Learning Dynamic Adaptation Strategies in Agent-Based Traffic Simulation Experiments - Multiagent System Technologies - page 85

Static scenario Dynamic scenario

No speed limit Speed limit: 130 km/h Classifier No regulation Forcing lane change Classifier

Fig. 4. Comparison: Classifier vs. fixed actions

actions. The results are shown in figure 3. The classifier chooses “no speed
limit” at low trac densities and switches to “speed limit 130km/h” for high
trac densities. Analyses have shown that the classifier chooses the best action
for 85% of the simulated situations. This leads to the mean overall velocities
for taking action “no speed limit”: 82 . 71, “speed limit 130km/h”: 82 . 83, and
classifier: 85 . 85km/h. The corresponding boxplot is shown in Figure 4 on the
left. A one-sided paired t-test with error level α =0 . 05 indicates a greater mean
velocity for the classifier ( p< 2 . 2 e 16 in both cases). Even if multiple testing

(two tests) is taken into account using the “familywise error rate” (Bonferronis
FWER), the statistical tests indicate a greater mean velocity of the classifier
compared to the fixed settings.

−

5.2 Dynamic Scenario: Forcing Lane Changes

In the second scenario, there is a part of a motorway divided into 20 segments
EI 0 ··· EI 19 , each with a length of 500m. Each EI has two lanes. The NI between
EI 14 and EI 15 is connected to the one lane link road EI 20 .Figure5shows
an excerpt of the road map. One kilometer ahead the junction point on the
motorway, a variable-message sign (VMS) is used to dynamically advise drivers
to use the left lane in order to let the ones coming from the right enter the
motorway. In our simulation, if this signal is given, drivers will change from the
right to the left lane in the corresponding region if it is possible in the current
situation. Therefore, a classifier is trained using the features density (EI 15 , 16 ),
density (EI 17 , 18 )and density (EI 20 ) as situation description. The two actions
“turn off VMS” and “turn on VMS” (show message to use left lane till junction
point) are compared and the better one is used as the target attribute value
for the training input. During simulation, varying trac densities on the two
entry points into the road network are simulated. Each training example is
generated from the analysis of 60s simulated time, according to the method