Calibration of Vehicle and Driver Characteristics in VISSIM and ANN-based Sensitivity Analysis

Congestion and traffic delays in urban road networks have become a characteristic feature of present cities. In developing economies, the heterogeneous mix of slow and fast-moving vehicles have further contributed to the chaos resulting in roads operating below existing capacities. Traffic engineers and transport planners often resort to traffic flow modeling techniques in order to evolve strategies for effective transport management. In this context, simulation-based approaches are found to be more effective in modeling the traffic flow on roads. Traffic flow modeling using micro-simulation approaches facilitate the study of bottlenecks and assists in the analysis of traffic flow characteristics, the movement of individual vehicles, and the study of vehicle and driver characteristics.

The present study focuses on performing investigations on assessing the influence of vehicle and driver characteristics on the prediction of traffic volumes at 18 mid-block sections of the road network in Mangalore city for VISSIM-based microsimulation modeling of heterogeneous traffic conditions. A 16-minute video-graphic data with information on traffic flow and speed at important junctions, in addition to speed and volume data on 18 mid-block sections of the city for the evening-peak hour were used in the analysis. Out of the 16-minute video-graphic data, information pertaining to 12-minutes was used in the calibration procedure, while the remaining data was used in the validation of VISSIM model.

The initial phase of this study includes preparation of a high-resolution map of the city to a scale of 1:5000 using Gimp (a high-end image processing software), which was imported to AutoCAD and later superimposed with an AutoCAD drawing of the road-junctions and linking-roads of the city. The AutoCAD drawing and the scaled image were then exported to VISSIM. Preliminary modeling in VISSIM was performed using default values of vehicle and driver characteristics to determine the best random seed. For the existing network and the existing traffic conditions in the city, the best random seeds were identified as 25 and 42. The simulated and observed values of traffic volume were compared using the GEH statistic originally developed by Geoffrey E. Havers. This approach has been recommended for modeling studies by several agencies (DMRB, 1996; Smith and Blewitt, 2010; UKDoT, 2014; Wisconsin, 2005; WSDoT, 2014).

In the next phase, the multi-stage first-level of calibration trials were conducted using random seed 42 in VISSIM starting with default values of vehicle and driver characteristics, followed by testing of various combinations of the same. A number of combinations of values assigned to vehicle characteristics such as the minimum lateral clearance, acceleration distributions, deceleration distributions, the desired speed distributions were analysed and fine-tuned in addition to driver characteristics such as the minimum look-ahead distance, minimum look-back distance, additive part of desired safety distance, multiplicative part of desired safety distance, average standstill distance, minimum longitudinal speed (for lateral movement), time-between direction changes, and minimum collision time-gain (for lateral movement to the adjacent lane).

The multi-stage first-level of calibration trials were performed in 14 major stages with sub-stages at the 2nd, 3rd, and the 7th stages such that the traffic scenario of roads in Mangalore city could be represented to a higher degree of accuracy. At each stage of calibration, an attempt was made to improve the accuracy of predicting traffic volumes at 18 mid-blocks of the city by verifying the mean absolute error (MAE) in addition to the GEH statistic as an indicator.

Subsequently, information on the vehicle and driver characteristics used in each stage of the first-level of calibrations and the details on the prediction errors for various trials were given as input to an Artificial Neural Network (ANN) model. The ideal configuration of the ANN model was arrived at based on a number of trials performed in EasyNN-plus software. Using the details on the connection weights, an ANN-based sensitivity analysis was performed using the approach recommended by Özesmi and Özesmi (1999).

Based on the sensitivity analysis performed, the relative importance of input variables was computed. Here, it was observed that the characteristics such as the average standstill distance, minimum look-ahead distance, and the desired speed: lower bound values assigned for speed distributions of each vehicle type were found to have a major influence on the accuracy of predictions in microsimulation modeling using VISSIM. The second-level of calibrations were performed by fine-tuning the values of the characteristics that were found to be highly sensitive in the sensitivity analysis and the prediction errors were found to decrease further. The VISSIM model was validated using the remaining four-minute video-graphic data for traffic flows on 423 links of the major road network of the city covering an area of about 3 square km and the prediction errors were found to be within tolerable limits.

The proposed two-level multi-stage calibration approach with an ANN-based sensitivity analysis introduced in between in the present study is expected to provide traffic engineers and planners with a novel framework for the development of reliable microsimulation models to be considered. The approach discussed above will provide the basis for evolving strategies for improving the prediction of traffic flows in the road network for better traffic management.

VISSIM is a powerful micro-simulation tool that can perform modeling and analysis of traffic flow on urban roads, as well as on inter-urban motorways. A number of studies have been performed using microsimulation techniques in modeling urban traffic.

Fellendorf and Vortisch (2001) calibrated and validated the VISSIM model for freeways in Germany and US by modeling driver behavior where the traffic regulations were very much different. The main focus of this study was to investigate the longitudinal movement of cars as part of the car following model by simulating various types of driver behavior. The study was performed considering traffic conditions with flows up to 2400 vehicles/h/lane characterized mainly by the movement of cars that follow a strict lane-discipline for freeways. Mosseri et al. (2004) performed studies on evaluating traffic management strategies at the Lincoln Tunnel Corridor in New Jersey focused on traffic movement on high-occupancy lanes, exclusive bus lanes, and toll lanes.

Park and Schneeberger (2003) suggested a nine-step procedure for testing, calibration, and validation of a VISSIM model based on studies on Lee Jackson Memorial Highway, Fairfax, Virginia. The authors found it essential to perform simulation trials using field data such as speed and volume for freeways in order to calibrate Wiedemann-99 (1974) car-following characteristics such as start-up lost time, sensitivity related to car-following, the time required to perform lane-change operations, acceptable gaps, and driver behavior that cannot be easily measured on field.

Dorothy et al. (2006) modeled a large-scale transport network in Columbus, Ohio using VISUM and VISSIM adopting measures of effectiveness such as queue length, average delay, and fuel emissions. Mathew and Radhakrishnan (2010) devised an approach to test and calibrate models simulating heterogeneous flows at road junctions. The vehicle and driver parameters to be calibrated were identified using a sensitivity analysis and the values of these were further optimized using the Genetic Algorithm (GA) to ensure lesser prediction errors. Data collected from two intersections in Trivandrum City and one intersection from Jaipur City was used for this purpose.

Siddharth and Ramadurai (2013) performed investigations on ANOVA-based sensitivity analysis and derived the optimal values of characteristics using GA while automatically calibrating a VISSIM model for heterogeneous traffic conditions for IT-Corridor State Highway in Chennai. Lu et al. (2016) adopted a video-based method for calibrating car-following characteristics in VISSIM for studies performed in Waterloo, Canada. Bede et al. (2017) performed studies on modeling look-ahead controlled vehicles moving at cruising speeds. In these vehicles, the speed is optimized based on the need to minimize energy consumption, travel-time for various terrain conditions, and speed limits. The study was performed for various classes of vehicles.

Karakikes et al. (2017) developed a VISSIM model to analyse a network of 500 links and 113 nodes in Bavaria based on travel-time determined using Bluetooth detectors. The methodology adopted in calibration and validation ensured accuracy of 96.5% for both cars and heavy vehicles. Sadat and Celikoglu (2017) developed a VISSIM model to study the impact of variable speed limits on Istanbul Freeway D100 over a distance of 5.2km. MATLAB was used to arrive at the variable speed limits based on volume, occupancy, and average speed. In this study, the vehicular movements in the network were monitored using the Remote Traffic Microwave Sensor (RTMS) provided by the Istanbul municipality.

Srikanth et al. (2020) conducted studies on lane-change behavior for multi-lane divided highways using VISSIM where it was observed that the number of lane-changes depends mainly on the traffic volume and the number of lanes along the direction of travel. The study also analysed lane-changes for traffic operating at maximum capacity. It was also found that random seed 40 performed much better than random seeds 41, 42, 43, and 44.

In the studies mentioned above, it can be seen that not many comprehensive studies covering the entire road network of cities have been performed especially in the case of heterogeneous traffic flow conditions.

In the present study, it was found necessary to perform simulation trials based on observed video-graphic data for calibration of Wiedemann-74 (1974) car-following model characteristics such as the minimum collision time-gain, time-between direction changes, minimum longitudinal speed, additive part of desired safety distance, multiplicative part of desired safety distance, minimum look-ahead distance, minimum look-back distance, minimum lateral clearance, and driver behavior, that are not easily measured in order to improve the prediction of traffic flows on the urban road network of Mangalore city. Additionally, traffic in Mangalore city is further characterized by a heterogeneous mix of slow and fast motorized vehicles and poor observance of lane-discipline similar to emerging tier-II cities in India.

In the multi-stage two-level calibration approach performed in this study, the values of the characteristics were varied one at a time in a manner similar to the approach adopted by Mathew and Radhakrishnan (2010). However, the sensitivity analysis is performed using an ANN-based approach in the present study in contrast to a GA-based sensitivity analysis performed by Mathew and Radhakrishnan (2010) and Siddharth and Ramadurai (2013).

As part of this study, it was also proposed to adopt the use of measures of effectiveness such as the GEH statistic value, the mean absolute errors (MAE), and the RNSE values in order to indicate the accuracy of simulations based on observed and measured volumes of vehicles passing through 18 mid-block sections of the road network. The existing guidelines for calibration of microsimulation models formalized by Dowling et al. (2004) and Park et al. (2005), cater to the simulation requirements of homogeneous traffic conditions. The present study will assist in evolving the framework for heterogeneous traffic.

Mangalore is a beautiful scenic city in the southern part of the State of Karnataka, India. It is an emerging city located at 12°51'55.4" N latitude and 74°50'24.7" E longitude, at an altitude of 22 m above sea-level, and is well-connected by air, rail, and road. Mangalore City Corporation (MCC) witnessed an increase in population from 195,000 in 1971 to 488,968 in 2011 (Census, 2011) and is spread over 60 urban wards, covering an area of about 132.45 square km (MCC, 2015). The current population in the city is about 601173 (GoK, 2013), with a total road length of 1133.31km (MCC, 2016).

The heterogeneous mix of vehicles in the city includes cars, light commercial vehicles (LCVs) with a gross vehicle weight of lesser than 7.5 metric tons, buses, auto-rickshaws, and motorized two-wheelers. The average annual increase in the total number of vehicles during 2012-18 was about 10.86%, while the total vehicle count stood at 676,920 in March 2018. Presently, the percentage of cars, motorized two-wheelers, auto-rickshaws, LCVs, buses, and trucks comprise 19%, 71%, 4%, 2%, 1%, and 3% (GoK, 2018).

The present study included 423 links of road network spanning over 3.08 square km of Mangalore city connecting major junctions such as Hampankatta, Navbharat, PVS, Bunt’s Hostel, Jyothi, Bendoorwell, Balmatta, and St. Theresa’s School Junction. A few of these junctions are provided with fixed time traffic signals with manual override options to handle peak-hour traffic, while a majority of the junctions are manually controlled.

Video-graphic data on vehicular movements at eight major traffic junctions of the city and at 18 important mid-block sections of the road network were recorded for the peak-hour between 17.30-18.30 hours, during 10-18 March 2015 on mid-week days. A net total of 16-minutes peak-hour data was considered for the development of VISSIM model. A 12-minute video-graphic data that constitutes 75% of the data was used for testing, development, and calibration of VISSIM model, while a four-minute video-graphic data that constitutes 25% of the data was used for model validation.

As part of data collection efforts, additional video-graphic recordings of vehicular movements were used to determine the approximate speeds of randomly selected vehicles on selected road sections of the city. The acceleration values were also computed based on the initial and final speeds on road sections. It was proposed to use part of this information in assigning the desired speed distributions and the desired acceleration distributions.

The following sub-sections provide descriptions on the methodology adopted in various phases of work undertaken in this study. Details on the creation of a high-resolution map of the city, prelimCinary steps in model development in VISSIM, measures of effectiveness adopted for the study, identification of the best random seed, first-level of calibrations, ANN-based sensitivity analysis, second-level of calibrations, and validation of the model are discussed below.

5.1. Creation of a High-resolution map for the city

An AutoCAD drawing with details on the road geometrics, widths of the road links, and dimensions of the junctions was prepared based on measurements made at various locations in the city. For the road-stretches of the city that remained unchanged for the past decade, the information contained in drawings prepared previously by Dalal Consultants and Engineers (2003) were used to update AutoCAD drawing in 2015.

Additionally, a high resolution .jpg digitized base-map (of 6284 x 6726-pixel size and 3.43MB size) of the city was developed to a scale of 1:5000. AutoCAD drawing prepared earlier in 2015 was then overlaid on the digitized base-map in AutoCAD, and re-aligned to ensure a perfect match. The .jpg base-map along with the re-aligned AutoCAD drawing was then saved as a composite .dwg image as shown in Figure 1.

Figure 1

Download asset Open asset

5.3. Procedure adopted for identification of the best random seed for simulation

Preliminary trial runs using random seeds 1 to 50 were performed to identify the most ideal random seed for simulation as part of the present study. Here, it was observed that random seeds 25 and 42 provided lesser prediction errors for simulations in Mangalore city. Figure 2 illustrates details on the results of trials conducted using random seeds 25, 28, 42, and 50 where the mean absolute errors (MAE) in the prediction of traffic flows at 18 mid-block sections were found to be lower. Simulation experiments reported by various researchers (Wang et al., 2012; Ishaque and Noland, 2005; Kamdar, 2004; TJPDC, 2007; Arkatkar et al., 2016) based on numerous trials also indicate that the use of random numbers 25, 28, 42, and 50 provide reliable estimates of the actual values.

Figure 2

Download asset Open asset

The trial runs indicated that simulations using random seeds 42 and 25 resulted in lower mean absolute errors (MAE) of 33.4% and 33.4% respectively when simulated with default values in VISSIM. Also, the corresponding GEH statistic values were computed as 7.06 and 7.02 respectively for random seeds 42 and 25. However, the random seed 42 was used for calibrations since Wang et al. (2012) observed that the use of random seeds higher than 25 could result in lesser standard errors. PTV (2014) also recommends the use of random seed 42 especially in the studies related to predicting inter-arrival time of vehicles. Additionally, it was observed in the present study that performing trial runs for random seed greater than 50 resulted in higher prediction errors. In the light of the above references and also based on simulations performed for random seeds 1-50, it may be considered that the likely set of best performing random variables has been evaluated. Thus, the selection of random seed and the calibration of vehicle and driver characteristics are based on a deterministic approach in this study. Liu (2002) too recommends that simulation trials may be performed using a selected set of random seeds in order to achieve the desired results. A similar approach has been adopted in this study.

On the other hand, Monte Carlo simulations depend upon the use of random numbers to identify a set of stochastic combinations of characteristics to be tested (Newman and Barkema, 1999). But such approaches are found to be useful in situations where the number of characteristics to be evaluated and the value ranges are within reasonable limits (Mattis, 1976). Also, the results obtained using Monte Carlo approaches may not be reliable or may involve more number of trials according to Park and Miller (1988).

The following sub-sections provide details on the results of the first-level of calibrations (performed in 14 major stages), sensitivity analysis, the second-level of calibrations (performed in three major stages), and the validation of simulation model.

6.1. Results and discussions on the first-level of calibrations

An initial test-run using random seed 42 which was identified as the best random seed was performed in VISSIM with default values assigned to the vehicle and driver characteristics. A number of informal trial runs were conducted by fine-tuning vehicle and driver characteristics before commencing the first-level of calibrations. The best sequence of refinement of various vehicle and driver characteristics was then evolved using a trial and error approach.

In Stage 1 of the first-level of calibrations, the minimum lateral clearance for each type of vehicle was assigned based on the recommendations provided by Arasan and Krishnamurthy (2008), Manjunatha et al. (2013), and Bains et al. (2013). In Stages 2a, 2b, and 2c of refinement, the values associated with the distributions for desired accelerations, maximum accelerations, and decelerations respectively were set for each vehicle-type as a part of refining vehicle characteristics using a trial and error method considering values adopted and modified based on Arasan and Krishnamurthy (2008). After these stages of refinements, the mean absolute error remained stable at around 33.4% and the GEH statistic also remained unchanged at 7.06.

In Stages 3a and 3b, the distributions for desired accelerations and maximum accelerations respectively were refined as a part of vehicle characteristics considering values adopted and modified based on Mathew and Radhakrishnan (2010), Mehar et al. (2014), and Bains et al. (2013). Through these refinements, the mean absolute error reduced from 33.4% to 32.9% and the GEH statistic decreased marginally from 7.06 to 6.94.

In Stage 4, the value associated with the minimum look-ahead distance was assigned as part of driver characteristics and the best value for vehicles plying on the city roads was identified as 40m. The mean absolute error at this stage reduced from 32.9% to 29.5% and the GEH statistic reduced significantly from 6.94 to 6.14 indicating that the assigned value improved significantly.

In Stage 5, the values associated with the distributions for maximum accelerations were further refined to increase the accuracy of prediction. At this stage, the mean absolute error reduced from 29.5% to 27.5%, and the GEH statistic reduced from 6.14 to 5.82. In Stage 6, the values associated with the distributions for desired accelerations were further refined to tally with the actual on-field acceleration conditions. Here, the mean absolute error reduced slightly from 27.5% to 27.2%, and the GEH statistic showed a minor reduction from 5.82 to 5.63.

In Stages 7a and 7b of refinement, the values related to the desired speed: lower and upper bounds for speed distributions of various vehicle types were assigned considering values adopted and modified based on Arasan and Krishnamurthy (2008). In these stages, the mean absolute error reduced from 27.2% to 24.5%, and the GEH statistic showed a reduction from 5.63 to 5.12, signifying a moderate improvement in the prediction of traffic volumes.

In Stage 8, the value associated with the minimum look-back distance was assigned as part of fine-tuning driver characteristics. At this stage, the best value for the minimum look-back distance of 30m was arrived at for vehicles plying on the city roads. The mean absolute error reduced from 24.5% to 22.4%, and the GEH statistic reduced from 5.12 to 4.65.

In Stage 9, the value related to the average standstill distance was assigned as part of driver characteristics and the best value for vehicles plying on the city roads was identified as 1m. The mean absolute error reduced from 22.4% to 20.1% and the GEH statistic reduced from 4.65 to 4.08. In Stage 10, the additive part of desired safety distance was refined to 0.4m as part of fine-tuning driver characteristics. At this stage, the mean absolute error reduced from 20.1% to 13.8% and the GEH statistic reduced from 4.08 to 2.74, showing a significant improvement in the accuracy of predictions.

In Stage 11 of refinement, the multiplicative part of desired safety distance was refined to 0.5m as part of fine-tuning driver characteristics. Here, the mean absolute error was found to reduce from 13.8% to 10.3% and the GEH statistic displayed a significant reduction from 2.74 to 2.0. In Stage 12, the time-between direction changes was refined to 5seconds as part of fine-tuning driver characteristics. The mean absolute error was found to be unchanged at around 10.3% while the GEH statistic reduced slightly from 2.05 to 2.01.

In Stage 13, the minimum longitudinal speed (required for lateral movement) was refined to 1.9kmph as part of fine-tuning driver characteristics. The mean absolute error reduced from 10.3% to 10.2% and the GEH statistic reduced from 2.0 to 1.97.

Finally, in Stage 14 of refinement, the minimum collision time-gain was fine-tuned to 0.57 seconds as part of driver characteristics. The mean absolute error reduced from 10.2% to 9.6% and the GEH statistic reduced from 1.97 to 1.89 as in Table 4 for the final stage of refinement.

Table 4

Direction of Traffic Flow		Simulated volume	Observed volume	Absolute Error (%)	GEH Statistic	RNSE Statistic
From	To
Jyothi	Hampankatta	465	516	9.9	2.30	2.25
Hampankatta	Jyothi	431	464	7.1	1.56	1.53
PVS	Bunts	282	293	3.8	0.65	0.64
Bunts	PVS	239	301	20.6	3.77	3.57
Navabharat	PVS	441	502	12.2	2.81	2.72
PVS	Navabharat	464	474	2.1	0.46	0.46
Hampankatta	Navabharat	382	445	14.2	3.10	2.99
Navabharat	Hampankatta	357	420	15.0	3.20	3.07
Jyothi	Bunts	343	404	15.1	3.16	3.03
Bunts	Jyothi	378	394	4.1	0.81	0.81
Jyothi	Balmatta	315	366	13.9	2.76	2.67
Balmatta	Jyothi	476	517	7.9	1.84	1.80
Balmatta	St. Theresa	132	142	7.0	0.85	0.84
St. Theresa	Balmatta	174	161	8.1	1.00	1.02
Bendoorwell	Balmatta	270	291	7.2	1.25	1.23
Balmatta	Bendoorwell	275	336	18.2	3.49	3.33
Bendoorwell	St. Theresa	241	241	0.0	0.00	0.00
St. Theresa	Bendoorwell	205	220	6.8	1.03	1.01
			MAE	9.6	1.89	1.83

1. Source: Authors own creation.

Table 5 provides a summary of the mean absolute errors and the values of the mean GEH statistic at each stage of refinement in the calibration procedure, while Figure 3 provides a pictorial representation of the same. The summary of the values of vehicle and driver characteristics at the end of the first-level of calibrations are provided in Table 6.

Table 5

Stages of Refinement	Mean Absolute Errors at the end of the stages (%)	Mean GEH Statistic
Initial trial (default value)	33.4	7.06
Stage 1	33.4	7.06
Stages 2a, 2b, and 2c	33.4	7.06
Stages 3a and 3b	32.9	6.94
Stage 4	29.5	6.14
Stage 5	27.5	5.82
Stage 6	27.2	5.63
Stages 7a and 7b	24.5	5.12
Stage 8	22.4	4.65
Stage 9	20.1	4.08
Stage 10	13.8	2.74
Stage 11	10.3	2.05
Stage 12	10.3	2.01
Stage 13	10.2	1.97
Stage 14	9.6	1.89

1. Source: Authors own creation.

Table 6

Sl. No.	Vehicle and Driver Characteristics	Calibrated Value
		Car	MTW	Auto-rickshaw	LCV	Bus	Truck
	Vehicle and Driver Characteristics at the End of First Level of Calibrations
1	Minimum Lateral Clearances (m)	0.5	0.5	0.5	0.5	0.5	0.5
2	Desired Accelerations at 50kmph (m/s2)	1.7	1.6	0.5	0.7	0.5	0.4
3	Maximum Accelerations at 50kmph (m/s2)	1.8	1.7	0.6	0.8	0.7	0.7
4	Deceleration Distributions (m/s2)	2.8	2.8	2.4	2.1	2.2	1.8
5	Desired Speed: Lower Bounds (kmph)	35	30	30	38	37	30
6	Desired Speed: Upper Bounds (kmph)	70	60	55	68	65	60
7	Minimum Look-ahead distances (m)	40	40	40	40	40	40
8	Minimum Look-back distances (m)	30	30	30	30	30	30
9	Average Stand Still Distances (m)	1.0	1.0	1.0	1.0	1.0	1.0
10	Additive Part of Desired Safety Distances (m)	0.4	0.4	0.4	0.4	0.4	0.4
11	Multiplicative Part of Desired Safety Distances (m)	0.5	0.5	0.5	0.5	0.5	0.5
12	Time-Between Direction Changes (s)	5.0	5.0	5.0	5.0	5.0	5.0
13	Minimum Longitudinal Speeds (kmph)	1.9	1.9	1.9	1.9	1.9	1.9
14	Minimum Collision Time-gains (s)	0.57	0.57	0.57	0.57	0.57	0.57
	Vehicle and Driver Characteristics Revised at the End of Second Level of Calibrations
1	Average Stand Still Distance (m)	0.4	0.4	0.4	0.4	0.4	0.4
2	Minimum Look-ahead distance (m)	25	20	20	25	30	30
3	Desired Speed: Lower Bounds (kmph)	35	30	30	30	37	30

1. Source: Authors own creation.

2. Notes. MTW = Motorized two- wheelers, LCV = Light commercial vehicles of gross weight lesser than 7.5 metric tonnes.

Figure 3

Download asset Open asset

Simulation approaches to modeling urban traffic are adopted due to the reason that the development of analytical models is cumbersome considering the involvement of a number of vehicle and driver characteristics that also depend to a large extent on the proportion of slow and fast motorized vehicles, the operating volumes and speeds, and the level of congestion. Additionally, microsimulation based approaches are found to be more suitable in analysing the interplay between various interacting variables, especially in the study of flow of vehicles through the bottlenecks in the road network of the city.

The first-level of calibrations for VISSIM-based microsimulation model developed in this study was performed in 14 major stages with 4 sub-stages, starting with the assigning of default values for vehicle and driver characteristics, and systematically making changes to ensure that the accuracy of prediction improved with each stage as illustrated in Figure 3. The second-level of calibrations were performed based on the observations made in ANN-based sensitivity analysis. The following are the important conclusions related to calibration, sensitivity analysis, and validation performed in this study:

• After performing the first-level of calibrations that ended in Stage 14, it was observed that in most of the 18 mid-block locations, the mean absolute error in prediction was much lesser than 15%, and the mean GEH statistic was lesser than 5.0. Although these observations satisfy the criteria for calibration specified by UKDoT (2014) and WSDoT (2014), it was considered ideal to perform a second-level of calibrations based on ANN-based sensitivity analysis in order to further improve the accuracy of predictions.

• An ANN-based sensitivity analysis of vehicle and driver characteristics was performed as part of the study, which revealed that the average standstill distance, the minimum look-ahead distance, and the desired speed: lower bounds for speed distributions with the relative importance values of 36.03%, 29.08%, and 10.02% as in Table 7 influenced the prediction accuracy of the VISSIM model in the second-level of calibrations. It was also found that further fine-tuning in the values of characteristics such as the additive part of desired safety distance, minimum look-back distance, desired accelerations at the desired speeds, maximum accelerations at the desired speeds, and the multiplicative part of desired safety distance did not yield any further improvement in the accuracy of predictions.

• A comparison of the first and the second-levels of calibrations indicate that there is a significant improvement in the accuracy of predictions as the mean absolute error (MAE) shows a reduction from 9.6% in the first-level of calibrations to 7.7% in the second-level of calibrations, and as the GEH statistic shows a reduction from 1.89 to 1.49.

• The validation of the model performed using 25% of the video-graphic data kept apart for the same showed a mean absolute error of 13.7%, with the mean GEH statistic value at 1.51, and the mean RNSE value at 1.47. The RNSE values were found to be much lesser than 3.0 for more than 85% of the mid-block sections validated which satisfied the requirements stipulated by the Wisconsin DoT (2005).

Table 9 provides a summary of final calibrated values for the vehicle and driver characteristics refined after the end of the second-level of calibrations in VISSIM for Mangalore city. The table also provides comparisons to similar studies performed in India.

Here, it can be observed that the calibrated value for the minimum look-ahead distance was 27.91m for the study performed in Chennai city (Siddharth and Ramadurai, 2013), while in the present study, it was found to be 20m for motorized two-wheelers and auto-rickshaws, 25m for cars and LCVs, and 30m for buses and trucks. The operating value ranges suggested for the same for the study performed in Chennai city (Siddharth and Ramadurai, 2013) ranges between 10m and 30m, while a range of values between 10m and 50m is suggested in the present study based on simulation trials performed in the first and second-levels of calibrations considering the traffic flow in the entire road network of the city.

Similarly, the calibrated value for the minimum look-back distance was 14.31m for the study performed in Chennai city (Siddharth and Ramadurai, 2013), while assigning a value of 30m for all vehicle types worked well in the present study. The operating value ranges suggested for the same for the study performed in Chennai city (Siddharth and Ramadurai, 2013) ranges between 10m to 30m, while a range of values between 10m and 50m is suggested in the present study.

Also, the calibrated value for the average standstill distance based on the studies in Chennai city (Siddharth and Ramadurai, 2013) and Trivandrum city (Mathew and Radhakrishnan, 2010) were 1.0m and 1.33m respectively, while in the present study, it was found that assigning a value of 0.4m for all vehicle types worked well. The operating value ranges suggested for the same for studies performed in Chennai city (Siddharth and Ramadurai, 2013) and Trivandrum city (Mathew and Radhakrishnan, 2010) ranges between 1 to 2m and 0 to 4m respectively, while the same for the present study ranges between 0.5m and 2.5m.

Similar, observations can be made in the case of additive part of desired safety distance for which the calibrated value is found to be 0.4m for all vehicle types in the present study, while the studies performed in Chennai city (Siddharth and Ramadurai, 2013) and Trivandrum city (Mathew and Radhakrishnan, 2010) arrived at calibrated values of 0.2m and 0.28m respectively.

Also, it can be found that in the case of multiplicative part of desired safety distance for which the calibrated value is found to be 0.5m for all vehicle types in the present study, the studies performed in Chennai city (Siddharth and Ramadurai, 2013) and Trivandrum city (Mathew and Radhakrishnan, 2010) arrived at calibrated values of 0.78m and 0.16m respectively.

The above-mentioned variations in the calibrated values are mainly attributed to the reason that the studies performed by Siddharth and Ramadurai (2013) focused only on a short stretch of road, namely IT Corridor State Highway in Chennai and the studies performed by Mathew and Radhakrishnan (Mathew and Radhakrishnan, 2010) were restricted to investigations at intersections in cities, whereas, the present study was focused on simulating the traffic flow for the entire road network spread across 3.08 square km of Mangalore city. Thus, it can be observed that the vehicle and driver characteristics need to be calibrated to suit local traffic conditions and driving styles for modeling urban traffic scenarios as recommended by Fellendorf and Vortisch (2001).

1. 1
   Study of the effect of traffic volume and road width on PCu value of vehicles using microscopic simulation

   1. VT Arasan
   2. K Krishnamurthy

   (2008)

   Road and Transport Research 17:32–48.

   • Google Scholar
2. 2
   Methodology for simulating heterogeneous traffic on expressways in developing countries: a case study in India

   1. S Arkatkar
   2. S Velmurugan
   3. R Puvvala
   4. B Ponnu
   5. S Narula

   (2016)

   Transportation Letters 51:1–16.

   https://doi.org/10.1179/1942787515Y.0000000008

   • Google Scholar
3. 3
   From traffic conflict simulation to traffic crash simulation: introducing traffic safety indicators based on the explicit simulation of potential driver errors

   1. V Astarita
   2. VP Giofré

   (2019)

   Simulation Modelling Practice and Theory 94:215–236.

   https://doi.org/10.1016/j.simpat.2019.03.003

   • Google Scholar
4. 4
   Effect of speed limit compliance on Roadway capacity of Indian Expressways

   1. MS Bains
   2. A Bhardwaj
   3. S Arkatkar
   4. S Velmurugan

   (2013)

   Procedia - Social and Behavioral Sciences 104:458–467.

   https://doi.org/10.1016/j.sbspro.2013.11.139

   • Google Scholar
5. 5
   Modeling and simulation based analysis of Multi-Class traffic with Look-Ahead controlled vehicles

   1. Z Bede
   2. B Németh
   3. P Gáspár

   (2017)

   Transportation Research Procedia 27:593–599.

   https://doi.org/10.1016/j.trpro.2017.12.125

   • Google Scholar
6. 6
   Population census 2011. government of India. Mangalore City census 2011 data. PUB. by commissioner and registrar General of the Indian census 2011, New-Delhi

   1. Census

   (2011)

   http://www.census2011.co.in/census/city/451-mangalore.html, 18 Nov 2015.

   • Google Scholar
7. 7
   Traffic Management Report for Mangalore town

   1. Dalal Consultants and Engineers

   (2003)

   Mangalore: Dalal Consultants and Engineers Ltd. Submitted to Karnataka Urban Infrastructure Development and Finance Corporation Ltd. (KUIDFC).

   • Google Scholar
8. 8
   Design manual for roads and bridges. UK highways Agency. Traffic appraisal of roads schemes. London

   1. DMRB

   (1996)

   http://www.standardsforhighways.co.uk/ha/ standards/dmrb/vol12/section2/12s2p1.pdf, 15 Aug 2019.

   • Google Scholar
9. 9
   Development and validation of large scale microscopic models. Transportation research board, 85th annual meeting No. 06-2392

   1. P Dorothy
   2. R Ambadipudi
   3. R Kill

   (2006)

   • Google Scholar
10. 10
    Calibration of Microsimulation Models. Traffic Analysis Toolbox Volume III: Guidelines for Applying Traffic Microsimulation Modeling Software, Federal Highway Administration (FHWA), FHWA-HRT-04-040, Washington, DC

    1. R Dowling
    2. A Skabardonis
    3. V Alexiadis

    (2004)

    Calibration of Microsimulation Models. Traffic Analysis Toolbox Volume III: Guidelines for Applying Traffic Microsimulation Modeling Software, Federal Highway Administration (FHWA), FHWA-HRT-04-040, Washington, DC.

    • Google Scholar
11. 11
    Logistic regression and artificial neural network classification models: a methodology review

    1. S Dreiseitl
    2. L Ohno-Machado

    (2002)

    Journal of Biomedical Informatics 35:352–359.

    https://doi.org/10.1016/S1532-0464(03)00034-0

    • PubMed
    • Google Scholar
12. 12
    Validation of the microscopic traffic flow model VISSIM in different real-world situations. Transportation research board 80th annual meeting, (January 2001), 1–9

    1. M Fellendorf
    2. P Vortisch

    (2001)
13. 13
    Assessing the relative importance of input variables for route choice modeling: a neural network approach

    1. T Feng
    2. T Arentze
    3. H Timmermans

    (2011)

    Journal of the Eastern Asia Society for Transportation Studies 9:341–353.

    • Google Scholar
14. 14
    Interpreting neural network connection weights

    1. GD Garson

    (1991)

    Artificial Intelligence Exp 6:46–51.

    • Google Scholar
15. 15
    Review and comparison of methods to study the contribution of variables in artificial neural network models

    1. M Gevrey
    2. I Dimopoulos
    3. S Lek

    (2003)

    Ecological Modelling 160:249–264.

    https://doi.org/10.1016/S0304-3800(02)00257-0

    • Google Scholar
16. 16
    Two-Way interaction of input variables in the sensitivity analysis of neural network models

    1. M Gevrey
    2. I Dimopoulos
    3. S Lek

    (2006)

    Ecological Modelling 195:43–50.

    https://doi.org/10.1016/j.ecolmodel.2005.11.008

    • Google Scholar
17. 17
    Projected population of Karnataka 2012-2021. Government of Karnataka (GoK). Directorate of economics and statistics Bangalore

    1. GoK

    (2013)

    http://des.kar.nic.in/docs/ Projected%20Population%202012-2021.pdf, 19 Nov 2018.

    • Google Scholar
18. 18
    Annual report of the transport department for the year 2017-18. Government of Karnataka (GoK). transport and road safety department

    1. GoK

    (2018)

    http://transport. karnataka.gov.in, 27 Apr 2019.

    • Google Scholar
19. 19
    Pedestrian modeling in urban road networks: issues limitations and opportunities offered by micro-simulation

    1. MM Ishaque
    2. RB Noland

    (2005)

    9th Annual Computers in Urban Planning and Urban Management 9:1–9.

    • Google Scholar
20. 20
    Rainfall-runoff modelling using artificial neural networks (ANNs): modelling and understanding

    1. A Kalteh

    (2008)

    Caspian Journal of Environmental Sciences 6:53–58.

    • Google Scholar
21. 21
    Evaluating the Transit Signal Priority Impacts along the U.S. 1 Corridor in Northern Virginia

    1. V Kamdar

    (2004)

    Virginia Polytechnic Institute and State University.

    • Google Scholar
22. 22
    Designing a Vissim-Model for a motorway network with systematic calibration on the basis of travel time measurements

    1. I Karakikes
    2. M Spangler
    3. M Margreiter

    (2017)

    Transportation Research Procedia 24:171–179.

    https://doi.org/10.1016/j.trpro.2017.05.086

    • Google Scholar
23. 23
    An introduction to computing: with Haskell

    1. RP Lippmann

    (1987)

    ASSP Magazine, IEEE 4:4–22.

    • Google Scholar
24. 24
    Monte Carlo Strategies in Scientific Computing

    1. JS Liu

    (2002)

    Department of Statistics, Harvard University.

    • Google Scholar
25. 25
    A video-based approach to calibrating car-following parameters in VISSIM for urban traffic

    1. Z Lu
    2. T Fu
    3. L Fu
    4. S Shiravi
    5. C Jiang

    (2016)

    International Journal of Transportation Science and Technology 5:1–9.

    https://doi.org/10.1016/j.ijtst.2016.06.001

    • Google Scholar
26. 26
    Methodology for the calibration of VISSIM in mixed traffic

    1. P Manjunatha
    2. P Vortisch
    3. T Mathew V

    (2013)

    TRB 2013 annual meeting paper revised from original submittal.

    • Google Scholar
27. 27
    Calibration of microsimulation models for nonlane-based heterogeneous traffic at signalized intersections

    1. TV Mathew
    2. P Radhakrishnan

    (2010)

    Journal of Urban Planning and Development 136:59–66.

    https://doi.org/10.1061/(ASCE)0733-9488(2010)136:1(59)

    • Google Scholar
28. 28
    Solvable spin systems with random interactions

    1. DC Mattis

    (1976)

    Physics Letters A 56:421–422.

    https://doi.org/10.1016/0375-9601(76)90396-0

    • Google Scholar
29. 29
    Geographical information. Mangalore City Corporation (MCC)

    1. MCC

    (2015)

    http://www.mangalorecity.mrc.gov.in/en/Geographical%20Information, 1 Dec 2018.

    • Google Scholar
30. 30
    Statistics on length of road under municipal Corporation and land use details 2014-2015. Mangalore City Corporation (MCC)

    1. MCC

    (2016)

    http://www.mangalorecity. mrc.gov.in/sites/mangalorecity.mrc.gov.in/files/City%20 Statistics_0.pdf, 1 Dec 2018.

    • Google Scholar
31. 31
    Highway capacity through vissim calibrated for mixed traffic conditions

    1. A Mehar
    2. S Chandra
    3. S Velmurugan

    (2014)

    KSCE Journal of Civil Engineering 18:639–645.

    https://doi.org/10.1007/s12205-014-0440-3

    • Google Scholar
32. 32
    Lincoln tunnel corridor study: traffic Micro-Simulation

    1. GMS Mosseri
    2. MA Hall
    3. JM Rivera

    (2004)

    Procedia - Institute of Transportation Engineers.

    • Google Scholar
33. 33
    Monte Carlo Methods in Statistical Physics

    1. MEJ Newman
    2. GT Barkema

    (1999)

    1–300, Monte Carlo Methods in Statistical Physics, Oxford University Press, p.

    • Google Scholar
34. 34
    Development and evaluation of a procedure for the calibration of simulation models

    1. B Park
    2. H Qi

    (2005)

    Transportation Research Record: Journal of the Transportation Research Board 1934:208–217.

    https://doi.org/10.1177/0361198105193400122

    • Google Scholar
35. 35
    Microscopic simulation model calibration and validation: case study of VISSIM simulation model for a coordinated actuated signal system

    1. B Park
    2. JD Schneeberger

    (2003)

    Transportation Research Record: Journal of the Transportation Research Board 1856:185–192.

    https://doi.org/10.3141/1856-20

    • Google Scholar
36. 36
    Random number generators: good ones are hard to find

    1. SK Park
    2. KW Miller

    (1988)

    Communications of the ACM 31:1192–1201.

    https://doi.org/10.1145/63039.63042

    • Google Scholar
37. 37
    VISSIM-7.0 User Manual

    1. PTV

    (2014)

    Germany: Planung Transport Verkehrs - Aktiengesellschaft, Karlsruhe.

    • Google Scholar
38. 38
    Another insight into artificial neural networks through behavioural analysis of access mode choice

    1. PVS Rao
    2. PK Sikdar
    3. KVK Rao
    4. SL Dhingra

    (1998)

    Computers, environment and urban systems 22:485–496.

    • Google Scholar
39. 39
    Simulation-Based variable speed limit systems modelling: an overview and a case study on Istanbul Freeways

    1. M Sadat
    2. HB Celikoglu

    (2017)

    Transportation Research Procedia 22:607–614.

    https://doi.org/10.1016/j.trpro.2017.03.051

    • Google Scholar
40. 40
    PsycCRITIQUES

    1. D Scarborough
    2. M Somers

    (2006)

    Neural Networks in Organizational Research: Applying Pattern Recognition to the Analysis of Organizational Behavior, PsycCRITIQUES, 51.

    • Google Scholar
41. 41
    Determining the relative importance of parameters affecting Concrete pavement thickness

    1. G Shafabakhsh
    2. H Naderpour
    3. R Noroozi

    (2015)

    Journal of Rehabilitation in Civil Engineering 3:61–73.

    • Google Scholar
42. 42
    Calibration of VISSIM for Indian heterogeneous traffic conditions

    1. SMP Siddharth
    2. G Ramadurai

    (2013)

    Procedia - Social and Behavioral Sciences 104:380–389.

    https://doi.org/10.1016/j.sbspro.2013.11.131

    • Google Scholar
43. 43
    Traffic modelling guidelines. office of the Mayor of London. TfL traffic manager and network performance best practice version 3.0. transport for London

    1. J Smith
    2. R Blewitt

    (2010)

    http://content.tfl.gov.uk/traffic-modelling-guidelines.pdf, 15 Aug 2019.

    • Google Scholar
44. 44
    Simulation of traffic flow to analyze Lane changes on multi-lane highways under non-lane discipline

    1. S Srikanth
    2. A Mehar
    3. KGNV Praveen

    (2020)

    Periodica Polytechnica Transportation Engineering 48:109–116.

    https://doi.org/10.3311/PPtr.10150

    • Google Scholar
45. 45
    Simulation model development and calibration. US-29 North corridor transportation study. Technical memorandum 5, PUB. by Thomas Jefferson planning district Commission (TJPDC), Charlottesville, Va

    1. TJPDC

    (2007)

    http://www.tjpdc.org/ pdf/places_29/techmemos/29N%20Tech%20Memo%205% 20VISSIM_Calibration.pdf, 5 Jun 2015.

    • Google Scholar
46. 46
    Model calibration, the GEH formula. Wisconsin department of transportation, Microsimulation guidelines

    1. Wisconsin

    (2005)

    http://www.wisdot.info /microsimulation/index.php?title=Model_Calibration, 10 Apr 2018.

    • Google Scholar
47. 47
    Protocol for VISSIM simulation. Washington state department of transportation (WSDoT). Olympia, Washington

    1. WSDoT

    (2014)

    https://www.wsdot.wa. gov/NR/rdonlyres/378BEAC9-FE26-4EDA-AA1F-B3A55F9C532F/0/VISSIMProtocol .pdf, 15 Aug 2019.

    • Google Scholar
48. 48
    Tag unit M3.1: highway assignment modeling. Department of transportation (UKDoT). Transport appraisal and strategic modeling division, London

    1. UKDoT

    (2014)

    https://assets.publishing.service.gov.uk/government/uploads/system/ uploads/attachment_data/file/427124/webtag-tag-unit-m3-1-highway-assignment-modelling.pdf, 15 Aug 2019.

    • Google Scholar
49. 49
    Simulation-Based Testbed Development for Analyzing Toll Impacts on Freeway Travel-Final Report

    1. Y Wang
    2. Y Lao
    3. C Liu
    4. G Xu

    (2012)

    Simulation-Based Testbed Development for Analyzing Toll Impacts on Freeway Travel-Final Report.

    • Google Scholar
50. 50
    Simulation des Strassenverkehrsflusses, 8

    1. R Wiedemann

    (1974)

    Karlsruhe, Germany: Schriftenreihe des Instituts für Verkehrswesen der Universität, (In German language).

    • Google Scholar
51. 51
    A basic introduction to neural networks. Department of computer science, University of Wisconsin, Madison

    1. WiscCS

    (1989)

    http://pages.cs.wisc.edu/ ~bolo /shipyard/neural/local.html, 17 Jan 2017.

    • Google Scholar
52. 52
    Illuminating the “black box”: a randomization approach for understanding variable contributions in artificial neural networks

    1. JD Olden
    2. DA Jackson

    (2002)

    Ecological Modelling 154:135–150.

    https://doi.org/10.1016/S0304-3800(02)00064-9

    • Google Scholar
53. 53
    An artificial neural network approach to spatial habitat modelling with interspecific interaction

    1. SL Özesmi
    2. U Özesmi

    (1999)

    Ecological Modelling 116:15–31.

    https://doi.org/10.1016/S0304-3800(98)00149-5

    • Google Scholar

Author details

1. Marsh M Bandi

   National Institute of Technology Karnataka, Surathkal, Mangalore, India

   For correspondence

   marsh.bandi@gmail.com

   Competing interests

   No competing interests reported

   "This ORCID iD identifies the author of this article:" 0000-0002-7856-1460

2. Varghese George

   National Institute of Technology Karnataka, Surathkal, Mangalore, India

   Competing interests

   No competing interests reported

   "This ORCID iD identifies the author of this article:" 0000-0003-1497-9603