This section compares the conflict rates for the different ML strategies and design options against heterogeneous traffic (S0) at the same MAR. The comparison is performed for conflicts of both types (rear-end and lane change) and involving the individual vehicle and aggregated conflicts as per the representation shown in
Figure 4. The last subsection in this discussion focuses on the five design options that were identified as viable ML design options based on traffic operation considerations [
3]. It should be noted that the comparisons of safety performance in this study are qualitative and based on mean conflict rates, where each mean value is calculated using the results of five simulation runs. The statistical significance of the differences noted between conflict rates for an ML design option and the case of heterogenous traffic at the same MAR was also assessed using a series of two-sample
t-tests.
Table 1 summarizes the percentage of ML design options that were significantly different from S0 at a 5% level of significance for each conflict type, vehicle interaction, and traffic demand level. For easy visualization, each cell has a background color, where green cells indicate a percentage greater than 50% and the darker the color the higher the percentage. On the other hand, blue cells indicate a percentage less than 50% and the darker the color the lower the percentage. For example, the 62.5% value in the cell corresponding to DV-DV-RE and
= 0.8 indicates that the mean DV-DV-RE conflict rate of 62.5% of all ML design options at this volume ratio was statistically different from the case of heterogeneous traffic. The table indicates that differences between all ML design options and S0 are mostly significant, except for conflicts involving HVs, especially for lane change conflicts.
4.1. Comparing Rear-End Conflicts of All Design Options
First,
Figure 5 summarizes the rates of aggregated rear-end conflicts for each following vehicle type, while
Figure A1,
Figure A2 and
Figure A3 in
Appendix A summarize the individual vehicle rear-end conflict rates for each combination of lead and following vehicle types. For the easy comparison of all strategies and design options, the
y-axis in these figures is scaled based on the maximum rate in each column of subplots in each figure. As mentioned previously, the conflict rates presented in these figures use different TTC threshold (
) values based on the type of following vehicle. First,
Figure 5 shows that the aggregated conflict rates are fairly low for S0, with a trend of slight increase as
increases. The highest aggregated conflict rates are XV-DV-RE (the following vehicle is a DV) even at the high MAR of 75%, where DVs make up only 25% of all PCs in the network. This can be attributed to the higher TTC threshold (
) for DVs (
= 1.5 s) compared to AVs (
= 0.75 s), causing all events with a TTC in the range of 0.75 to 1.5 s to be excluded from the conflict count when the following vehicle is an AV but included when the following vehicle is a DV. In addition, as the MAR increases, both the number of XV-DV-RE and the DV volume decrease, with a net result of relatively high conflict rates.
The same trend can be observed in the individual vehicle conflicts when breaking down the conflicts by the lead vehicle type, as shown in
Figure A1,
Figure A2 and
Figure A3. These figures also show that, in examining each type of following vehicle, the highest individual vehicle conflict rates are DV-DV-RE, AV-AV-RE, and DV-HV-RE. In addition, for the two continuous-access strategies (S1 and S2), the AV-AV-RE conflict rates are highest in Case b(1AVML50), which could have resulted from having 50% of all AVs traveling in one lane. This trend also causes the DV-DV-RE conflict rates to be lowest for these two design options. Notably, S3b(LR-1AVML50) has a considerably lower AV-AV-RE conflict rate than the same option in strategies S1 and S2, which is likely caused by the inability of some AVs to merge onto the AVML within the ingress area in S3(LR). On the other hand, for the two restricted-access strategies (S3 and S4), the AV-AV-RE conflict rates are highest in Case d(1DVML75), which is associated with the highest number of AVs in the vehicle fleet.
Figure 5 also shows that Case a(1AVML25) in all strategies (S1–S4) has comparable aggregated conflict rates to S0, indicating that all ML strategies do not cause a considerable increase in this conflict type at a 25% MAR compared to heterogeneous traffic. Additionally,
Figure A1,
Figure A2 and
Figure A3 also illustrate that ML strategies are associated with increased rates of conflicts involving the same vehicle type as the lead and following vehicle (i.e., AV-AV and DV-DV). Such a trend is intuitive since ML implementation would cause more vehicles of the same type to travel in the same lane(s), thus increasing the potential for conflicts involving a pair of vehicles of the same type. Similarly, increasing the number of DVs in the GPLs, where HVs travel, causes an increase in DV-HV conflicts.
As the MAR increases to 50%, Case b(1AVML50) still shows low aggregated conflict rates in S1(LC), S2(RC), and S4(RR), while Case c(2AVML50) shows low aggregated conflict rates only in S2(RC). These four design options are comparable to, albeit higher than, the corresponding heterogeneous traffic case S0(Het-50). Other trends related to the increase in conflicts AV-AV, DV-DV, and DV-HV are also valid for all these design options. The remaining design options experience a considerable increase in the conflict rates. For example, S3b(LR-1AVML50) shows considerably higher rates of all conflicts involving DV and HV as the following vehicle type. At a 75% MAR, only the design option S1d(LC-1DVML75) is comparable to the corresponding heterogeneous traffic case S0(Heterogenous MAR at 75%), while the remaining options cause an increase in the conflict rates.
4.2. Comparing Lane Change Conflicts of All Design Options
First,
Figure 6 summarizes the rates of aggregated lane change conflicts for each following vehicle type, while
Figure A4,
Figure A5 and
Figure A6 in
Appendix A summarize the individual vehicle lane change conflict rates for each combination of lead and following vehicle types. As mentioned previously and similar to rear-end conflicts, the conflict rates presented in these figures use different TTC threshold (
) values based on the type of following vehicle. First,
Figure 6 shows that the aggregated conflict rates in heterogeneous traffic (S0) increase with the increase in
. This can be explained by increasing traffic volumes and congestion prompting more vehicles to try to change lanes in an attempt to reduce their travel time. The figure also shows that the XV-DV-LC conflicts (where the following vehicle is DV) generally experience the highest conflict rates. As shown in
Figure A4, these high XV-DV-LC conflict rates are caused mainly by high DV-DV-LC conflict rates.
Figure A4 also shows that, as the MAR increases, the dominant lead vehicle type in these conflicts changes from DVs at a 25% MAR to AVs at a 75% MAR, which is expected, since the majority of the vehicular fleet is made up of AVs.
Figure 6 also shows that the ML design options generally reduce the aggregated lane change conflict rates. The notable exceptions are Case c(2AVML50) in all strategies except for S2(RC) and all cases in S3(LR) except for S3d(LR-1DVML75).
Figure A5 also shows that Cases c(2AVML50) and d(1DVML75) in all strategies S1-S4 increase AV-AV-LC conflict rates. Finally, as indicated earlier in
Table 1, compared to heterogeneous traffic (S0), most design options did not produce significantly different lane change conflict rates involving HV as the following vehicle.
4.3. Comparing Select Design Options
As mentioned earlier, five specific ML design options were deemed feasible based on the traffic operational performance assessed using network output volumes and average vehicle travel times [
3]. Four of these options corresponded to a 25% MAR; these were Case a(1AVML25) in all strategies (S1–S4), while only one option corresponded to a 50% MAR, namely S1b(LC-1AVML50). This subsection examines closely these five options relative to the heterogeneous traffic conditions at the corresponding MARs. Therefore,
Figure 7 summarizes the aggregated conflict rates for these five options along with the heterogeneous traffic conditions S0(Het-25) and S0(Het-50). For brevity, in the figures, these two heterogeneous conditions are shown as Cases S0a and S0b, respectively.
Figure 8 and
Figure 9 summarize the individual vehicle conflicts for the considered options for rear-end and lane change, respectively.
Table 2 summarizes the
p-values for the two-sample
t-tests to compare the rate of each conflict type in each ML design option against the corresponding heterogenous traffic case at the same MAR and
values. Cells with a
p-value less than 0.05 are highlighted to indicate a statistically significant difference at a 5% level of significance.
Figure 7,
Figure 8 and
Figure 9 confirm the trends previously observed when comparing all strategies. Specifically, XV-DV exhibits the highest aggregated rear-end and lane change conflict rates. These higher rates result mainly from high DV-DV conflict rates, which could be due to the higher TTC threshold value. The figures also show that, when the following vehicle is an AV, the highest rear-end conflict rates are AV-AV-RE. However, for the lane change conflicts, Case a(1AVML25) in all strategies has DV-AV-LC conflict rates higher than those for AV-AV-LC. This trend deviates from the general trend in
Figure A5, where the AV-AV-LC conflict rates are generally higher than those for DV-AV-LC. The figures also show different trends in relation to the change in conflict rate with ML deployment relative to S0. First, at the low traffic demand level (
= 0.8), the aggregated XV-XV-RE conflict rates are lower for all five ML options than for the case of heterogenous traffic. As the traffic demand level approaches or exceeds capacity (
= 1.0 and 1.2), the aggregated rear-end conflict rates may increase for some options. Still, S1a(LC-1AVML25) experiences lower aggregated rear-end conflict rates of all types than S0. However, as shown in
Table 2, the differences in the rear-end conflict rates are mostly insignificant for conflicts involving only DVs and HVs, while the differences are mostly significant for conflicts involving AVs as the lead or following vehicles. For the overall aggregation (XV-XV-RE), only 33% of all comparisons at 25% MAR returned statistically significant differences from the case of heterogeneous traffic. For the only ML option at a 50% MAR, which is S1b(LC-1AVML50), the rear-end conflict rates are generally lower than those for heterogenous traffic at the low traffic demand level (
= 0.8) but mostly higher at the higher demand levels (
= 1.0 and 1.2). As shown in
Table 2, most of these differences are statistically significant. This result could be due to the increase in traffic turbulence generated by the number of AVs on the freeway as they seek an opportunity to enter the left-side ML at higher demand levels.
For the lane change conflicts shown in
Figure 7 and
Figure 9, the aggregated rates at the low demand level (
= 0.8) are lower than in S0 only for S2a(RC-1AVML25), S4a(RR-1AVML25), and S1b(LC-1AVML50), and the differences are mostly significant for conflicts involving AVs. This trend suggests that deploying AVML at a low traffic demand level can reduce potential collisions at MARs of 25% and 50%. As the traffic demand level approaches or exceeds capacity (
= 1.0 and 1.2), the aggregated lane change conflict rates may increase for some options, similar to the trend of rear-end conflicts. S1a(LC-1AVML25), which was shown to have lower aggregated rear-end conflict rates, has comparable aggregated lane change conflict rates to S0, where the differences in the XV-XV-LC conflict rates are statistically insignificant. At the 50% MAR, S1b(LC-1AVML50) has significantly lower aggregated lane change conflict rates at
= 0.8, significantly higher rates at
= 1.0, and comparable rates that are not significantly different at
= 1.2. Similar to rear-end conflicts, an increase in conflicts is expected at higher
values since the higher vehicular volume on the network would generate more turbulence in the traffic stream as vehicles attempt to change lanes to enter or exit the ML or freeway. It is evident that there is a reduction in the number of rear-end and lane change conflicts at MARs of 25% and 50% in most cases when S1(LC) is adopted. Hence, there is a possibility that, by implementing this type of ML at these MARs, traffic collisions could be reduced, which would, in turn, improve safety for all road users during this transition phase.