You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
<pclass="lead color-5">Inspired by the rapid decrease in lanes on the San Francisco-Oakland Bay Bridge, we study a bottleneck that merges from four lanes down to two to one.</p>
61
-
<p>We demonstrate that the AVs are able to learn a strategy that increases the effective outflow at high inflows, and performs competitively with ramp metering.</p>
62
-
60
+
<pclass="lead color-5">Inspired by the rapid decrease in lanes on the San Francisco-Oakland Bay Bridge, we study a bottleneck, depicted in Fig. 1 that merges from four lanes down to two to one. We come up with a scheme that could save on the costs of deploying traffic light infrastructure by relying on freely available autonomous vehicles.</p>
61
+
<figure>
62
+
<imgsrc="../assets/images/experiments/bottleneck_control.png" class="img-fluid result" alt="Bottleneck control design">
63
+
<figcaptionclass="small text-center"> Fig. 1, a bottleneck in which four lanes go to two then to 1. Red vehicles are autonomous, white are human. </figcaption>
64
+
</figure>
65
+
66
+
<p>In this system, too large of an inflow into the system leads to irreversible congestion. We demonstrate that we can increase the total flow rate of this system by using AVs. The AVs are able to learn a strategy similar to a traffic light, deciding which lanes go and whch stop. This increases the effective outflow at high inflows, and performs competitively with ramp metering, the solution that is currently deployed on the Bay Bridge. </p>
67
+
<p>In the video below, we depict the effects of this congestion.</p>
<imgsrc="../assets/images/experiments/bottleneck_control.png" class="img-fluid result" alt="Bottleneck control design">
90
-
<figcaptionclass="small text-center"> Control structure of the bottleneck. Scale of segments are distorted for visualization. </figcaption>
91
-
</figure>
85
+
<p>The effects of congestion can be visualized in the following graphic, where we plot the number of entering vehicles per hour vs. the number of exiting vehicles per hour. As can be seen, the exiting number gradually increases but then sharply dips down.</p>
92
86
93
87
<figure>
94
88
<imgsrc="../assets/images/experiments/bottleneck_congestion.png" class="img-fluid result" alt="Bottleneck control design">
95
-
<figcaptionclass="small text-center">Without control, congestion rapidly forms in the bottleneck.</figcaption>
89
+
<figcaptionclass="small text-center">Entering vehicles/hour vs. exiting vehicles per hour. Around 1600 entering vehicles per hour, the exiting number begins to decrease.</figcaption>
96
90
</figure>
97
91
92
+
<p>Now we convert 10% of the vehicles to AVs and let them choose their accelerations and actions. As you can see in the attached video, </p>
<p>Now we compare the entering vehicles/exiting vehicles curves with and without autonomous vehicles. The curves with 10% autonomous vehicles is depicted in yellow and the 0% autonomous vehicles i.e. the fully human case, is depicted in blue. At high inflows, we see a 25% improvement in the outflow. </p>
111
+
98
112
<figure>
99
113
<imgsrc="../assets/images/experiments/uncontrolVRLBigAxes.png" class="img-fluid result" alt="Bottleneck control design">
100
-
<figcaptionclass="small text-center"> Control structure of the bottleneck; at high inflows the outflow is improved by 25%. </figcaption>
114
+
<figcaptionclass="small text-center">Comparison of the inflow/outflow curves with and without control. </figcaption>
101
115
</figure>
102
116
117
+
<p>For purposes of curiosity, we compare this with a ramp metering strategy in which we add a set of traffic lights at the entrance to the bottleneck. This is the currently deployed control strategy on the Bay Bridge. We see that the two perform comparably at high numbers of entering vehicles, but the traffic lights outperform at low inflows. However, in this low-inflow regime we would not actually apply control as this is before congestion sets in. Thus, we claim that the two schemes perform comparably. </p>
103
118
<figure>
104
119
<imgsrc="../assets/images/experiments/RLVAlineaBigAxes.png" class="img-fluid result" alt="Bottleneck control design">
105
120
<figcaptionclass="small text-center"> Comparison of inflow, outflow curves for AV control vs. ramp metering. At high inflows they perform comparably. </figcaption>
<td><b>"Simulation to Scaled City: Zero-Shot Policy Transfer for Traffic Control via Autonomous Vehicles"</b>, K. Jang, E. Vinitsky, B. Chalaki, B. Remer, L. Beaver, A. Malikopoulous, A. Bayen,<emclass="color-primary">ICCPS</em>, 2019</td>
<td><b>"Dissipating stop-and-go waves in closed and open networks via deep reinforcement learning"</b>, A. Kreidieh, C. Wu, A. Bayen, <emclass="color-primary">ITSC</em>, 2018</td>
<td><b>"Benchmarks for reinforcement learning in mixed-autonomy traffic"</b>, E. Vinitsky, A. Kreidieh, L. Flem, N. Kheterpal, K. Jang, C. Wu, F. Wu, R. Liaw, E. Liang, A. Bayen, <emclass="color-primary">PMLR, Volume 87</em>, 2018</td>
77
+
<td><b>"Benchmarks for reinforcement learning in mixed-autonomy traffic"</b>, E. Vinitsky, A. Kreidieh, L. Flem, N. Kheterpal, K. Jang, C. Wu, F. Wu, R. Liaw, E. Liang, A. Bayen, <emclass="color-primary">PMLR, Volume 87, 2018 <spanclass="badge badge-success">When citing the benchmarks, please cite this paper</span></em></td>
0 commit comments