THE BENEFITS OF SIGNAL GROUP ORIENTED CONTROL Authors: Robbin Blokpoel 1 and Siebe Turksma 2 1: Traffic Engineering Researcher at Peek Traffic, robbin.blokpoel@peektraffic.nl 2: Product manager research at Peek Traffic, siebe.turksma@peektraffic.nl ABSTRACT There are many different traffic light control algorithms. They can be roughly classified as fixed time, vehicle actuated or traffic responsive. All of these control algorithms have stages, blocks or similar concepts at their basis that impede flexibility. This paper will therefore present a signal group oriented algorithm that is not bound by any constraints. This means that any combination of signal groups that is allowed by the conflict matrix can be actuated in any order. To isolate the effects of shifting to signal group oriented control, this paper tests a version of the same control algorithm that does respect stages and stage orders. These tests revealed that the overall time lost for all vehicles can be reduced by up to 21% by switching to signal group oriented control. INTRODUCTION Traffic light control algorithms have evolved towards more flexibility over the years. The first algorithms were fixed time, but soon it was noticed that a lot of green time was wasted especially with lower traffic demand. The first improvement to fixed time control was vehicle actuated control. Basically this method uses detectors to decide whether to cut off the green phase or extend it. More details about vehicle actuated control can be found in [1]. These algorithms are still among the best for isolated intersections. Although they are often bound by local, regional or national policy constraints, they are very flexible. Examples of national constraints are the ring-barrier structure of the USA [2] and the German RiLSA standard [3]. The major drawback of vehicle actuated control is that it only optimizes for a single intersection. Traffic responsive algorithms solve this problem by trying to coordinate consecutive traffic lights. These strategies usually work well when the traffic is busy, because the coordination between intersections will reduce the number of stops of the traffic. This reduces or even nullifies the acceleration loss at the beginning of a green phase and therefore more vehicles can pass during a green phase. Apart from this increased capacity, the travel time for vehicles will also be reduced when there is proper coordination. When there is less traffic, however, the benefits of coordination do not outweigh the loss of flexibility caused by the same coordination. Examples of traffic responsive control are SCOOT [4], Utopia [5] and Rhodes [6]. With vehicle actuated and traffic responsive algorithms available, a road operator can select either coordination or flexibility, while it would be best to have both at the same time. A possible solution to this would be to switch to a traffic responsive algorithm during the rush hour and to vehicle actuated one when there is less traffic. Still, this does not have the benefits - 1 -
of both systems at the same time period. The experimental control algorithm in this paper can coordinate major traffic streams while maintaining flexibility for the other streams. This is possible because the algorithm does not have to respect constraints given by stage orders, block structures or similar concepts. This paper will first explain the theory of the experimental control algorithm, together with how it can be programmed to respect stages and stage orders or be fully signal group oriented. Then the test network will be presented, which is a small network from a real city. In the results section an extensive comparison between Utopia, the signal group oriented experimental algorithm and the stage oriented experimental algorithm will be made. This research about signal group oriented control versus stage oriented control is part of the EcoMove project in the EcoBalancedPriority application and will as such also be part of the overall EcoAdaptiveBalancing and Control (EcoABC) application. See [7] for more details about EcoMove. THE SIGNAL GROUP ORIENTED TRAFFIC CONTROL ALGORITHM The architecture of the experimental control algorithm contains several important building blocks. In Figure 1 these blocks and their connections are shown. Figure 1: Architecture of the traffic light control algorithm As can be seen from the figure, information about the queue state and predictions of the future state play an important role in this algorithm. The predictions also contain coordination information from neighboring intersections. This queue information forms the basis for the solution generator that will generate possible control plans according to current traffic demand on all signal groups. The cost function calculates how good a certain solution is. This is based on waiting time, stops and the number of vehicles being serviced. The evaluation block determines whether the optimal final solution has already been reached, or whether more solutions must be generated to reach the optimum. The queue information is based on data from inductive loops, which are positioned at the stopline of every signal group and at the borders of the network for the busiest roads are queue entry detectors. For two consecutive intersections the stopline detector of the first serves as the entry detector for the next. A queuing model combines and enriches the data from the loops, while predictions are also based on the planning of neighboring intersections. - 2 -
When the experimental control algorithm works in signal group oriented mode, it can basically select any combination of signal groups to create the next stage. This means that the algorithm only has to respect the conflict matrix, which also helps to reduce the complexity, since an intersection without a conflict matrix of 12 signal groups would have 4095 possible solutions: ( ) ( ) ( ) (1) This number is reduced considerably when a conflict matrix is taken into account. For instance a T-intersection (see Figure 2) only has 6 signal groups and should have 63 possible combinations according to formula (1), but with a conflict matrix this number is reduced to 18. This number of 18 still includes some combinations which do not seem to be very efficient at first sight: (4), (4,6), (4,6,7). However, only having signal group 4 green can be efficient during the transition phase from one larger set of signal groups to another. Additional groups for pedestrians on an intersection also do not add many more possibilities, because they cannot be green at the same time with many different vehicle signal groups. So even though the number of possible combinations of signal groups that can be green at the same time seems to be very large, it gets to manageable numbers when the conflict matrix adds constraints. Figure 2: Conflict matrix of a T-intersection Using these 18 possible stages implies that it is not possible to define a cycle or cycle time anymore. Some of them might even be used only a few times a day, but the individual signal groups will of course get their green time from other stages when there is a demand. Still, there is no guarantee anymore that every signal group gets at least one turn during a cycle (there is no cycle anymore after all), which guarantees a maximum waiting time. Therefore, a mechanism is implemented that guards against excessive waiting times on each signal group. Without this mechanism solutions might be found especially in saturated conditions - that would completely ignore directions with little traffic. This mechanism will increase the weight in the cost function sharply after a certain configurable waiting time and in that way force a green phase for those vehicles. To put the experimental algorithm in stage oriented mode with fixed order, there are two solutions. The easiest one is simply give the solution generator only one option for every switch. That option will be the next stage according to the preconfigured order. The only thing the controller can vary is the duration of each stage and when there is no traffic on any signal group of a certain stage, it can be skipped. Additionally, there is a minimum and maximum - 3 -
stage length that can be configured. Since this greatly reduces the complexity for the controller, there is no problem letting it run in this stage oriented mode. However, contemporary algorithms with fixed stage order still have some flexibility inside a stage. Utopia, for example, has the mechanism of quick actuations and the Dutch standard of CCOL [8] allows alternatives inside a block. Going back to the intersection of Figure 2 the stages could be like this: Option 1: (2,7,8), (2,3,4), (4,6,7) Option 2: (2,7,8), (2,7,4), (2,3,4), (4,6,7) Option 1 represents the minimum amount of stages in a straightforward manner. Option 2 has an additional stage in case there is no traffic at 8, but there is traffic at 4. This works well when this actually happens, but when there is a lot of traffic at 2 and 3, while there is no traffic at 4 and 8, this extra stage will lead to a loss of performance. The cost function will indicate that both stage (2,7,8) and (2,7,4) have the same weight, so it will extend the stage (2,7,8) to the maximum stage length and then switch to (2,7,4) since there still is traffic on 2. As soon as it is in (2,7,4) it notices that (2,3,4) is better and after the minimum green time it switches to (2,3,4). To prevent situations like described above, smart stage skipping is introduced as a second method of the stage oriented approach. In this case the rules of skipping a stage are different: The difference between the current green stage and the next are compared and if there is no traffic on the additional signal groups of the next stage, then that stage can be skipped. So when at (2,7,8) and possibly going to (2,7,4), the difference is 8 and 4 (since 2 and 7 are common in both stages). And since there is no traffic at 4, this stage is skipped and the cost function will continue to evaluate the following stage which is (2,3,4). That one will be better than (2,7,8) and therefore the controller will switch directly after the minimum green time. This way it is still guaranteed that every signal group will get at least one turn during a cycle given that there is a demand. At the same time more flexibility is added especially when the intersection is more complicated, like a + shaped intersection with separate bicycle and pedestrian signal groups. Additionally, this method will also look ahead through all other stages after half of the maximum stage length has expired. When there is another stage with a higher weight then it switches to the next one (still respecting the stage skipping rules). That way it balances the green allocation between the stages better and it has more means to achieve a good coordination between two intersections. The experimental control algorithm uses the same method for coordination in both signal group and stage oriented mode, which is different from Utopia, SCOOT and other traffic responsive control strategies. Those strategies synchronize the cycle times of neighboring intersections and try to coordinate the most important signal groups within those cycles and therefore plan at least one cycle ahead in the future. The experimental control algorithm, however, plans only one transition ahead and even in the case of fixed stage order, there is no cycle time coordination. Still, coordination is easily achieved because of the comprehensive queue model that forms the basis of the optimization process. When a large group of approaching vehicles is detected in the queue model, the algorithm tries to prevent that they have to stop. This effort can be increased by configuring additional weight on the coordination between two signal groups of different intersections. - 4 -
This kind of coordination based on the short term horizon requires flexibility, so the fully flexible signal group oriented algorithm without fixed stage order will excel in this situation. This can best be illustrated by the example intersection of Figure 2 and the stages previously given by option 2. Now suppose the controller is at stage (4,6,7) and there is a lot of traffic approaching for signal group 3 with extra coordination weight, but all other signal groups have light traffic as well. In this case the fully flexible controller can switch as soon as the minimum green time has expired. The smart stage skipping alternative, however, first has to spend half of the maximum green time at stage (4,6,7) and (2,7,8) followed by the minimum allowed time at (2,7,4) before it can finally switch to (2,3,4). This is not so efficient and therefore Utopia simulations will also been carried out on the network to see whether this algorithm with fixed stage order is still competitive. TESTING ENVIRONMENT Figure 3: The test network of Veenendaal, with controlled intersections indicated by a circle around them. - 5 -
All tests were carried out in the network of Veenendaal, as shown in Figure 3. The traffic simulator used is Vissim version 5.30.08 [9]. The network contains 5 intersections of the ring road, of which two are really close together and connect the ring road to the highway on- and off-ramps. The 6 th intersection is one that is very close to the last intersection of the ring road in the model and has to be modeled as well for realistic results. This is because bad coordination between these two intersections will affect saturation flow of several signal groups on the ring road. Utopia controls the intersections that are very close together as one intersection in the reference situation. This was to ensure proper coordination between them, since the preconfigured stages only have coordinated signal groups among them. However, the experimental algorithm will control them as two separate intersections in order to have more flexible control options. The distance between 61 and 62 is approximately 70 meters and the distance between 64.5 and 64 is 90 meters. Taking a closer look at the network reveals that also intersection 63 is really close to 64.5. This intersection, however, is one with very few traffic that conflicts with the main flows between 62 and 64.5. The road on the right is the entrance to two restaurants and the road on the left in only accessible from the north while there is no traffic that can enter the ring road from there. The test scenario for this network is the evening rush hour, which is very busy with an intensity of around 5000 vehicles per hour in the network. The busiest intersection is 62, a detail view is shown in Figure 4: Figure 4: Detailed view of Intersection 62-6 -
Signal group 8 has a traffic volume of 729 vehicles per hour, effectively requiring a green share percentage of 40%. The conflicting streams are 5, 11 and 3 (double lane) which in their turn also conflict which each other and require 10, 15 and 15% of the green time. This totals to 80% of the time that has to be spent on these conflicting signal groups. This implies that there is only 20% of the time left for stage transitions or inefficiencies in the control program. With an amber time of 3 seconds and a clearance time usually around 3 seconds and at least 4 stages per cycle, this means that the cycle time has to be at least 120 seconds to fulfill the demand. This can be reduced when part of the amber time is used for traffic as well. Also, the pedestrians and bicycles cannot have their own stage, but have to be combined with nonconflicting vehicle groups. Of the other groups, signal group 2 has most traffic with 54% of green time required, but this group can be combined with both 8 and 3, which have 55% of the green time together. The same holds for 4, 7 and 9. These groups can be combined more easily with groups that have higher demand and no problems are expected on these groups. Another challenge in the network is the traffic that comes from intersection 64.5 and wants to turn left at 64 to access the highway. This signal group to turn left requires green for 37% of the time and good coordination with the traffic coming from 64.5. Otherwise a lot of green time can get wasted at 64, resulting in long queues. RESULTS The first simulations in the Veenendaal network of six intersections have revealed an additional problem that is a challenge for the traffic control algorithms: the trucks in the network have a very slow acceleration which leads to difficulty estimating the queue length from time to time and a reduced saturation flow. This effect is visualized in Figure 5: Figure 5: Large gaps caused by trucks As can be seen in the figure, the gap is 110 meters and the truck arrived 11 seconds later at the stopline. The queue entry detector has a circle around it in the figure and is at a distance of 300 meters. This means that the queue is actually measured as 35% shorter than it actually is, because if that truck would accelerate just as fast as the vehicles, then all vehicles behind the entry detector would have been detected already. Of course both experimental traffic control algorithm and Utopia estimate the traffic behind the entry detector, but this estimate is less accurate. Especially for a well performing algorithm this proved to be a problem, because the queue did not often reach beyond the entry detector and therefore the queue was often underestimated in these cases. Additionally, the gap time of 11 seconds is difficult for queue correction algorithms; because they might think the queue estimate was wrong and reset it to zero. The clearest problem with these big gaps is of course that 11 seconds of green time are - 7 -
wasted when the controller decides to continue the green phase. A vehicle actuated algorithm would in this case surely cut off the green phase and would actually do that from lower gap times already. This gap time problem also occurs for the experimental control algorithm at signal group 5 (see Figure 4) which is configured as vehicle actuated, because that queue is very hard to estimate. However, when a truck is near the end of the queue it often faces a red light once it has finally reached the stopline. Still this turned out to work better than trying to work with the inaccurate queue estimates. When the queue was over-estimated the controller would sometimes wait up to 10 seconds for a vehicle that isn t in the queue at all. When on the other hand the queue was under-estimated, the green phase could be cut off too early for up to three times in a row. Considering the vital importance of this signal group for the performance of the intersection, putting this queue to vehicle actuated was very important. For the Utopia configuration this problem did not occur, because intersections 61 and 62 (see Figure 3) were controlled as one intersection and therefore green phases were always coordinated. Another important concern for the experimental algorithm is the calculation time. Because many stages are possible in any order, the complexity may explode resulting in slow simulations and high system requirements for controllers on the street. Simulations showed that the six controllers can run together with the Vissim simulator at 9x real time speed for the fully flexible mode and 10x real time when the stage order is fixed. Note that the simulator was zoomed in at an empty part of the network, because drawing moving vehicles takes a lot of processing time. This speed was achieved at a 2.53 GHz dual core 64 bits processor while only 50% of the processor capacity was used. Memory usage was always under 150 MB for Vissim and the controllers together. For Utopia only 2x real time speed was possible. It should be noted that this does not imply that the experimental algorithm is 4.5 times faster, because Utopia requires a fixed timestep, so this 2x real time is a peak calculation time, while for the experimental controller the timesteps varied in speed between 8.5 and 9.5x real time speed. Results for average loss time and number of stops per vehicle are presented in table 1. For every controller type, 10 simulation runs of 2 hours each were done after a warm-up time of 300 seconds to allow traffic to enter every part of the network. This warm-up time was not included in the results. Multiple runs were necessary because the variation between two runs was quite large. For example the total delay time for all types for Utopia was between 123.7 and 136.5 seconds. The results are mostly positive for all modes of the experimental controller. The only common problem that appears from this is that the pedestrians have to stop around 50% more often, but have 40% less loss at the same time. This is due to the inefficient detection and signal group configuration at some intersections. Looking at signal group 35 and 36 in Figure 4 this becomes clear. Pedestrians waiting in the middle put a request to 36 independent of the fact if they want to go south or north. However, if they would go north, its green phase can still be combined with 7,8 and 9. Therefore, it would be much more efficient to have four separate signal groups for that pedestrian crossing. The network was not adapted to make this possible because Utopia simulations were already completed with the two-signal group configuration. The problem with a pedestrian wanting to cross from 35 is that the green phase is long enough to reach the middle, but not long enough to start crossing from the middle to the end. When there would be 4 signal groups, a high priority coordination request can be added to the next signal group required for the pedestrian to finish the crossing. Therefore, this problem is not - 8 -
something to worry about, especially because there are 8400 cars/trucks, 1400 bicycles and 120 pedestrians that pass during the 2 hours of simulation. The loss time results for the pedestrians in the case of signal group oriented mode is simply a matter of configuration and since waiting times of 38 seconds are not exceptional this is not a problem. Controller type Vehicle type Average loss time - 9 - Improvement compared to Utopia Average number of stops Improvement compared to Utopia Utopia All types 128.6 NA 2.36 NA Utopia Cars 150.6 NA 2.79 NA Utopia Bicycles 29.9 NA 0.61 NA Utopia Pedestrians 74.3 NA 0.97 NA Signal group All types 66.1 49% 1.51 36% Signal group Cars 69.5 54% 1.65 41% Signal group Bicycles 38.4-29% 0.72-18% Signal group Pedestrians 44.8 40% 1.53-58% Stage smart All types 84.3 34% 2.00 15% Stage smart Cars 94.7 37% 2.26 19% Stage smart Bicycles 28.4 5% 0.772-26% Stage smart Pedestrians 71.9 3% 1.45-50% Stage simple All types 105.8 18% 2.28 3% Table 1: Results for loss time and number of stops in the evening peak hour Another concern is the standard deviation of the waiting time and the number of stops. A high standard deviation means that the predictability of the travel time is low and this is generally not appreciated by road users. This might be an unwanted side effect of signal group oriented control, because there is more freedom to favor the main direction over the side directions. Therefore, the standard deviation for the experimental algorithm with and without fixed stage order was compared. The results for average loss time and average amount of stops are presented in Table 2. Controller Avg. loss time Std. loss time Avg. stops Std. Stops Signal Group 66.1 47.2 1.51 1.10 Stage Smart 84.3 54.5 2.00 1.38 Table 2: Comparison of the standard deviation for average loss time and average amount of stops for all vehicle types. Another interesting situation is what would happen when the algorithms are operating under off-peak conditions. The algorithms may be too much focused on the peak hour and therefore not being flexible enough during low traffic volumes. Results of the off-peak hour are presented in Table 3. These results for off-peak hour show comparable differences between the different control strategies. Since the configuration was kept exactly the same, there is of course still the same problem with pedestrians having to cross in two steps and the slightly lower priority for bicycles. Additionally, there is of course less of a challenge and small inefficiencies in the control program do not immediately lead to large queues. That is why the improvements are smaller than in the peak situation. It should also be noted that the fixed stage order algorithm still had an improvement on delay time compared to Utopia, but didn t do well on the number of stops, because coordination was harder to achieve with low traffic density.
Controller type Vehicle type Average loss time Improvement compared to Utopia Average number of stops Utopia All types 50.8 NA 1.15 NA Utopia Cars 61.8 NA 1.39 NA Utopia Bicycles 25.1 NA 0.68 NA Utopia Pedestrians 51.4 NA 0.96 NA Signal group All types 32.7 36% 0.87 25% Signal group Cars 33.4 46% 0.94 33% Signal group Bicycles 31.4-25% 0.66 3% Signal group Pedestrians 23.3 55% 1.40-47% Stage smart All types 43.0 15% 1.28-11% Stage smart Cars 50.1 19% 1.51-9% Stage smart Bicycles 23.0 9% 0.78-16% Stage smart Pedestrians 73.6-42% 1.68-75% Table 3: Results for the off-peak hour Improvement compared to Utopia CONCLUSION The new algorithm presented in this paper performed better than the current state of the art control algorithm Utopia. During off-peak hours the average loss time was almost halved with a 49% decrease. The average amount of stops, which greatly contribute to pollutant emissions, was decreased by 36%. The verification checks regarding standard deviation and the off-peak scenario proved that signal group oriented control does not suffer from a drawback common in the world of traffic control algorithms. The standard deviation of both the waiting time and the average number of stops was lower in the case of signal group oriented control compared to the same algorithm constrained to respect stage orders. The performance improvement of signal group oriented control over stage oriented control was 28% for the loss time and 32% for the stops. From the larger improvement on number of stops it becomes clear that flexibility is very important for coordination when only planning on a short term horizon. When comparing the experimental algorithm with fixed stage order to Utopia it is remarkable that it still performs better. From this fact it can be concluded that the higher quality queue model that formed the basis of this research outweighs the benefit of planning on a longer horizon. The experimental algorithm is very promising, since not only travelers will reach their destinations faster, there is also an increased network capacity and less pollutant emission due to traffic. Authorities need to rethink their usual constraints on the signal group sequencing to allow for improved traffic control that trades road user predictability for performance. - 10 -
REFERENCES [1] A.G.R. Bullen, Effects of Actuated Signal Settings and Detector Placement on Vehicle Delay, Transportation Research Record 1244, pp. 32-38, 1989. [2] M. Mladenovic, Milos and M. Abbas, Modeling Ring-Barrier Traffic Controllers Using Colored Timed Stochastic Petri Nets, IEEE ITS conference, September 19-22, 2010. [3] Forschungsgesellschaft für Straßen und Verkehrswesen, Richtlinien für LichtsignalAnlagen RiLSA, 2007, Köln [4] P. B. Hunt, D. I. Robertson, R. D. Bretherton and R. I. Winton, SCOOT - a traffic responsive method of coordinating signals, TRL Laboratory Report 1014, 1981. [5] V. Mauro and D. Di Taranto (1990), UTOPIA, Proceedings of the 6th IFAC/IFIP/IFORS Symposium on Control and Communication in Transportation, Paris, France. [6] K.L. Head and P.B. Mirchandani, RHODES: A Real-time Traffic Signal Control System: Architecture, Algorithms and Analysis, Transportation Research, 9C(6), pp. 415-432, 2001. [7] N. Eikelenberg, Cooperative systems and services for energy efficiency: From inefficiency to efficiency, ITS Europe conference proceedings, 2011, Lyon. [8] CROW, Handboek verkeerslichtenregelingen, 2006, (in Dutch) [9] PTV Vision, Vissim 5.30-05 user manual, 2011. - 11 -