=Paper=
{{Paper
|id=None
|storemode=property
|title=Creating driving behavior for artificial agents in a social augmented micro-world
|pdfUrl=https://ceur-ws.org/Vol-928/0336.pdf
|volume=Vol-928
|dblpUrl=https://dblp.org/rec/conf/csp/SchneiderB12
}}
==Creating driving behavior for artificial agents in a social augmented micro-world==
https://ceur-ws.org/Vol-928/0336.pdf
Creating driving behavior for artificial agents in
a social augmented micro-world
M.C. Schneider and H-D. Burkhard
Humboldt-Universität zu Berlin, Germany
mschneid@informatik.hu-berlin.de, hdb@informatik.hu-berlin.de
Abstract. This paper presents first results on the implementation of
driving agents in the socially augmented micro-world (SAM). Time se-
ries 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 classifications done by hand. Further approaches
for future research are presented and discussed.
Keywords: artificial agents, driving behavior, micro-world, time series
1 Introduction
This paper presents the current work on the ongoing interdisciplinary project of
Psychology and Informatics in the field of human factors. The central goal of the
work in this paper is to construct artificial agents which can replace humans in
a socially augmented micro-world. The socially augmented micro-world (SAM),
was developed by the group of H. Wandke at the Institute of Psychology at
Humboldt-Universität zu Berlin [1]. The aim is to study design problems for
complex human operated systems.
Fig. 1 shows the setup of the ATEO lab system (ALS), consisting of two
humans which interact with SAM through a joystick, as seen in the left part of
the figure. The humans are called micro-world inhabitants (MWI).The human
operator sits on the right in the figure, hidden from both MWI. He supervises
both MWI through the ATEO Master. ATEO refers to the German title of the
umbrella project for developing the ALS: ArbeitsTeilung Entwickler-Operateur.
So far, determinism has been the main deficit of simulations. To avoid deter-
minism there are two humans included, to cooperatively performing a pursuit
tracking task. Each human partially influences a vehicle, which drives a long a
street. Both MWI are instructed with different orders, to drive fast or to drive
accurately. The operator’s task is to improve the overall performance of the
MWI’s by visual or acoustical hints. For a broader range of experiments, it shall
be possible to replace human MWI by artificial agents.
In the next chapter we will review previous work. Then we will show the
final architecture of the agent. The next two parts focus on the two approaches
which were made to analyze the human MWI behavior. We thank the colleages
from Department of Mathematics and Informatics, University of Novi Sad, for
2 M.C. Schneider et al.
Fig. 1: setup of ATEO lab system
providing the data of their series analysis. The last chapter presents the results
of the comparison of both approaches and the first results of artificial agents and
outlines future work.
2 Earlier Work
2.1 Agent architecture
In the last year we finished the proposed architecture [2] [3]. The architecture
focuses on usability, efficiency, maintainability and portability. Usability can be
understood as clarity of the used concepts and the effort needed to use the agents.
Efficiency concerns the real time behavior of each component. The next point
is the maintainability, which refers to the cost of finding and correcting errors.
Modifiability is also part of maintainability and is the cost of changing parts of
the agent architecture. The last part is portability and describes the difficulty
to port the agent from one environment to an other.
One important aspect of the usability is the effort which is needed to implement
new agent behavior. To simplify the architecture we use a design pattern called
strategy [4]. This design pattern gives a lot of possibilities, new agents can easily
be added and the behavior can be changed during run time.
2.2 Classification by hand
As mentioned in [2], the behavior of human MWI need to be analyzed, as a
prerequisite for the development of artificial agents. There are 26 data sets from
former studies with human probands. One data set contains 11 tracks. For each
Title Suppressed Due to Excessive Length 3
track there exists one file in the csv data format, which contains information
about the state of SAM, taken every 39ms.
The first approach focused on only one track, because this track exhibits charac-
teristics like forks or obstacles for the first time and has no intervention from the
operator. We only analysed the critical situations (forks, the avoiding of obsta-
cles). Data gained in the situations provides more variance and hence supports
the identification of selective types of interaction behavior. Six different steering
behavior were found:
1. Adapted navigator: MWI adjusts his steering to characteristics of the track.
2. Extreme steering navigator: steers with maximum deflection alternating ex-
tremely 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 nav-
igator.
5. Indecisive navigator: makes no decision or decides too late.
6. Residue: remaining data sets which did not fit in any other classification or
does not show any characteristic steering behavior.
Fig. 2 shows an example for the visualization that support the classification
by hand. The driver types were classified by similarities of the curves.
2.3 Classification with time series
Instead of analysis by hand as described above, the analysis of time series could
lead to automatized classification. Such an approach was done with the help of
FAP [5]. FAP is a system for time-series anlysis, developed at Department of
Mathematics and Informatics, University of Novi Sad [6]. Three time series have
been investigated which described the path, the acceleration and the deviation
from the ideal path. The data have been derived from the CVS file. Three dif-
ferent similarity measures were chosen: Dynamic Time Warping (DTW), Edit
Distance on Real sequence (EDR) and Edit Distance with Real penalty (ERP).
For each of the three time series a distance matrix was calculated. A hierarchical
clustering algorithm was used for clustering and creating dendrograms. This can
be seen in Fig. 3.
As each of the three similarity measures follows a different methodology, they
lead to different clusters. Two distinct clusters were to be identified distinguish-
ing navigators focusing on speed from those focusing on accuracy. To do this, the
appropriatness of the three similarity measures was investigated. The different
instruction should result in different navigation styles. This difference 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 defined
as:
4 M.C. Schneider et al.
Fig. 2: The acceleration-, speed- and steering curves of a data set at a fork
(left). The road segment with fork (right). The classification sytem of parameters
(below).
P
n
1 |si − g| if m = 0
Pm
|r − g| if n=0
1 i
ERP (R, S) = min{ERP (Rest(R), Rest(S)) + distE RP (r1 , s1 ),
ERP (Rest(R), S) + distERP (r1 , gap), otherwise
ERP (R, Rest(S)) + distERP (s1 , gap)}
|ri − si | if ri , si not gaps
distE RP (ri , si ) = |ri − g| if si is a gap
|si − g| if ri is a gap
This measure borrows ideas from the domain of strings. Two strings are
aligned so that they become identical with the smallest number of added, deleted
or changed symbols.
Title Suppressed Due to Excessive Length 5
Fig. 3: A dendrogram created from hierarchical clustering, with ERP as similarity
measure for the path series on ”learning track”
3 Results and Discussion
3.1 Comparison between classification by hand and classification
with time series
As described in the introduction, the goal is to create artifical agents which can
be substituted human MWI. Those artifical agents should show characteristic
human driving behavior.
The ERP similarity measure is used to compare the human and agent behav-
ior. We looked to confirm the hypothesis that MWI in the same classification,
which was done by hand, are more similar to each other by the ERP similarity
measure. In the first approach, the MWIs were compared using ERP by tracks
were they have interacted separatledly with SAM (track 1-4). The classification
by hand was done by another track (track 7) where steering were done coopera-
tively. 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 Artifical agents first 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. [7] shows that with a reactive agent and
simple rules it is possible to create human-like driving behavior.
P-controller, or proportional controller, is well know as a control system and can
mathematically described.[8]
6 M.C. Schneider et al.
at = Kp (yt − xt )
• at is the resulting controller signal
• Kp is the propotional gain
• yt is the desired position of the vehicle at time t
• xt is the actual position at time t
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 dif-
ference 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 − positioncurrent )
acceleration = Kpa ∗ (1024 − 2 ∗ |steering|)
Both functions have a value range from -1024 till 1024. We tested two different
value pairs.
Fig. 4: steering P-controller (left) and acceleration P-controller (right)
Fig. 4 shows the P-controller function for both Typs, Typ1 (Kps = −1, Kpa =
1) and Typ2 (Kps = −1.28, Kpa = 1). For the steering P-controller deviation
(x-axes) was printed against the steering (y-axes) and for the acceleration P-
controller deviation (x-axes) was printed against acceleration (y-axes).
We calculated the ERP distances for those two artificial agents compared to
the human MWI. As result we got that the ERP distance of the artificial agents
to any of the human MWI is greater then 18000. For example the greatest
distance between two human MWI is 10237 (between Experiment 10 Rider 2
and Experiment 25 Rider 2). This means that the human MWI are far away
from P-controller.
Title Suppressed Due to Excessive Length 7
3.3 Future work
In this paper, we presented the first attempts to create artifical agents which
show human-like driving behavior. To compare the performances, we used the
ERP similarity measure. In the end, we showed a first result for the artificial
agents compared to human MWI. To improve the performance of the artificial
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 PID-
controller. 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
different navigation behavior. With human-like we mean a lower ERP similarity
value. Further work is necessary for the usage of time series for analysis.
References
[1] Wandke, H. ; Nachtwei, J.: The different human factor in automation: the devel-
oper behind vs. the operator in action. In: Waard, D. de (Hrsg.) ; Flemisch, F. O.
(Hrsg.) ; Lorenz, B. (Hrsg.) ; Oberheid, H. (Hrsg.) ; Brookhuis, K. A. (Hrsg.):
Human factors for assistance and automation. Maastricht, the Netherlands: Shaker
Publishing, 2008, S. 493–502
[2] Burkhard, H-D. ; Jahn, L. ; Kain, S. ; Meyer, C. ; Muetterlein, J. ;
Nachtwei, J. ; Niestroj, N. ; Rougk, S. ; Schneider, M.: Artificial Subjects
in the Psychological Experiment Socially Augmented Microworld (SAM). In: Pro-
ceedings of International Workshop CS&P 2011, 2011, S. 55–66
[3] Jahn, L. ; Schneider, M.: Dokumentation der Entwicklung von Agenten zur Erset-
zung von Mikroweltbewohner on der Socially Augmented Microworld. Studienarbeit
an der Humboldt University Berlin, 2012
[4] Gamma, Erich ; Riehle, Dirk ; Helm, Richard ; Johnson, Ralph ; Vlis-
sides, John: Entwurfsmuster: Elemente wiederverwendbarer objektorientierter Soft-
ware. [6. Aufl.]. München and Boston and Mass. [u.a.] : Addison-Wesley, 2011
http://www.worldcat.org/oclc/693880985
[5] Kurbalija, Vladimir ; Burkhard, Hans-Dieter ; Ivanović, Mirijana ; Meyer,
Charlotte ; Nachtwei, Jens ; Fodor, Lidija: Time Series Mining in a Prschological
Domain. 2012
[6] Kurbalija, V. ; Radovanović, M. ; Geler, Z. ; Ivanović, M.: A framework for
time-series analysis. In: Proceedings of the 14th international conference on Artificial
intelligence: methodology, systems and applications, 2010, S. 42–51
[7] Patrick A.M. Ehlert and Leon J.M. Rothkrantz ; Patrick A.M. Ehlert
(Hrsg.): A REACTIVE DRIVING AGENT FOR MICROSCOPIC TRAFFIC SIM-
ULATION. http://www.kbs.twi.tudelft.nl/docs/conference/2001/Ehlert.P.A.M-
ESM2001.pdf. Version: 2001
[8] Russell, Stuart ; Canny, John F.: Künstliche Intelligenz: Ein moderner Ansatz.
1. München and and Boston [u.a.] : Pearson Studium, 2004. – ISBN 978–3–8273–
7089–1