This paper presents rst results on the implementation of driving agents in the socially augmented micro-world (SAM). Time series analysis are used for comparing the agent behavior with human behavior. We will discuss the results from time series analysis/clustering and compare them with classi cations done by hand. Further approaches for future research are presented and discussed.
providing the data of their series analysis. The last chapter presents the results of the comparison of both approaches and the rst results of arti cial agents and outlines future work.
Agent architecture
In the last year we nished the proposed architecture [
One important aspect of the usability is the e ort which is needed to implement
new agent behavior. To simplify the architecture we use a design pattern called
strategy [
Classi cation by hand
As mentioned in [
The rst approach focused on only one track, because this track exhibits characteristics like forks or obstacles for the rst time and has no intervention from the operator. We only analysed the critical situations (forks, the avoiding of obstacles). Data gained in the situations provides more variance and hence supports the identi cation of selective types of interaction behavior. Six di erent steering behavior were found: 1. Adapted navigator: MWI adjusts his steering to characteristics of the track. 2. Extreme steering navigator: steers with maximum de ection alternating extremely from one direction to the other within a short time. 3. No steering navigator: this MWI hardly shows own steering, the partner steers. 4. Parallel navigator: produces comparable navigation lane to an adapted driver.
The parallel navigator exists only in combination with a second parallel navigator. 5. Indecisive navigator: makes no decision or decides too late. 6. Residue: remaining data sets which did not t in any other classi cation or does not show any characteristic steering behavior.
Fig. 2 shows an example for the visualization that support the classi cation by hand. The driver types were classi ed by similarities of the curves. 2.3
Classi cation with time series
Instead of analysis by hand as described above, the analysis of time series could
lead to automatized classi cation. Such an approach was done with the help of
FAP [
As each of the three similarity measures follows a di erent methodology, they lead to di erent clusters. Two distinct clusters were to be identi ed distinguishing navigators focusing on speed from those focusing on accuracy. To do this, the appropriatness of the three similarity measures was investigated. The di erent instruction should result in di erent navigation styles. This di erence should be observable by related clusters derived from the time series. As result the ERP measure is the most suitable to group the MWI. The ERP measure is de ned as: 8>P1n jsi gj >>>>><P1m jri gj ERP (R; S) = minfERP (Rest(R); Rest(S)) + distE RP (r1; s1); ERP (Rest(R); S) + distERP (r1; gap); if m = 0 if n = 0 otherwise
ERP (R; Rest(S)) + distERP (s1; gap)g distE RP (ri; si) = 8>jri <
jri >:jsi sij if ri; sinot gaps gj if siis a gap gj if riis a gap 3
Comparison between classi cation by hand and classi cation with time series As described in the introduction, the goal is to create arti cal agents which can be substituted human MWI. Those arti cal agents should show characteristic human driving behavior.
The ERP similarity measure is used to compare the human and agent behavior. We looked to con rm the hypothesis that MWI in the same classi cation, which was done by hand, are more similar to each other by the ERP similarity measure. In the rst approach, the MWIs were compared using ERP by tracks were they have interacted separatledly with SAM (track 1-4). The classi cation by hand was done by another track (track 7) where steering were done cooperatively. The tracks 1-4 were used since there was no interdependence between the drivers just that their characteristic behavior could be observed more clearly. As a negativ result, the MWIs in one of the classes found by hand did not show more ERP similarity than the others. The question stands, if other results could be achieved by comparing the same tracks. 3.2
Arti cal agents rst results
We decided to use P-controller for our agents because of its simplicity. We want
to start from a simple and easy to understand agent and add more complex
concepts step by step. Ehlert et. al. [
In general, the agent behavior can be called seeking behavior, since it tries to reach a desired point, in this case, the middle of the street. We take the difference from the actual position of the vehicle to the desired position as input for a p-controller. So we get the following equation: steering = Kps (positiondesired acceleration = Kpa (1024
positioncurrent) 2 jsteeringj) Both functions have a value range from -1024 till 1024. We tested two di erent value pairs. In this paper, we presented the rst attempts to create arti cal agents which show human-like driving behavior. To compare the performances, we used the ERP similarity measure. In the end, we showed a rst result for the arti cial agents compared to human MWI. To improve the performance of the arti cial agents foresight will be used. In this way, we come from a pure reactive agent to more deliberative agent. A further point is the step from P-controller to PIDcontroller. We will add more parameters which have to be tuned, but in the end we should get a more human like steering behavior and more parameters to create di erent navigation behavior. With human-like we mean a lower ERP similarity value. Further work is necessary for the usage of time series for analysis.