Learning Dynamic Adaptation Strategies in Agent-Based Traffic Simulation Experiments - Multiagent System Technologies - page 87

Classified as:
Class No limit Max 130km/h
No limit

Classified as:
Class turn off VMS turn on VMS
turn off VMS 705 352 12185 3049

Max 130km/h 158 1266 turn on VMS 6709 8057

Fig. 7. Confusion matrices of classifiers in static and dynamic scenario (cross-
validation on training data)

6Con lu on

In this paper, we have presented an approach to learning dynamic adaptation
strategies using the example of trac simulation experiments. Simulation pro-
vides the clear advantage to perform “what if” studies and thus, allows for
investigating alternative behaviors in identical settings. In the evaluation we
have tested two different behaviors in two scenarios (speed limit vs. no speed
limit and forcing cars to change the lane at a feeder road vs. letting them drive
as usual) in order to check if a strategy based on tra c features can be used to
gain an advantage in tra c flow. The results have shown that the trained clas-
sifiers lead to a greater mean velocity in both scenarios in comparison to fixed
strategies where the current situation is not taken into account. Although the
increase of the average velocities per road user is not very high (especially in the
second setting), it is worthwhile taking into account the potential for reduction
of CO 2 emission as well as for gas consumption for the total amount of road
users by such dynamic adaptations. The results of the evaluation might not be
too surprising - e.g., having learned a classifier what speed limit to use in depen-
dence on the trac flow or trac densities (Fig. 6) - but they show the principal
feasibility of the approach generating comprehensible decision rules or trees. In
future work, it would be interesting to investigate more complex settings.
One of the disadvantages of the approach as presented here is that with an
increasing number of options to be taken into account, the number of simula-
tion runs will increase constantly. Even more problematic is a situation where
multiple criteria are to be evaluated, e.g., forcing lane change behavior, different
speed limits, and overtaking prohibition for trucks. In this case the Cartesian
product of options could be used for investigation. However, due to the huge
number of combinations, such an approach would not be feasible any more. We
are currently working on a new approach where successful settings are identified
from a set of independent simulation runs with different behaviors of different
aspects (potentially selected probabilistically). The basic idea is to estimate
the quality of behaviors by taking into account results of similar situations and
thus, identifying behavior rules which are expected to be advantageous in simi-
lar settings. This approach has the advantage that even an unstructured set of
experiments could be taken into account.

Acknowledgments. This work was partly supported by the MainCampus
scholarship of the Stiftung Polytechnische Gesellschaft Frankfurt am Main .