Transport sector is one of the largest CO2 emitters. As the economy dramatically grows, the travel demand is also significantly increased in developing countries. A huge number of vehicles traveling on the road have caused a large amount of CO2 emission into the air. Moreover, it causes heavy traffic congestions and traffic accidents that hinder the economic growth. Thus, it is necessary to integrate public transport into transport network. Public transport reduces CO2 emission, traffic congestion and also fatality by traffic accident. The most considerable public transport in developing countries is public bus, which is suitable for booting up public transport and raising public awareness of advantages of public transport. Responding to the issues above, Phnom Penh, capital city of Cambodia has put public bus in service in 2014 to reduce traffic congestion as well as car ownership. In order to promote bus ridership, there are many measures needed to improve bus operation, especially travel speed. A study report by JICA (2014) proves that the traffic congestion is a significant issue that causes the travel speed of vehicles declined. As public bus in Phnom Penh is operated in mixed traffic, the bus travel speed is also affected by congestion.

Bus rapid transit (BRT) is regarded as an emerging attractive alternative to improve the bus operation and travel speed at a low-to-moderate cost [1-4]. However, until recently heavy investments have been made exclusively in building metro and light rail systems in Asia [4]. Levinson 2003 reviewed 26 case study cities of BRT in United States and Canada, Australia, Europe, and South America, but not including Asia. Hensher and Golob 2008 assessed 44 BRT systems in operation throughout the world, and only one, TransJakarta at Jakarta, Indonesia in Southeast Asia is included although systems in China, Japan, South Korea and Taiwan are included. Maeso-Gonzalez and Pérez-Ceron 2014 concluded that there are substantial differences between the systems at developed countries in North America, Oceania and Europe and those at developing countries in South America, Asia and Africa. The former has the technologically advanced systems while the latter gets more consolidated among the population and more developed in terms of service.

However, the systems in Asia are not comparable in size or performance to those of South America, and some systems also lack innovation and are limited to unsuccessfully adapting the operations not suitable for the local needs [4]. Although Hensher and Golob 2008 stated the cost of providing high capacity integrated system is an attractive option in many contexts, BRT scheme consists of many features and requires a large amount of budget for developing compared to ordinary bus scheme from the view point of developing countries. Thus, as Nikitas 2015 stated, it is difficult to transform a concept that is often misunderstood into new local applications even if they could genuinely improve road traffic conditions, especially at developing countries in South Asia.

From the view point of low implementation cost, there are some features of BRT system that are worth to consider for better speed improvement. They are:

Dedicated right-of-way: Bus speed is increased if bus runs in a dedicated lane. However, other possible combination of vehicles running in bus lane should also be considered for optimizing both bus and other vehicles’ speed.

Bus stop distance: At each bus stop, bus decelerates and accelerates causing delay thus it is crucial to optimize distance of bus stops to improve bus speed.

Express, limited and local services: Bus speed is increased by providing limited service bus that stops only at high demand station and leaves low demand station to local service bus.

This study aims to find appropriate measures to improve bus speed with minimum budget at Phnom Penh, Cambodia. In order to minimize cost for improving bus speed, some policies such as lane management, optimizing bus stop distance and limited bus service are considered in this study.

Kuukka-Ruotsalainen found that travel speed and punctuality are improved by bus lane by 15% to 20% and delays caused by traffic lights are reduced by transit signal priority (TSP) by 40% to 50% [5]. Two-thirds of BRT speed increases in Los Angeles were result of less stops, whereas one-thirds were of TSP [6]. Distance from bus stop to intersection affects bus travel speed. Wang concludes that far-side stop (bus stop located after-crossing intersection) is better than near-side stop (bus stop located close to before-crossing intersection) with TSP application [7]. The study was made at signal controlled intersection with/without TSP application. Bus performance can be improved by minimizing bus stops [8]. Their study claims that travel time is reduced from 3.75 to 2.69 min/km by reducing bus stops from 5 to 3.75 per kilometer and dwell time from 20 to 15 seconds. However, these studies targeted on BRT only, and the effects on other vehicles were not investigated well. In order to fully investigate the effect of BRT on other vehicles, traffic simulation models are very useful. Lan and Chang and Meng developed cellular automata (CA) models of mixed traffic flow of car and motorcycle, and the former is applied to a highway in Taiwan [9,10]. CA model is also developed by Hu for mixed traffic flow of car and electric bicycle, and applied to an arterial in China [11]. Arasan and Koshy developed a microscopic traffic simulator, HETEROSIM, and applied to the mixed traffic flow of multiple types of vehicles including bus, truck, car, motorcycle and bicycle at an arterial in India [12]. Mu and Yamamoto used a simulation package, VISSIM, to examine the mixed traffic flow of car and microcar at an urban arterial network in Japan [13]. BRT was not on the focus of the simulation analyses mentioned above, but they suggest the traffic simulation models are suitable for investigation of the mixed traffic flow. In our study, VISSIM is used as a simulation tool as it is one of the most useful and reliable simulators [14].

On the effect of BRT, Tranhuu used a simulator, SATURN, to investigate the effect of motorcycle on bus lane in Vietnam [15]. They applied the model within four distinct situations caused by the motorcycle. They are very strong level, medium level, weak level, and no violation of motorcycles in bus lane. Their research concluded that strong enforcement in general traffic can reduce the bus travel time. Arasan and Vedagiri applied HETEROSIM to investigate the effect of bus stop and exclusive bus lane in India [16]. They obtained threshold values of traffic flow for provision of exclusive bus lanes under different roadway and traffic conditions. Wei applied CA model to simulate the movement of car and bus, and evaluated the effect of TSP on vehicles’ waiting time at intersection, thus fuel consumption and exhaust emission [17].

Khan and Maini stated in their review that studies of traffic flow for non-lane- based mixed traffic conditions in developing countries are limited, and that the flow characteristics of mixed traffic is dependent on roadway geometry, traffic conditions, and static and dynamic properties of vehicles in the traffic stream. Thus, the properties of motorcycle as well as other vehicles should be well represented in the simulator for our purpose since the motorcycle is dominant in Phnom Penh, Cambodia. Several studies have been carried out to investigate the properties of motorcycle in the mixed traffic on many countries including Kenya, UK, Vietnam, Indonesia and India, but not in Cambodia [19-24].

One way to make the simulator adjust to the local context is the calibration of the parameters in the simulator, and the calibration and the validation are a critical step although sometimes informally practiced [25]. Park and Schneeberger 2003 applied a Latin hypercube sampling (LHS) design to sample an orthogonal array of parameters randomly from the entire design space, and estimated a regression model for the surface function of the target variable such as travel time with the outputs of the simulations with LHS parameters [26]. Then, the candidate parameters were evaluated for validation. LHS was also used by Park and Qi in combined with genetic algorithm (GA) for optimizing the parameters [27]. GA was used by Mathew and Radhakrishnan and Manjunatha too [28, 29]. However, the focus of the studies was a signalized intersection, and their target variable was univariate.

In our study, VISSIM simulator is applied to a road segment containing several intersections without signal control in Phnom Penh. Since the focus of our study is not only bus speed but also the effects on other vehicles, we need to validate simulated travel times of other vehicles including car and motorcycles in addition to bus. Conventional GA for univariate target is not suitable while GA for multivariate target may take more computation time than practically acceptable. Thus, we develop a structural equation model to obtain the multivariate surface function after simulation runs with LHS parameters. Then, the candidate parameter sets are evaluated for validation. Proposing the calibration process for multivariate target is one of our contributions.

Empirically, our contribution is to identify appropriate measures to improve bus speed at a low-to-moderate cost in Phnom Penh, Cambodia. Phnom Penh is in motorcycle dominant mixed traffic condition, which is typical for several South Asia as well as India. Thus, the findings of the current study can offer fundamental insights to these megacities.

Two types of data are required for this study. One is necessary for running VISSIM simulation which includes road geometry, vehicle classification, traffic volume and bus stop. Road geometry such as road length, lane width, number of lane and intersection are crucial data which are converted into network link in VISSIM. Vehicle classification clearly describes vehicle types and their size. Traffic volume defines the number of vehicles by vehicle class that are traveling within the network. Location of bus stop, bus route and bus stop marking are also needed for link input. The other data type is travel speeds by vehicle types. The travel speed is used for VISSIM parameter calibration and validation.

This study is conducted on an arterial road (Monivong Blvd.) in Phnom Penh, Cambodia. The road connects from ChroyChangvar Roundabout to Canadia Tower as shown in (Figure 1). It is bi-directional road: southbound and northbound. There are 11 junctions along 1,500 m road length. Southbound link is selected for this study.

Data on road geometry is extracted from road inventory table which was made by PPUTMP project team (JICA, 2014). The width of arterial road is 18m divided into 2 directions and each direction consists of three lanes: 2.50 m shoulder lane, 3.25 m middle lane and 3.25 m left lane. The road has 6 m width sidewalk at both sides of carriageway. Collective road width is normally ranging from 7 to 10 m. However, in this study collective roads are assumed to be 8 m carriageway with two directions, 4 m each direction.

Location and marking of bus stops along southbound road were collected. There are 4 bus stops along study road: 1. Old stadium roundabout, 2. In front of Calmet hospital, 3. In front of Ministry of Information, and 4. In front of Medical Science University as shown in (Figure 1). The bus marking size is 10 m by 2.75 m.

Figure 1: Study link.

The vehicles in mixed traffic are classified into 4 categories. The size of vehicle by vehicle class is described in (Table 1).

Table 1: Vehicle classification.

The traffic volume is acquired by visually and classified into vehicle type using manual counters by surveyors stationed on site. Two types of traffic survey have been conducted. One is road side traffic survey, and the other is intersection traffic survey. For road side traffic survey, traffic counting was conducted at the beginning of the link. Unlike road side traffic survey, traffic volume is counted by turning direction for intersection traffic survey. Traffic counting was conducted during morning peak hour (7:00AM~8:00AM) on weekdays from 2015 Dec 29th to 2016 Jan 6th. During the survey period, the weather was sunny and there was no flood or other environment impacts on traffic flow. The observed traffic volume by vehicle type at the peak hour was 7005 for motorbike, 1250 for car, and 6 for bus at the beginning of the study link (ChroyChangvar Roundabout). It clearly shows the congested flow dominated by motorbikes at the study link.

GPS devices were equipped on car, bus and motorbike to record travel data including travel time, travel speed and distance. The measurement was made three times for each vehicle during peak hour. The results of travel speed survey are summarized in (Table 2).

Table 2: Observed travel speed.

The simulation framework in this study consists of three main parts: a simulation model, a model calibration and a model validation.

Simulation Model

Tool for microscopic simulation of the traffic flow used in this study is VISSIM 5.40. It is capable to analyze private and public transport operations under constraints such as lane configuration, vehicle composition, and public transport stops. It is a powerful tool for the evaluation of various alternatives based on transportation engineering and planning measures of effectiveness. Car following model integrated in VISSIM is psycho- physical driver behavior model developed by Wiedemann [30, 31]. The model concept is that the driver decelerates to speed of preceding vehicle running at slower speed and accelerates gradually again after passing the vehicle. The model has been calibrated through multiple field measurements at the Technical University of Karlsruhe since 2009

KIT: Karlsruher Institut fur Technologie), Germany. Calibration of parameters can be found in calibration procedure.

In order to assure that a simulation model is usable, it’s necessary to make the simulation produce the result matching to real condition at a significant level. Some inputs that explain driver behavior and vehicle characteristics are not easy to collect on site thus a calibration is required. Calibration is a process that involves in parameters adjustment which interacts traffic flow in simulation more likely to real traffic flow.

Calibration Flow

Calibration flow is explained as the following. Every simulation model has an uncalibrated parameter set inherent to the simulation model that is so called default parameter set. It may produce satisfying result in some cases but it is impossible to guarantee. However, if the default parameter produces satisfying result, then it is fine to skip the calibration and validation procedure and use the simulation model with the default parameter set for further analysis. This advantage reduces a huge amount of time and effort by skipping the main calibration procedure. After the simulation model setup process, it’s worth to try the simulation with default parameter set. If the default parameter set is not acceptable, calibration and validation procedures need to be conducted. It is crucial to select parameters that produce the effective result matched to real traffic conditions, generally they are known as driver’s behavior parameters and vehicle performance parameters. Range of each selected parameter is defined based on traffic data or user experience. Suppose we have 5 parameters and each parameter has 5 levels thus total number of simulation run is 55 = 3125 times. It consumes much time for running simulation therefore in order to reduce the number of combinations of parameter set into a reasonable level while still reasonably covering the entire parameter surface, LHS algorithm is applied. LHS is one of space-filling methods that maximally cover space.

Multiple times of simulation runs are conducted. The results from simulation are then used for feasible study. Feasible study is conducted in order to identify whether parameter sets produce the results matched to values observed by field survey or not. In case they don’t produce satisfying results, range of parameter sets has to be redefined and return to experiment design process stage again.

If the range of parameter sets is feasible, calibrated model is created and solved for candidate parameter set. After that, multiple times of simulation runs with calibrated parameter set are conducted. The results for simulation runs with calibrated parameter set are then used for model validation.

Simulation Scenarios

Vehicles are set to run in heterogeneous traffic. Motorbikes are able to take over other vehicles driving on the same lane from left or right. The bus is assumed to stop each bus stop and the average dwelling time is assumed as 20s. When the space is available, vehicles can freely change the lane. Stopping for alighting/pick up is made on the network and abrupt lateral vehicles. Therefore, several virtual parking lots are created on the network to represent vehicles stopping for alighting/pick up on road side.

The desired speed distribution affects roadway capacity and achievable travel speeds. The desired speed is the speed that driver desires to travel when there are no obstacles. It is not necessarily the speed at which the vehicle travels in the simulation. The driver will travel at his/her desired speed (with a small stochastic variation called oscillation) when there are no abruptions from other vehicles. Any vehicles with a higher desired speed than its current travel speed will check for the opportunity to pass without endangering other vehicles. Minimum and maximum values can be entered in VISSIM for the desired speed distribution. According to Cambodia traffic law, four wheel vehicles are allowed to drive up to 40km/h while two wheel vehicles are allowed to drive at maximum speed of 30km/h. However, in reality two wheel vehicle drivers are seeking to drive at maximum speed of 40km/h. Therefore, in this study all vehicles are assumed to have desired speed ranged 35

~ 40km/h except some motorbike taxis that intend to travel slower seeking for their customer on road side. Therefore, in order to represent those motorbike taxis on the link, 10% of motorbikes are assumed to travel at the desired speed distribution of 12 ~ 15km/h. Conflict area is intersection where decision of priority has to be made in order to avoid traffic congestion. The priority of vehicle can be determined by signal base or non- signal base. This study focuses on non-signal based intersection by using right side priority rule. The vehicles traveling on the study road always have right of way while vehicles traveling in access road yield. The vehicles attempting to turn left or right have to follow right side priority rule. For example, suppose vehicle A traveling to north and vehicle B traveling to west. If both vehicles arrive at intersection at the same time, vehicle B has right of way and vehicle A yields. It is assumed that there is no traffic abruption caused by pedestrian crossing the road or walking at road side.

Parameter Set

Trial simulation is conducted to test if default parameter can produce satisfying results. The results from trial simulation show that the average travel times of car, bus and motorbike are 424s, 437s and 416 respectively while the average travel time from field data of car, bus and motorbike are 394s, 441s and 305s. Therefore, parameter calibration procedure is necessarily conducted.

Driving behavior composes of car following model, lane changing model, and lateral model. Each model consists of a collective parameter. For car following model, as mention already, Wiedemann (1974) is used in VISSIM and it has been calibrated through multiple field measurements. The parameters of car following model and lane changing model have a strong influence on traffic flow thus they are selected for calibration. For Wiedemann 74 model, safety distance is computed by

???? = ???????? + (????????add + ????????mult × ????) × √???? Where ax is average standstill distance, bxadd is additive part of desired safety distance, bxmult is multiplicative part of desired safety distance, z is a value of range [0, 1] which is normal distributed around 0.5 with a standard deviation of 0.15, and v is the vehicle speed [m/s].

Ranges of parameter values selected for calibration are shown in (Table 3). On the car following model, average standstill distance and additive part of desired safety distance are selected for calibration. Average standstill distance is the distance between stopped cars. The default value for average standstill distance is 2.0 m. For motorbike, this parameter value is set to 1.0 m and for other vehicles, the parameter value ranges from 1.0 m to 4.0 m. Additive part of desired safety distance has a large effect on computation of the safety distance. The default value of additive part of desired safety distance is 2.0m. For motorbike, this parameter value is set to 1.0 m and for other vehicles, the parameter value ranges from 1.0 m to 4.0 m. On the lane changing model, look ahead distance and minimum headway are selected for calibration. Look ahead distance is visibility distance that vehicle can see forward and begin to attempt to change lane. The default value is 0 m to 250 m. As the desired speed distribution is low in the study area, traffic density is high thus headway vehicles are closer. Therefore, look ahead distance should be less than 250 m. The value of look ahead and back is assumed to be the same, and look ahead distance value range is adapted to 0m to 100 m. The minimum headway distance defines the minimum distance to the vehicle in front that must be available for a lane change. The default value is 0.5 m. The value of parameter range in the case study is between 1 m and 5.0 m for all vehicles. The default value appeared too small for vehicles to attempt a lane change. It did not appear realistic that a vehicle would attempt a lane change given headway of 0.5 m. As a result, larger values are assumed to be more reasonable. Other parameters are set to default value.

Table 3: Parameters for calibration.

Simulations Based On Experimental Design

124 parameter sets are generated from LHS within parameter range. Correlation among parameters is calculated to ensure that the relationship between each parameter pair is minimized. The calculated correlation coefficients are very low, at most 0.186 in the absolute value. Therefore, the surface of the parameter is considered to be adequately covered. Five random seeded runs are performed in VISSIM for each of the 124 parameter set that translates into 620 runs. The simulation runtime is 1hour. Simulation output is collected in 4 stages: 1st at 15mn, 2nd at 30mn, 3rd at 45mn and 4th at 1hour of simulation run. The average travel times of car, bus and motorbike of southbound link are collected for each of the 620 runs. The average travel time for each of 124 parameter sets is derived from calculating average of five results of multiple runs.

A structural equation model is developed with calibration parameters as independent variables and average travel times of car, bus and motorbike as dependent variables. SPSS Amos is used for parameter estimation. During the estimation process, minimum headways of four wheels vehicle and motorbike are found to have low t-statistic values in all three equations, meaning not statistically significant, so discarded from the model. The results of the estimation is shown in (Table 4). The results show that the proposed model is acceptable in terms of chi-square value (c2), goodness of fit index (GFI), adjusted GFI (AGFI), and root mean square error of approximation (RMSEA). The error covariances are estimated as significant, which means the structural equation modeling provides efficient parameter estimates rather than estimating each equation separately. The results also show that R-squares are high for motorbike and car while low for bus. It means that the variance of travel times of motorbike and car are well represented by the functions, but the function for the bus accounts only for 30.3% of the variance. The reason for this low goodness-of-fit for the bus is the small sample size of bus travel time for each run: only four bus travel times are collected in each simulation run according to the bus schedule. On the effects of explanatory variables on the travel time, average standstill distance and additive part of desired safety distance have positive parameter estimates for car and bus while negative for motorcycle. It is because the average standstill distance and additive part of desired safety distance are varied from 1.0 to 4.0 m only for car and bus while fixed as 1.0 m for motorcycle. The results suggest that the longer standstill distance and the longer additive part of desired safety distance cause longer travel times for car and bus, and that motorcycles can utilize the longer headways of car and bus to overtake, which contributes the shorter travel time of motorcycle. Comparing between car and bus, the results show that bus has smaller parameter estimates for both average standstill distance and additive part of desired safety distance than car. One of the reasons is that bus stops at bus stops even if there are no vehicles in front, which makes the effect of the headway on bus smaller than that on car. Look ahead distance has negative parameter estimates for all three vehicle types. It is reasonable since the longer look ahead distance makes the lane change smoother, which contributes shorter travel time. The parameter estimate for car is smaller in the absolute value than those for motorcycle and bus, which means the look ahead distance has small effect on car than motorcycle and bus, but the reason is not clear.

Table 4: Estimated surface function^a.

Error covariance: car and bus, 6.747 (3.82); car and motorbike, 4.702 (4.52); bus and motorbike, 8.791 (3.33)

Model fit: c2 = 2.880 (df = 6), p = 0.824, GFI = 0.994, AGFI =

0.965, RMSEA = 0.000 ^at statistic is shown in parenthesis.

Validation

Validation is conducted by comparing travel time of car, bus and motorbike acquired from field survey with the results produced by calibrated model. Root mean square error (RMSE) and root mean square error percentage (RMSEP) are used as measurement tools for estimating the value of error. Matlab is used to generate the candidate parameter sets according to the estimated simultaneous equations shown in Table 4 within corresponding parameter ranges. We didn’t use the equation for the bus because of the low goodness-of- fit. Seven sets of candidate parameter are derived from the model, and five random seeded runs are performed for each candidate parameter set.

The seven candidate parameter sets and the results of the validation are shown in (Table 5). The results suggest that Set 1 gives the smallest RMSE and RMSEP for motorbike and bus travel time while Set 2 gives the smallest for car. However, Set 2 has the largest RMSE and RMSEP for motorbike and the second largest for bus. Moreover, the deviation of simulated motorbike travel time by Set 2 from the observation is statistically significant at 95% confidence, and that of bus travel time is at 90%. It means that Set 2 is not acceptable. Thus, Set 1 is used for the evaluation of bus stop and lane management in this study. Looking at the values of parameters in Set 1, the average standstill distance is set at 1.7 which is smaller than the default value, 2.0 while the additive part of desired safety distance is set at 3.7 which is larger than the default value, 2.0. It is consistent with our field observation that vehicles make a queue with a smaller headway when stopping or at lower speeds while take a longer headway at higher speeds for a larger heterogeneous traffic condition. The look ahead distance is set at 38, which is much smaller than the default value, 250. It is also consistent with our expectation that the traffic density is so high that the look ahead distance becomes shorter. Looking at RMSE and RMSEP of Set 1, it is clear that travel times for motorbike and car are very well represented by the simulation since RMSEP is only about 0.7% and 0.5% respectively and much better than that for bus while RMSEP at 2.7% for bus is still acceptable. The reason for the lower goodness-of-fit for bus may be the same as the low R-squares: the small sample size of bus travel time for each run, and that the equation for the bus was not considered to generate the candidate parameter sets.

Table 5: Evaluation of candidate parameter set^a

^aDifference from the observation is significant at the 90% (+) and 95% (*) confidence interval.

The calibrated simulation model is used for evaluating bus travel time in various bus stop and lane management schemes. The effects of bus stop distance, express, limited and local services, and dedicated right-of-way are evaluated in this study. In order to evaluate bus stop distance and express, limited and local services, the number of bus stops in the simulation is decreased from four to three (Case 1) and two (Case 2). The boarding and alighting time of the skipped bus stops are distributed equally to the remaining bus stops in order to keep the total boarding and alighting time within the segment the same.

On the other hand, in order to evaluate the dedicated right-of- way the bus lane is introduced into the curb side lane in the simulation as dedicated for bus only (Case 3), for bus and car (Case 4) and for bus and motorcycle (Case 5). Other two lanes are occupied by all vehicle classes. The current lane width near curb side is 2.50 m which is shorter than bus stop marking width (2.75 m), so the lane width is redistributed. The current widths of the other two lanes are 3.25 m, so the lane width is set as 3.0 m equally for the three lanes in the simulation. The redistribution of lane width is applied for the cases investigating the effects of the number of bus stops as well as the cases for bus lanes to make them comparable.

The results of the simulation for each case are shown in (Table 6). For the case with the decreased number of bus stops from four to three (Case 1), bus stop No. 3 in Figure 1 is skipped, and the results suggest that the improvement of bus travel time is not very significant although the deceleration and acceleration time is reduced for a deactivated bus stop. Also, the travel times of motorcycle and car doesn’t change significantly. Similar results are obtained for the case with further decrease in the number of bus stops to two by skipping bus stops No. 1 and 3 in Figure 1 (Case 2). These results suggest that the reasons for the delay in the bus travel time is not related to the deceleration and acceleration at the bus stop, and that it is no possible to reduce the bus travel time significantly by introducing the bus stop distance, or express, limited and local services on this road segment.

For the cases with the bus lanes, the effects on the travel times for bus as well as motorcycle and car are statistically significant. When the bus lane is introduced into the curbside lane as dedicated for bus only (Case 3), the travel time of bus is reduced by 49.14% compared to the current situation, which is highly statistically significant. However, the travel times of motorcycle and car increase by 33.79% and 14.41% respectively compared to the current situation. It is because the motorcycles and cars have to take over each other to maximize their desired speed in narrower space provision of two lanes only.

When the bus lane is dedicated to bus and car (Case 4), the travel times of car as well as bus are statistically significantly reduced by 13.96% and 20.98% respectively. However, the travel time of motorcycle increases by 15.06%. The results suggest that the two lanes are insufficient to accommodate the traffic volume of motorbikes. On the other hand, the travel time of car improves in this case much more than that in other cases. When the bus lane is dedicated to bus and motorcycle (Case 5), the travel time of bus is reduced by 32.82% from the current situation. The improvement is smaller than the case with the bus lane dedicated to bus lane (Case 3), but much larger than the case with the bus lane dedicated to bus and car (Case 4). Moreover, the travel times of both motorcycle and car are reduced. The improvement is statistically significant for both motorcycle and car although the size of the improvement is much smaller than that of bus. The results suggest that motorcycles can change lanes from the second and third lanes from the curbside to the first lane to maximize their speed in this case. Thus the density on the second and third lanes become lower which facilitates cars running on these lanes faster, too. The results suggest that the bus lane dedicated to bus and motorcycle (Case 5) gives a palate improvement, so is the most acceptable scheme to improve bus travel time in this empirical analysis.

Table 6: Travel time by case.

^aStandard error is shown in parenthesis. Difference from the current case is significant at the 90% (+), 95% (*), 99% (**) and 99.9% (***) confidence interval.

The strategies on the design and operations of bus lane during peak hours were examined for the heterogeneous traffic flow condition with dominant motorcycles in this study. The operations of bus in different settings of lane management and bus stops distance were investigated in order to improve bus travel speed. A simulation model was developed to represent the heterogeneous traffic flow on a corridor in Phnom Penh, Cambodia.

VISSIM simulator was applied to a road segment containing several intersections without signal control in Phnom Penh. Since the focus of our study was not only bus speed but also the effects on other vehicles, we developed a structural equation model to obtain the multivariate surface function after simulation runs with LHS parameters. LHS parameters provided sufficient variations for parameter settings with very low correlation coefficients, which is desirable for estimation of structural equation models. The structural equation model identified the key explanatory variables and sufficient goodness-of-fit statistics for motorcycle and car. We found that average standstill distance except motorcycle, additive part of desired safety distance except motorcycle and look ahead distance affected the simulated travel times significantly in our settings. The candidate parameter sets were evaluated for validation, and then the combination of shorter average standstill distance, longer additive part of desired safety distance and shorter look ahead distance gave the smallest RMSE and RMSEP. The results are consistent with the field vehicles make a queue with a smaller headway when stopping or at lower speeds while take a longer headway at higher speeds for a larger heterogeneous traffic condition. It is also consistent with our expectation that the look ahead distance becomes shorter at high traffic density. The results of the simulation on the evaluation of the number of bus stops and the bus lanes suggested that the bus lanes significantly affect the travel time of bus while the effects of the number of bus stop is not significant. For the cases with the bus lanes, the effects on the travel times for motorcycle and car as well as that for bus are statistically significant. When the bus lane is introduced into the curbside lane as dedicated for bus only, the travel time of bus is reduced, but the travel times of motorcycle and car increased. The results are consistent with the literature. Arasan and Vedagiri 2010 in their simulation study on a corridor in India found that the running speeds of car and motorcycle decreased when the bus exclusive lane was introduced while that of bus significantly increases at the traffic conditions with high demand capacity ratios. When the bus lane is dedicated to bus and motorcycle, the travel times of motorcycle and car as well as bus are reduced. The results suggest that the bus lane dedicated to bus and motorcycle gives a palate improvement, so is the most acceptable scheme to improve bus travel time in this empirical analysis. The results are different from Tranhuu 2007 who found that the bus lane with poor enforcement made the bus travel time higher than that in general traffic. They considered four levels of traffic enforcement ranging from the case where motorcycles violate the bus lane very often to no violation of motorcycles in bus lane. The bus lane dedicated to bus and motorcycle in our study is similar to the case with lowest enforcement on motorcycle in Tranhuu, but our results suggest that the bus lane dedicated to bus and motorcycle improve the travel times of bus, car and motorcycle. There are potential reasons for this inconsistency. At first, the traffic conditions between the two studies are different although both are South Asian countries. Our study investigates the heterogeneous traffic in Phnom Penh, Cambodia while they investigated that in Hanoi, Vietnam. Both traffic conditions are dominated by motorcycles, but the ratios of motorcycles to total volume are different as well as total traffic volumes are different, so the effects of the shared use of bus lane with bus and motorcycle must be dependent on those traffic conditions. Also, our study concentrates on one corridor and applies a microscopic traffic simulator, VISSIM, while they considered a network with more than 1,000 links and applied a mesoscopic traffic simulator, SATURN. The microscopic simulator is believed to be better to represent the heterogeneous traffic flow in detail, but the overall effects on traffic network is not considered in this study, which is one of our limitations.

The first author would like to thank Japanese Grant Aid for Human Resource Development Scholarship (JDS) for the financial support to study in Japan.

1. Levinson HS, Zimmerman S, Clinger J, Gast J. Bus rapid transit: Synthesis of case studies. Transportation Research 2003; 1841: 1-11.
2. Hensher DA, Golob Bus rapid transit systems: a comparative assessment. Transportation. 2008; 35: 501-518.
3. Maeso-Gonzalez E, Perez-Ceron P. State of art of bus rapid transit transportation. European Transport Research Review. 2014; 6: 149-156.
4. Nikitas A, Karlsson A worldwide state-of-the-art analysis for bus rapid transit: looking for the success formula. Journal of Public Transportation. 2015; 18: 1-33.
5. Kuukka-Ruotsalainen V, Airaksinen S, Lehmuskoski M, Musto M, Murole P. Speeding up public transportation as a central competitive factor. Liikenne-Ja Viestintaministerion Julkaisuja (Finland Ministry of Transport and Communications). 2007; 53:
6. Levinson HS, Zimmerman S, Clinger J, Rutherford S, Smith R, Cracknell J. et al. TCRP Report 90: Bus Rapid Transit: Case Studies in Bus Rapid Transit and, Implementation Guidelines, Transportation Research Board of the National Academies, Washington DC. 2003.

7. Wang R, Seo M, Nakamura F, Okamura T, Tanaka S. A performance evaluation of bus stop placements near a signalized intersection by a microscopic traffic simulation. Journal of the Eastern Asia Society for Transportation 2013; 10: 1270-1280.
8. Levinson Analyzing transit travel time performance. Transportation Research Record. 1983; 915: 1-6.
9. Lan LW, Chang Inhomogeneous cellular automata modeling for mixed traffic with cars and motorcycles. Journal of Advanced Transportation. 2005; 39: 323-349.
10. Meng JP, Dai SQ, Dong LY, Zhang JF. Cellular automaton model for mixed traffic flow with motorcycles. Physica A. 2007; 380: 470-480.
11. Hu X, Wang W, Yang Mixed traffic flow model considering illegal lane- changing behavior: simulations in the framework of Kerner’s three-phase theory. Physica A. 2012; 391: 5102-5111.
12. Arasan VT, Koshy RZ. Methodology for modeling highly heterogeneous traffic Journal of Transportation Engineering. 2005; 131: 544-551.
13. Mu R, Yamamoto T. Analysis of micro-cars’ influence on traffic network using a microscopic simulator. Journal of Transportation Systems Engineering and Information 2013; 13: 44-51.
14. Rrecaj AA, Bombol KM. Calibration and validation of the VISSIM parameters - state of the art. TEM Journal. 2015; 4: 255-269.
15. Tranhuu M, Montgomery F, Timms P. Modeling bus lane priorities in a motorcycle environment using Transportation Research Record. 2007; 2038: 167-174.
16. Arasan VT, Vedagiri P. Microsimulation study of the effect of exclusive bus lanes on heterogeneous traffic flow. Journal of Urban Planning and 2010; 136: 50-58.
17. Wei L, Zhang L, Wang Z. A cellular automata model on bus signal priority strategies considering resource Practical Applications of Intelligent Systems. 2014; 279: 689-698.
18. Khan SI, Maini Modeling heterogeneous traffic flow. Transportation Research Record. 1999; 1678: 234-241.
19. Oketch T. New modeling approach for mixed-traffic streams with nonmotorized Transportation Research Record. 2000; 1705: 61-69.
20. Lee TC, Polak J, Bell MGH. New approach to modeling mixed traffic containing motorcycles in urban Transportation Research Record. 2009; 2140: 195-205.
21. Cao NY, Sano Estimating capacity and motorcycle equivalent units on urban roads in Hanoi, Vietnam. Journal of Transportation Engineering. 2012; 138: 776-785.
22. Nguyen LX, Hanaoka A microscopic simulation model for motorcycle traffic safety assessment in Vietnam. Proceedings of Infrastructure Planning 45 CD-ROM. 2012
23. Ambarwati L, Pel AJ, Verhaeghe R, van Arem B. Empirical analysis of heterogeneous traffic flow and calibration of porous flow model. Transportation Research Part C. 2014; 48: 418-436.
24. Kanagaraj V, Asaithambi G, Toledo T, Lee TC. Trajectory data and flow characteristics of mixed traffic. Transportation Research Record. 2015; 2491: 1-11.
25. Sacks J, Rouphail NM, Park B, Thakuriah P. Statistically- based validation of computer simulation models in traffic operations and management. Journal of Transportation and 2002; 5: 1-24.
26. Park B, Schneeberger Microscopic simulation model calibration and validation: case study of VISSIM simulation model for a coordinated actuated signal system. Transportation Research Record. 2003; 1856: 185-192.
27. Park B, Qi H. Development and evaluation of a procedure for the calibration of simulation Transportation Research Record. 2005; 1934: 208-217.
28. Mathew TV, Radhakrishnan Calibration of microsimulation models for nonlane-based heterogeneous traffic at signalized intersections. Journal of Urban Planning and Development. 2010; 136: 59-66.
29. Manjunatha P, Vortisch P, Mathew Methodology for the calibration of VISSIM in mixed traffic. Transportation Research Board 89th Annual Meeting, Compendium of Papers DVD. 2013.
30. Wiedemann Simulation des Strassenverkehrsflusses. Schriftenreihe des Instituts fur Verkehrswesen der Universitat Karlsruhe, Band 8, and Karlsruhe, Germany. 1974.
31. Japan International Cooperation Agency (JICA) The project for comprehensive urban transport plan in Phnom Penh capital city (PPUTMP) Final 2014.