection Control, arterial strategies that control the flow along congested arterial roads, and grid strategies that control the flow along major arterials and along minor cross-streets in order to prevent "gridlock" conditions. External Metering strategies, on the other hand, restrict the inflow of traffic along the periphery of a control area, while servicing demand at an acceptable level to improve the overall quality of traffic flow within the control area. The overall performance of the affected traffic should be improved. Rathi and Lieberman (1991) indicate that the external metering control strategies have the potential to improve traffic operations within and on the approaches to a congested control area (Quinn 1992). Although metering strategies appear to have potential, many factors against implementing metering control schemes should be considered. For example, there are likely adverse impacts on business in the affected area and equity concerns since not all metered road users are necessarily contributing to the congestion. It can also be argued that the effect of metering may simply transfer the congestion from inside to the outside of the control area without any change in overall travel time, air quality improvement, or other benefits. There are other implementation issues with metering, including the existence of sufficient storage, the potential of re-routing of delayed traffic, and the potential for increased red-light violations at the queue storage locations. Quinn also cautions that while traffic metering can be applied to protect a busy area from the sudden influx of morning peak traffic, it is less likely to be effective during the evening peak when much of the traffic originates from within the control area (e.g., parking garage) (Quinn 1992). Lieberman et al. (2000) developed a real-time traffic control policy for congested arterials based on the concept of gating. The algorithm, denoted RT/IMPOST (real-time/internal metering policy to optimize signal timing), is designed to control queue development on every saturated approach by suitably metering traffic to maintain stable queues. The control objectives are to (a) maximize system throughput, (b) fully use storage capacity, and (c) provide equitable service. Consistent with this approach, bounds on queue lengths and signal offsets are determined using a mixed-integer linear program (MILP). A simulation study was conducted by Lieberman to Operation of traffic signal systems in oversaturated conditions Page A-34

compare four signal timing tools, including RT/IMPOST, PASSER, TRANSYT, and SYNCHRO. The result showed that RT/IMPOST policy yielded improved network travel speed and delay during oversaturated conditions. RT/IMPOST considers the turning movements at the intersection and precisely controls the duration of each phase at every cycle length to ensure that downstream intersections queue lengths lie within the levels defined earlier in the optimization process. Hence, detector “blackout” effects (where the queue occupies the detector) should be limited by carefully locating the advance detectors. However, the procedure has a limitation that it does not support lagging left-turn phases and still requires detailed knowledge of approach volumes over time. Girianna and Benekohal (2002) developed a procedure based on genetic-algorithms (GA) that produced signal coordination timing for grid-networks of oversaturated one-way arterials. The algorithm provides signal timings that are responsive to traffic demand variations. The proposed procedure applies an online load-balancing mechanism to protect critical intersections from becoming oversaturated. Therefore, positive and/or negative progression strategies can be employed depending on the location of the critical intersections (i.e., entry or exit points). The algorithm adopts a two-stage strategy where queues are dissipated first before progression is achieved. The duration of the first stage depends on the location of the critical intersection. When critical intersections are located at exit points of the network, all upstream signals’ entering traffic volumes are metered by setting lower green times. The metering strategy is combined with setting negative offsets at the exit signals. Later, the offsets are gradually set to positive values as the algorithm promotes forward green bands. When critical signals are located at entry points, the negative offsets are maintained for a longer duration at the entry signals than they would be at the exit locations. This will ensure that all local queues are cleared before more vehicles arrive at downstream signals. The common cycle is changed depending on arrival volumes and the effective green times that are optimal to provide for the queue-dissipation process. The algorithm requires an efficient use of the green time and avoidance of “de-facto” red. De-facto red exists when the signal is green but traffic cannot proceed because of backed-up traffic on a receiving street (Abu-Lebdeh et al., 2000). The de facto red can be avoided by allocating less green time to upstream intersections than downstream intersections. The main issue with the Girianna and Benekohal’s algorithm is that it does not explicitly accommodate turning movements from side streets. In addition, the de facto red constraint makes it difficult to apply it in a two-way-street network. The procedure would be ineffective if congestion occurs in both entering and exiting points, since the algorithm assumes that congestion will occur in either the entering or exiting points, but not both. In a later effort, Girianna and Benekohal (2003) extended the work with a procedure for dissipating queues on two-way arterials. While the method is intriguing, its reliance upon known volumes is problematic for application in practice. Recovery from Severe Congestion Most, if not all, traffic signal control strategies were developed to prevent, or at least delay, the onset of congestion. Some strategies have been developed to manage situations of oversaturated Operation of traffic signal systems in oversaturated conditions Page A-35

conditions. Even fewer control strategies have been developed specifically to speed recovery from congested conditions. Many standard traffic management strategies can increase capacity and hence prevent or postpone the onset of congestion. However, once secondary congestion occurs, alternative methods are required. In terms of recovering as quickly as possible from severe congestion, Quinn identified the following approaches: 1) modify the control system to disperse the critical queues, 2) provide reserve capacity to relieve congested links, and 3) reduce temporarily the level of demand (Quinn, 1992). These recommendations are not followed by detailed analytical algorithms to achieve these objectives. Daganzo (1995) introduced a cell transmission model (CTM) to capture traffic flow dynamics. CTM is used as a numerical discrete approximation of hydrodynamic traffic flow theory (Lighthill, et al, 1955). Lo et al. (1999, 2001, and 2004) applied the CTM to control oversaturated networks by introducing a Dynamic Intersection Signal Control Optimization algorithm (DISCO) that uses the CTM as the calculation engine. DISCO produces timing plans that untie the gridlock in a few cycles. Due to its dynamic nature, DISCO supports fixed green splits in fixed-cycles or fixed-time plans as well as timing plans with variable green splits (variable cycle times). A genetic algorithm is used to find a near-“optimal” solution. However, the DISCO model is, like most past work, sensitive toward the quality of traffic volume data. Another limitation for practical implementation of DISCO is the substantial computational power needed to solve the optimization problem for a large network. Chang and Lin (2004) presented a dynamic control method incorporating a bang-bang control strategy to improve TRANSYT-7F’s performance in dealing with oversaturated signalized networks. Their method considers all the over-saturated conditions until all intersections become under-saturated. The cycle length of every oversaturated intersection in the network is assigned to be equal to the cycle length of a “pivot” intersection, which is defined as the most congested intersection at a certain cycle period. A branch and bound search method is utilized to look for progression routes that maximize the throughput of the network considering the progression priority. The proposed pivot search method then moves from one intersection to another as intersections change from over-saturated to under-saturated. When a new pivot intersection is identified, a fresh timing plan is generated on the basis of the new pivot intersection’s conditions. Finally, a smoothing procedure was introduced to minimize the effects of transitioning between timing plans. Dynamic Optimization Algorithms A number of strategies have been developed to improve the signal operation in oversaturated conditions (Abu-Lebdeh et al., 1997; 1999; 2000; 2003; 2001, Park et al., 1999; 2000). These strategies generally attempt to accomplish the following: 1. Identify the queue and the queue discharge time. 2. Identify the downstream storage available for queue discharge. Operation of traffic signal systems in oversaturated conditions Page A-36

3. Maximize throughput by avoiding the provision of green time that cannot be used or is inefficiently used because traffic cannot flow during the green periods. Abu-Lebdeh et al. (1997) developed a dynamic algorithm to obtain an optimal or near optimal-signal control trajectory (i.e. changes to offsets, splits, and cycle length) so system throughput is maximized subject to constraints on state and control variables (green times, and offsets) designed to prevent occurrence of de-facto red. Abu-Lebdeh et al. (2003) extended the previous research with a dynamic traffic control algorithm that can be customized for different priorities to arterial and cross street traffic in order to attain a desired traffic management strategy. The proposed procedure consists of two components: (1) a dynamic signal control algorithm that utilizes queue information to set different signal parameters to maximize the system throughput, and (2) a disutility function that evaluates the algorithm response based on the selected system performance goals. A real-time queue information feedback mechanism is needed for practical application of the strategy in the field. With queue estimation of field conditions, the algorithm could then compare the information with its projected values, and re-solve the optimization problem. Park et al. (2000) developed an extension to TRANSYT that finds signal timing parameters (cycle, split, and offset) in oversaturation conditions using a genetic-algorithm as the search routine. Kovvali et al. (2002) developed mesoscopic simulations software (Arterial Signal Coordination Software, ASCS) that can optimize diamond interchanges in oversaturated conditions. The strategy also uses a genetic algorithm approach to obtain near optimal solutions that encompass cycle length, phase sequences, and ring lag/ internal offset. Both efforts were focused on the search algorithm and modeling of oversaturation more so than any generalizations or findings about the differences in the resulting timings versus the timings that is produced by naïve TRANSYT or PASSER III. In both cases, demand volumes over time are necessary to run the models and obtain the “optimal” timings and phase sequence. Reduced Cycle Times Long cycles reduce the overall proportion of time lost during phase changes, and in general, increase the capacity of an intersection. There is some debate, however, that longer cycles may not be as appropriate in oversaturated conditions. Quinn (1992) speculates that short cycles have the following advantages over long cycles: a) Short cycles allow a high saturation flow to be maintained throughout the green period (the saturation flow falls if the exit side is blocked, which is less likely for short cycles since more vehicles can be stored in the downstream link if it is red); b) Short cycles are useful for clearing intersections blocked by turning traffic; and c) Short cycles provide more frequent opportunities for pedestrians to cross. Operation of traffic signal systems in oversaturated conditions Page A-37

Since a reduced cycle time decreases the capacity of the intersection as a whole during free flow conditions, it must be allowed to revert to its original value as soon as it has achieved its objective to manage the oversaturated condition (Quinn, 1992). Multi-Objective Analysis In undersaturated conditions, delay minimization and bandwidth maximization are the two main strategies that are used to optimize traffic signals in arterial networks. Targeting one or both of these strategies may not fully provide effective control strategies during oversaturated conditions. In oversaturated conditions, queues develop and grow in length, and the total delay increases exponentially as a function of the elapsed time. Congestion can spread out spatially and temporally due to queue spillback and may cause gridlock. Minimizing delay alone is not sufficient to resolve queues and their impacts, most simply because it cannot be directly measured. In order to possibly maintain optimal states of traffic during over-saturated conditions or transitions between unsaturated and over-saturated conditions, some control strategies have been developed in the past and reformulated based on an integrated criterion that combined delay minimization and queue management through a multi-objective analysis framework. Lan et al. (1992) proposed the COMBAND model to simultaneously minimize delay and maximize progression bandwidth on arterial networks to deal with the conflict between delay minimization and provision of bandwidth. The objective function is to optimize a linear, weighted combination of the delay/stop and bandwidth, subject to maximum queue constraints. Sayers et al. (1998) presented a Multi-Objective Genetic Algorithm (MOGA) for the signal control problem. The aim of their study was to optimize the signal controller off-line with respect to a number of diverse criteria, including reductions of specific vehicle emissions, pedestrians’ waiting time, vehicle stops, and vehicle delay. Since the users (vehicles and pedestrian) are competing against each other’s for a share of the resource (green time), a value referred as urgency is derived from the raw traffic data using fuzzy logic. This measure of urgency is then assigned for each respective user, which reflects its entitlement on the limited resource (green time). While the MOGA studies are intriguing, the researchers do not include any actual implementation results and queue management criteria are not considered. Chang-Jen et al. (2003) adapted multi-criteria decision making (MCDM) methods, including deviation minimization and compromise programming, to develop “compromise” signal control strategies and investigate the system performance of signalized intersections under various criteria. Their method is claimed to be capable of generating effective timing solutions fairly close to Pareto optimality for a given objective function. However, to achieve robustness of the control strategies, the stochastic process in which vehicular arrivals and queues dynamics are generated are not adequately considered. Operation of traffic signal systems in oversaturated conditions Page A-38

Abbas et al. (2007) proposed using multi-objective evolutionary programming to examine the effectiveness of alternate strategies for timing oversaturated intersections. The approach allows the optimization of several objectives simultaneously. Unlike traditional methods of assigning pre-defined weights to each objective function, multi-objective evolutionary algorithm produces the Pareto front of all objectives at the same time. This approach allows the analyst to explore the relationship between different objectives (delay, stops, throughput, queue length, cycle failures, etc.) and identify the time and locations where a shift from delay-minimization to throughput-maximization may be necessary to alleviate oversaturated conditions. Summary of Literature Review Significant effort has been invested in studying oversaturated conditions in the traffic control and traffic modeling community. Most of the work has been focused on identifying how to measure delay or model the effects of oversaturation. These efforts have marginal utility for this research. It has long been recognized that queue measurement and/or v/c ratio estimation is the key to managing oversaturation. There have been a myriad of approaches for performing queue estimation, most of which are input-output models that can estimate queues downstream of the detection point but have little predictive capability if the queue extends past the point of detector placement. Several encouraging research projects such as (Mueck, 2002) were found that show promise for methods to estimate queues using point detection that do not assume “exit” detection is available. The work begun by Liu (Liu and Ma, 2007) and extended in this project, continues along this direction to provide capability to measure queue length without exit detection. This is important since the typical agency in North America cannot afford to install “exit” detection at every intersection (or even at a subset of critical locations). While adding some detection is unavoidable in most situations, the challenge is that adding additional detection, while helpful, does not inherently solve the problem of not being able to measure the input volumes when the queue has extended past the detection point. Additional detectors are needed upstream of that location, and so on, and so on. Strategically placed detection is important, but we believe that methods such as will be described in Task 2 are necessary (until, at least, we have ubiquitous IntelliDrive technologies on a significant percentage of passenger vehicles and trucks). Approaches based on this type of technology are outside the scope of this research. Most of the remaining research in estimating oversaturated conditions has occurred within the context of the development of adaptive control systems. These systems must operate in both regimes (over and under-saturated conditions) and thus must at least approximate the effects of oversaturation. Because of the proprietary nature of these commercial applications, only limited open literature is available that describes their methodologies for estimation as well as strategies for control. Both SCOOT and SCATS approximate the degree of saturation with detector occupancy information. SCOOT models a link flow profile which is a combination of the volume and occupancy data at an exit detector location. SCATS models the level of saturation using short stop-bar detectors by counting the “spaces” between vehicles and comparing the amount of unused green to that amount of green that would be unused at saturation flow. Thus, SCATS can Operation of traffic signal systems in oversaturated conditions Page A-39

determine when a phase is oversaturated and has an estimate of how much, but does not directly model queue length. SCOOT more accurately approximates the queue on a link by modeling the dynamics of flow on the link. When the dynamics are not very dynamic, so to speak, SCOOT can estimate the severity of the congestion as compared to the link performance during saturation flow. ACSLITE and other systems that measure link flows at advance locations can approximate this type of oversaturation estimate, but have not developed algorithms to estimate the level of congestion or strategies to take advantage of the information. Several theoretical formulations of control approaches (“optimal” controls) for oversaturated conditions have been developed. This includes off-line search algorithms extending popular design tools (PASSER and TRANSYT) for finding cycle, split, and offsets that are “optimal” for oversaturated conditions. The main drawback of most theoretical formulations is the assumption that demand volumes are known and measurable in the real-world, which is simply not the case during oversaturation. Only limited study of the transferability of these formulations is presented in the literature, thus it is not clear how to take advantage of these efforts in this research work. Only a few research projects surveyed in the literature seem to have developed manageable, real-world approaches to control strategies for oversaturated conditions. These projects are: • Real-time internal metering policy to optimize signal timing (RT/IMPOST) was formulated explicitly for oversaturated arterials. RT/IMPOST controls the queue growth on saturated approaches by metering traffic utilizing the network storage capacity. (Lieberman, E. B., Chang, J. & Prassas, E. S., 2000). Although this method is driven by measurement of volumes, it may have value if it could be modified to accept queue length measurements as input values. • Diamond interchanges operation strategies in oversaturated conditions, which include Texas urban diamond signal control, the “Arlington” approach, and the approach developed by Kim and Messer. These strategies basically seek to maximize the system throughput by switching between two-phase, three-phase, and four-phase operations based on traffic demand variations through the day. Texas diamond mode in either three-phase or four-phases is the only strategy that is known to have been implemented in the field. The Arlington approach was used many years ago in the City of Arlington and is not known to have been continued operation. Kim and Messer’s approach never made it to the implementation stage. (Kim, Y. & Messer, C. J., 1992). • Rathi et al. (1988) developed a control approach that avoids spillback in urban grid networks by adjusting signal offsets. The control approach was implemented in Manhattan CBD in NY (5th Avenue between 63rd and 54th Streets) and provided 20% reduction in overall travel time. Combinations of strategies were used (i.e. simultaneous offset for N-S streets, negative progression with flared green for E-W streets to avoid spillback and capacity maximization). (Rathi, A. K. et al, 1998). Since this methodology was implemented in the real world, it should be investigated further for inclusion into the list of potential application strategies. Operation of traffic signal systems in oversaturated conditions Page A-40

• The if…then queue management concepts in SCOOT and SCATS. These approaches for initiating gating operation and triggering jumps or reduction in cycle times are directly relevant to this research project. Even without the adaptive operations that slowly adjust timings to react to changes in flow, the if…then strategies can be fairly easily implemented by agencies to accommodate specific recurrent conditions with the deployment of just a few queue estimation detectors. The algorithms developed in Task 2 for estimating queues and oversaturation levels combined with the approach discussed for development in Task 7 can leverage the concepts that are used in SCOOT to apply to real-world situations. Many other recent works have focused on individual intersections and simple rules for truncating phases due to downstream restrictions, or identifying that cycle times that should be just short enough to clear turn bay storage, but no longer. These methods are useful to include into the catalog of strategies, but need more systematic definition. Other research on strategies was found to be rather vaguely descriptive of a general approach such as “use negative offsets” or “use flared green times,” but little additional detail is provided as to describe how to mathematically construct such strategies. Research Directions from the Literature Review Previous research in diagnosis of oversaturated conditions and strategies for addressing these conditions is not extensive. This is mainly due to the challenging nature of the problem and the limited capabilities of existing state-of-the-practice detection systems. Most models for diagnosis and evaluation of oversaturated conditions have volumes as an input variable. This is especially problematic since the limitations of existing detection systems make it rather difficult to measure volumes during oversaturation, and particularly when queues grow past the detection point. Thus, the various “optimal” control formulations that we found in the literature are difficult to apply or use in practice since they need this volume information that is highly unreliable. We also found little to no research on what is an appropriate time for a queue to be persistent before a corrective action(s) needs to be taken. From these findings, we identified that it is important to develop diagnosis techniques that can identify queues that grow past the detection point (measurement of the extent of oversaturation) and identify quantitative measures for the relationships between queue lengths and green time allocation. Similarly, it will be important to characterize the degree to which a traffic facility is oversaturated, versus simply indicating that it is oversaturated or not. In particular it is important to identify when an oversaturated condition on one traffic facility (movement, link, or intersection) is detrimentally affecting the traffic flow on other facilities. This dovetails with the development of a framework for recommending appropriate strategies to practitioners since certain actions (such as phase truncation, for example) can have a direct and reverse effect on the intended outcome, if applied incorrectly. Similarly, if the bar is set too low to indicate when alternative control measures are required, actions might be Operation of traffic signal systems in oversaturated conditions Page A-41

implemented by practitioners when the condition is intermittent. Undue delays might be induced when the corresponding improvement to throughput is not significantly realized. Literature on strategies for handling oversaturation could be criticized in several areas: • not having enough description to replicate, • requiring information on traffic arrivals and volumes that is not possible to measure, and/or • not having been supported by sufficient research performed to identify transferability of the results to other networks or traffic scenarios. Realized performance improvements due to application of strategies for handling oversaturated conditions have not in the past been described adequately. For example, the concept of gating or metering appears promising for further research although existing literature describes the concept very qualitatively. This also indicates that it is critically important to develop a diagnosis methodology that identifies when gating is warranted, and when it is not. Simultaneous and negative offsets were also identified in the literature as promising strategies for arterials, although there is little research on the implementation issues such as when to implement a different offset pattern and how to mitigate the transition effects. Some recent work has focused on individual intersections and simple rules for truncating phases due to downstream restrictions, or identification that cycle times should be just short enough to clear turn bay storage, but no longer. These methods are useful to include into the catalog of strategies developed in this project, but they need more systematic definition. Other more dated research on strategies was found to only provide general approach such as “use negative offsets” or “use flared green times”, but little additional detail is provided as to describe the details of such approaches. The same is true of the methodologies used by adaptive control systems, notably SCOOT and SCATS and to some extent RHODES. From these findings, we identified that it will be important in this project to provide quantitative evidence that certain strategies improve total throughput and other performance measures. In addition to this, the research will focus on study of the rationale necessary to identify when it is appropriate to switch objectives from minimizing delay (normal operations) to maximizing throughput, and mitigating the effects of queues. Operation of traffic signal systems in oversaturated conditions Page A-42