Negotiation in Ride-hailing between Cooperating BDI Agents
¨
Omer Ibrahim Erduran
∗ a
, Marcel Mauri
∗ b
and Mirjam Minor
∗ c
Department of Computer Science, Goethe University, Frankfurt am Main, Germany
Keywords:
Multi-agent Cooperation, Environmental Sustainability, Distributed Problem Solving, Ride-hailing, Mobility
as a Service, BDI Architecture.
Abstract:
These days, ride-hailing is an emerging trend in Mobility as a service (MaaS). First services involving human
taxi drivers such as Uber, Lyft and DiDi are commercially successful. With the rise of autonomous vehicles,
self-organized fleets for ride-hailing systems come into the focus of research. Multi-agent systems (MAS)
provide solutions for many challenges of this application scenario. Especially, the communication of coop-
erating agents is beneficial for a structured and well planned task distribution. In this paper, we investigate a
MAS for autonomous vehicles in MaaS and put the focus on a negotiation based assignment of customer trips.
An agent model concept is introduced where the main type, the vehicle agent is designed following a BDI
architecture. The communication system for the MAS is implemented by using the contract net protocol. We
develop the negotiation process and furthermore evaluate the agent communication with respect to its impact
on pickup time satisfaction and environmental sustainability using two quality measures, which calculate the
average travel distance and the order dropout rate. An experimental setup including historical trip data in a
simulation demonstrates the feasibility of our approach.
1 INTRODUCTION
Due to frequent traffic congestions and problems with
parking space in urban areas, city residents tend to re-
frain from owning private cars (Pavone, 2015). Mo-
bility as a service (MaaS) ”stands for buying mobility
services as packages based on consumers’ needs in-
stead of buying the means of transport” (Kamargianni
et al., 2016). E-mobility with autonomously driving
vehicles will contribute to this ambitious goal. Today,
however, there are still many open issues in this field
and MAS are capable of providing novel solutions
for MaaS applications. This paper addresses a real-
world scenario of ride-hailing using multi-agent co-
operation and negotiation for a fleet of autonomously
driving vehicles. In ride-hailing, a single passenger
is served by a single autonomous vehicle (Qin et al.,
2020). The main task of the MAS is to provide trip
services for customers, who can call an autonomous
vehicle, for this purpose on demand. We address in
our work the self-organizing management which has
its roots in distributed problem solving by consider-
a
https://orcid.org/0000-0002-1586-0228
b
https://orcid.org/0000-0002-4135-1945
c
https://orcid.org/0000-0002-6592-631X
∗
The authors have contributed equally to this work.
ing trip requests with negotiating agents acting as a
single fleet. We further compare this method with a
centralized approach where the trip processing takes
place in a greedy manner. One aspect that needs to
be considered in mobility scenarios is its impact to
the environment. The distribution of customer trips as
tasks among the fleet leads to reduction of power con-
sumption through less driven distances. We propose a
novel domain specific concept of a multi-agent model
with different types of agents, an application of the
contract net protocol (CNP) for the communication
between the BDI agents representing the autonomous
vehicles. The core contribution of this paper is to
investigate the benefits of negotiation in a decentral-
ized MAS for the given scenario of autonomous taxi
fleet. As a proof of concept, we measure the improve-
ment in terms of missed trips and battery consump-
tion while letting a fleet operate in our experiments.
The latter is directly derived from the overall travel
distance to serve the customer requests. The research
question leads to the formulation of the following hy-
pothesis:
A utility based negotiation solution for delegating
trip assignments among a fleet of autonomous vehicle
agents will reduce the rate of missed customer trips
as well as decrease the energy consumption.
Erduran, Ö., Mauri, M. and Minor, M.
Negotiation in Ride-hailing between Cooperating BDI Agents.
DOI: 10.5220/0010973700003116
In Proceedings of the 14th International Conference on Agents and Artificial Intelligence (ICAART 2022) - Volume 1, pages 425-432
ISBN: 978-989-758-547-0; ISSN: 2184-433X
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
425

The remainder of this paper is structured as fol-
lows: In Section 2, we discuss related work consid-
ering research topics of autonomous vehicles for trip
services. Section 3 deals with the design of the MAS
with different agent types containing the architecture
of the vehicle agents as well as the state model for the
trip request processing. The trip assignment process
of the MAS during negotiation is evaluated in Sec-
tion 4.1. In Section 5, we discuss future work and
conclude with a summary.
2 RELATED WORK
The research area for this work is twofold: On the
one side, we discuss existing work studying the multi-
agent approach of trip assignments for vehicle sharing
systems. On the other side, we investigate recent ap-
plications and works which consider autonomous ve-
hicles as sharing systems in e-mobility.
Applying the BDI agent architecture (Silva et al.,
2020) for vehicle agents has been investigated re-
cently in (R
¨
ub and Dunin-Keplicz, 2020), which tends
to be the most relatable work to our approach. The
authors realize traffic agents considering a subsump-
tion architecture in the agent design step and focus
on small-scale traffic. The difference to our agent
architecture is that the work mainly focuses on the
agent design and implementation process without an
evaluation of the performance. However, in our work
we investigate the fleet performance concerning ne-
gotiation with critical measurements which are typi-
cal in the ride-hailing scenario (e.g. estimated time
of arrival (ETA) (Fu et al., 2020)). Experimental re-
sults are introduced in (Certicky et al., 2014), where
the authors study the travelled distance and success
rate with the support of a simulation. Further results
are shown in (Jaroslaw Kozlak and Zabinska, 2013),
where the agent simulation is processed with JADE,
which is also considered in our work. The focus in
both works lies on a setting similar to the ride-hailing
scenario. However, they both do not use the BDI ar-
chitecture for their vehicles. In contrast, our work fo-
cuses on the efficient distribution of trip requests so
that the vehicle fleet takes advantage of negotiation
considering its own utility. Moreover, we consider
free-floating data, which reflects the property of trips
carried out by autonomous vehicles. The research
problem addressed in (Malas et al., 2016) is similar
to ours but here, neither BDI agents are considered
nor the CNP is used for trip assignment. The specific
usage of BDI agents is considered in (Deljoo, 2017)
where the agents plan their work based on their util-
ity function. As mentioned before, we set the focus
on negotiation, which is significant for the MAS dur-
ing processing and not explicitly on agent planning.
Multi-agent approaches in context of traffic scenarios
are discussed in (Bazzan and Kl
¨
ugl, 2014).
The considered application system is also well
known as Autonomous Mobility on Demand (AMoD),
which relates to a mobility type, where autonomous
vehicles provide transportation services for cus-
tomers (Pavone, 2015). In (Pavone, 2015), the authors
present a centralized stochastic solution using a spa-
tial queuing model for the trip assignment problem,
which differs from our multi-agent approach. They
focus on an optimal routing of high-scale AMoD sys-
tems as well as their economical viability and ac-
ceptability in society. In (Danassis et al., 2019), a
heuristic-based approach for solving the trip assign-
ment problem is presented. The decentral nature of
the proposed heuristic leads to no communication be-
tween participants as well as a high scalability. For
their computation of trip assignment matching, they
use a deviant utility function for the trip assignment
process considering only the time factor. Further-
more, they optimize the assignment problem with lin-
ear programming, whereas we use distributed prob-
lem solving by means of cognitive agents and negoti-
ation. The usage of CNP is investigated in (Yu and
Zhang, 2010), where the scheduling of truck vehi-
cles is developed in a pickup and delivery scenario.
The evaluation therefore delivers results concerning
the scheduling, whereas in our work, we use the CNP
for the trip vehicle assignment.
3 DESIGN OF THE MAS
The main task of the MAS is the collaborative man-
agement of trip requests and battery power. An agent
model with different agent types has been designed
for this task. Currently, the agents are running in
a simulation environment. In future, the agents are
supposed to run as well in a real environment with
sensory inputs from autonomously driving vehicles.
The vehicles will also serve as actuators for physi-
cal movements in addition to a communication sys-
tem between the agents.
3.1 Agent Model and Communication
The agent model comprises the following agent types:
a vehicle agent represents an autonomous vehicle
within a fleet, an area agent represents a section of
the outdoor grounds where the vehicles operate, a taxi
office agent represents the interface to the customers
for a larger territory. The taxi office agent serves as
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
426

a preliminary broker for the incoming customer re-
quests. The agent uses a zone model for the area
agents in its entire territory and also provides a reg-
istry service for the vehicle agents. The zone model
determines the area agent in whose zone the customer
has sent the trip request. Further, it knows approxi-
mately where the vehicle agents are since they regis-
ter and de-register themselves at the area agents when
entering or leaving a zone. From time to time, the
area agents inform the taxi office agent on the num-
ber of registered vehicle agents. The taxi office agent
forwards incoming customer requests to one of the
vehicle-populated zones that are in close proximity
to the customer request. The registration data has a
higher time accuracy than in the zone model of the
taxi office agent. The area agent initially assigns in-
coming trip requests to an agent within its zone and
might select the agents arbitrarily according to their
recent location, in rotating order, or just choosing the
most recently registered agent. In addition to the
registry and pre-assignment service, the area agent
provides a neighborhood service that determines the
vehicle agents in the proximity of the vehicle agent
which is searching for negotiation partners. The latter
reduces the communication load during the negotia-
tion. The vehicle agent decides for each trip request
whether it tries do delegate it or whether it takes over
the trip itself. It is responsible for scheduling all trips
to which it has committed itself, including trips for
battery recharging. The delegation process follows
the CNP (Smith, 1980) for agent negotiation. A vehi-
cle agent that aims to delegate a trip request takes the
role of a manager announcing the trip request to be
negotiated on. The neighboring vehicle agents take
the role of a contractor bidding for the trip request.
The manager evaluates the bids and awards a contract
to the bidder it determines to be most appropriate.
3.2 Agent Architecture
All agent types of the agent model follow a cyclic in-
finite sequence of iterations of the observe-think-act
agent cycle (Kowalski and Sadri, 1999). The taxi
office agent and the area agents adopt a stimulus-
response architecture (Kowalski and Sadri, 1996),
which means that the decision is a direct response to
a sensory input, for instance to an incoming message.
Thus, the agent shows a reactive behavior. The ve-
hicle agents are designed in the BDI agent architec-
ture (Rao and Georgeff, 1995), which is a more so-
phisticated specialisation of the observe-think-act cy-
cle (Kowalski and Sadri, 1999) with a richer ’think’
phase than stimulus-response. B stands for beliefs and
represents the agent’s assumptions on its own, inter-
nal state and the state of the environment. A sample
belief of a vehicle agent is the current battery charge
level or the own location in the environment. D stands
for desires and represents the agent’s objectives to be
accomplished. An example desire is to fulfill a new
trip request within reasonable time. It might be in
conflict to other desires, such as maintaining a certain
level of battery charge or serving already committed
trip requests. I stands for intentions and represents the
currently chosen course of action. A sample intention
is to delegate a trip request by negotiation to another
agent. BDI agents show a deliberative behavior and
they are able to plan and pursue multiple objectives in
parallel.
3.3 Utility-based Decisions
A i-th trip request tr
i
contains the following informa-
tion:
tr
i
= (id
i
,type
i
,VATime
i
, l
start
i
, l
end
i
) (1)
It comprises a trip ID id
i
, the type of trip type
i
set
to ”CUSTOMER T RIP”, the desired vehicle arrival
time VATime
i
to pickup the customer, a start location
l
start
i
, and an end location l
end
i
of the trip. The vehi-
cle agent aims to commit itself only to tr
i
’s that seem
beneficial. A j-th vehicle agent va
j
is defined as
va
j
= (id
j
, l
j
, battery
j
) (2)
containing an id
j
, a current geolocation l
j
and a bat-
tery value battery
j
. It has different options for a trip
request tr
i
that is advertised for bids or that is pre-
assigned to va
j
. The utility function for a trip request
u(tr
i
) balances three relevant criteria, the journey to
the customer u
dist
, the battery power status u
battery
and
the trip history u
pts
which are weighted by w
1
, w
2
and
w
3
. It evaluates each option to deal with tr
i
in the
agent’s current situation as follows:
u(tr
i
) = w
1
∗ u
dist
(tr
i
)+
w
2
∗ u
bat
(tr
i
) + w
3
∗ u
pts
(tr
i
)
(3)
The calculation of the first component is based on
a distance measure d for geolocations. A geoloca-
tion l is defined as l = (longitude, latitude) in deci-
mal degrees (DeMers, 2008). The euclidean distance
approximates the distance between decimal degrees
with a 1 meter variation in every 2,500 meters dis-
tance (cmp. the discussion in (Erduran et al., 2019)).
The agent measures the euclidean distance d from its
current location va
j
.l to the location to pickup the cus-
tomer tr
i
.l
start
. The utility u
dist
normalizes the dis-
tance value by means of a bounding box around the
entire territory of the MAS in order to achieve values
Negotiation in Ride-hailing between Cooperating BDI Agents
427

between 0 and 1. dmax denotes the maximum possi-
ble distance between two points at the borders of the
territory:
u
dist
(tr
i
) =
dmax − d(va
j
.l, tr
i
.l
start
)
dmax
(4)
As the second component, u
battery
considers the
battery consumption of a potential trip in a rough ap-
proximation. Assuming a linear decrease of battery
during traveling, the battery consumption in terms of
number of charge units is directly derived from the
travel time. The travel time between two geolocations
l
x
, l
y
at a constant velocity v is estimated as:
travel time(l
x
, l
y
) =
d(l
x
, l
y
)
v
(5)
The agent calculates the travel time for a potential
round trip (tr
i
.type = round) summing up the time for
the journey to the customer d(va
j
.l, tr
i
.l
start
), the trip
itself d(tr
i
.l
start
,tr
i
.l
end
) and the journey back to the
initial position d(tr
i
.l
end
, va
j
.l).
The utility measures the decrease of the current
battery level va
j
.battery by the battery consump-
tion for tr
i
under consideration of bpc
i, j
which is
the battery power consumption of tr
i
processed by
va
j
. It is derived by the sum of d(va
j
.l, tr
i
.l
start
)
and d(l
start
, l
end
). We assume that a full battery con-
tains 100% of power neglecting specific energy units.
Since bpc
i, j
is proportional to the distance driven, we
consider this as a percentage value reflecting the bat-
tery consumption related to the full battery. In case
the current battery level va
j
.battery is too low to ful-
fill the trip, the utility takes the value −∞. Since we
consider a constant velocity, the battery consumption
is derived by the distance driven and the time needed.
In u
battery
the battery consumption is multiplied with
a weight B
f actor
resulting to a utility score. We con-
sider a distinction concerning the battery power level
va
j
.battery of the vehicles in 3 levels:
B
f actor
=





1.0, va
j
.battery > 80%
0.75, 80% ≥ va
j
.battery ≥ 30%
0.1, va
j
.battery < 30%
(6)
The B
f actor
is set to rate battery lifetime friendly
thresholds higher. These thresholds are also used in
other works (Ahadi et al., 2021; Zhang et al., 2020;
Dlugosch et al., 2020). A va
j
.battery beyond the
threshold gets a higher score to create incentives for
the trip to reach the threshold. The function for the
battery utility is defined as follows:
u
bat
(tr
i
) =
(
−∞, va
j
.battery < bpc
i, j
B
f actor
∗ (
bpc
max
−bpc
i, j
bpc
max
), else
(7)
For bpc
max
a vehicle agent will consider the min-
imum of his own current capacity va
j
.battery and the
battery consumption for the maximum possible trip
length.
For the third component u
pts
, we consider the
punctuality of the vehicle agent arriving to the trip
starting position. It is divided into 4 levels. Here
we consider θ as a threshold in minutes the cus-
tomer is ready to wait until the trip is canceled and
η is a buffer time where the vehicle should arrive in
this. The punctuality is the result of the difference
between the vehicle arrival time desired by the cus-
tomer VATime
i
and the estimated time of departure
etd(tr
i
.l
start
), which is calculated before driving to the
trip request. Thus, we consider the following distinc-
tions:
u
pts
(tr
i
) =









0.0, etd(tr
i
.l
start
) + θ > tr
i
.VATime
0.2, etd(tr
i
.l
start
) − θ < tr
i
.VATime < etd(tr
i
.l
start
)
0.6, 0 < tr
i
.VATime − etd(tr
i
.l
start
) < η
1.0, tr
i
.VATime − etd(tr
i
.l
start
) > η
(8)
The punctuality utility is 0.0, when the vehicle
would not be on time to the customer even with the
customer waiting which is defined by θ. It is 0.2 when
the vehicle has a slight delay, 0.6 when there is a small
buffer time defined by the threshold η and 1.0 when
there is enough buffer time.
4 EXPERIMENTS AND
EVALUATION OF TRIP
ASSIGNMENTS
The feasibility of the proposed agent model is investi-
gated by some experiments with sample data. The ex-
perimental setting comprises 10 sample data sets de-
rived from historical trip data as described in Subsec-
tion 4.3. The benefits of the trip assignment method in
a MAS (cmp. Section 3) is investigated by comparing
its results with those of a centralized trip assignment
method as a baseline. We denote the utility-based ne-
gotiation method for trip assignment neg. The base-
line method denoted by greedy just assigns each in-
coming trip request to the agent that is recently in the
closest proximity to the customer.
4.1 Evaluation Criteria
To evaluate our hypothesis, we will consider the cus-
tomer satisfaction and the battery consumption. Two
measures are defined as evaluation criteria for the
agent behavior. The order dropout rate ODR mea-
sures the rate of trip requests that have been dropped.
This is a measure for the customer satisfaction. The
average travel distance AT D measures the average
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
428

travel distance to serve a trip. Thus, AT D is a mea-
sure for battery consumption (cmp. the discussion
on the linear dependency between distance and bat-
tery in Section 3.3). Both measures are based on the
notion of a single agent’s event trace σ. An event
trace σ(va
j
) =< e
0
, e
1
, ..., e
k
> records a sequence
of events where a single event e comprises an event
type e type ∈ {START , PICKUP, DROP, PASS BY ,
REFILL}, a geolocation l the agent has visited, an
arrival time ta and a departure time td for l. For ex-
ample, an event trace of an agent va
j
with an initial
geolocation l
0
where the agent starts to visit further k
geolocations is denoted as follows:
σ(va
j
) =
*





e type
0
ta
0
l
0
td
0





,





e type
1
ta
1
l
1
td
1





, ...,





e type
k
ta
k
l
k
td
k





+
(9)
The individual order dropout rate odr is calcu-
lated counting the number of trip requests an agent
va
j
has missed divided by the number of trip requests
it has committed to fulfill for the time interval covered
by event trace σ.
odr(σ(va
j
)) =
# Dropped trips
# All committed trips
(10)
The order dropout rate of the entire MAS ODR
is the average of the individual odr’s of its agents.
Please note that trips which have been delegated
to another agent by negotiation are not counted by
odr(σ(va
j
)) since they are under the responsibility of
the delegate.
The second measure uses the individual average
travel distance atd an agent is driven per successfully
served trip according to the event trace σ. The overall
travel distance otd calculates the sum of the partial
routes from geolocation to geolocation:
otd(σ(va
j
)) =
k−1
∑
i=0
d(l
i
, l
i+1
) (11)
atd normalizes the otd of an agent by dividing it
through the number of served trips.
atd(σ(va
j
)) =
otd(σ(va
j
))
#Servedtrips
(12)
Analog to ODR, the ADT of the entire MAS is
formed by the average of the individual atd values.
4.2 Simulating the Event Trace
It would be very expensive to assess a committed as-
signment in situ, i.e. within a real cyber-physical sys-
tem. Instead, we simulate trip requests for the eval-
uation of the trip assignment process. Every agent
records its committed trip requests in a trip plan. A
trip plan w includes trip requests to serve a customer
request as well as refill trips to the charging station.
Therefore we denote:
w = (tr
1
,tr
2
, ...,tr
m
) (13)
where tr
i
.tr
type
is set to CUST OMER T RIP in
case tr
i
refers to a trip request from a customer and
REFILL T RIP if the trip request comes from the
agent itself to recharge battery. Note, that the end lo-
cation is empty for trips of type REFILL
T RIP. From
the simulated trip plan, an event trace can be com-
puted. It is built from processing w sequentially to
create the events for each tr
i
. Each event records the
simulated time of arrival eta as ta and the simulated
departure time etd as departure time td. As a rule of
thumb, trip requests are considered missed if the sim-
ulated time of arrival has a delay above a threshold
θ. A more detailed description of the build process of
the simulated event trace σ
e
(va
i
) can be found in the
appendix as an algorithm in pseudo code (Alg. 1).
4.3 Experimental Trip Data
Trip requests from customers arise on a specific time
and place in the considered area. In our ride-hailing
experiment, the trips start and end at the University
campus having a certain length with origin and desti-
nation coordinates. We create 10 samples containing
the customer trip requests, where the largest sample
contains 318 trips and the smallest 178 trips. The co-
ordinates of the trips as well as the samples are gener-
ated from a station based bike sharing data set
1
. The
data set is prepared and generated with random coor-
dinates, to realize a free floating scheme, based on a
method which is described in (Erduran et al., 2019).
To achieve a higher frequency of requests, the real
data from a week have been merged to a one-day data
sample. A single trip tr in a sample is described as
introduced in Section 3.3.
4.4 Experimental Prototype and Results
The MAS is implemented with JADE (Bellifemine
et al., 2008). We set up the experiments for compar-
ing agent communication according to the following
structure: Each data sample described in 4.3 is used in
4 different configurations leading to 40 runs in total.
We use two different amounts of participating vehicle
agents 3 and 7 and both with and without negotiation
denoted by neg
3
, greedy
3
, neg
7
, and greedy
7
. In all
configurations, the trips are initially assigned by the
1
https://data.deutschebahn.com/dataset/data-call-a-
bike, last access: Sept 19, 2021.
Negotiation in Ride-hailing between Cooperating BDI Agents
429

area agent to the vehicle agent that is located closest
to the customer. In the configuration where the agents
are allowed to negotiate the trikes will use the util-
ity function (cmp. Section 3.3) to decide if they will
commit the trip by themselves or start the contract net
protocol to delegate them to another vehicle agent. In
the case where negotiation is not allowed every vehi-
cle agent is forced to commit the trips as they were
assigned to them by the area agent. In every setup the
vehicle agents drive with a velocity of v = 3.6 km/h.
For the evaluation, we use a θ of 4 minutes to decide
if a vehicle agent arrives in time on the requested lo-
cation. For each run, the vehicle agents start with a
fully charged battery. The battery level is considered
by the utility function. The refill trips, however, have
not yet been implemented in our experimental proto-
type. The experiments are processed with a desktop
computer, which contains an Intel Core i5-9500 and
32 GB DDR4-RAM. The source code can be found
on GitHub
2
.
The results for the considered ten samples are
summarized in Table 1, where the AT D, ODR as well
as the amount of lost and committed trips for every
configuration is shown. In Fig. 1, only in one sam-
ple (VIII) the greedy approach leads to less lost trips
than the approach with negotiation. In the 9 other
sample runs, using negotiation leads to less lost trips
than the greedy approach. In Fig. 2, the negotiating
approach has in eight samples better results than the
greedy. In the samples VI & VII neither of the two
approaches cause lost trips. The overall results for
7 vehicle agents are better than for 3. These results
are not only evident when looking at the summary.
As expected, a higher amount of participating vehicle
agents decreases the amount of lost trips. The more
agents are located in the environment, the higher is
the chance that there is an unused vehicle when a new
trip request occurs. Comparing the results of 3 versus
7 vehicle agents, there is a small improvement of the
AT D with 3 agents but a decline with 7 agents. This
is probably due to the fact that 7 agents cause a very
densely populated operational area. Since the amount
of committed trips in both cases are the same, the
ODR for among all samples is significantly smaller
even in cases without negotiation. Obviously, more
available vehicle agents reduce the cases in which the
customer can not be reached in time. However, a
small reduction of ODR can be observed, when nego-
tiation is considered. Furthermore, we presume that
negotiation is worthwhile with less amount of vehi-
cles since the ODR decreases more for 3 vehicles.
The interpretation of the results of our experiments
requires a thorough analysis of the considered utility
2
https://github.com/WI-user/ICAART22-submission
function of the vehicle agent. Concerning the equal
weighting of the three components with
1
3
, we used
this as a preliminary value. Each component of the
utility function, which is explained in Section 3.3, in-
centives the vehicle agent towards a rational behavior.
The distance utility u
dist
reflects the situation that the
closer a vehicle is to the start of the trip request, the
higher is the utility score. Next, the battery utility
u
bat
impedes the vehicle to run out of battery and in-
centives the more balanced usage of all participating
vehicles as well as staying in a healthy battery power
range. For the last component u
pts
, the behavioral in-
tention is to cross out upcoming trips, where the vehi-
cle arrival delay is bigger than the assumed time, the
customers would wait. Those trips are also delegated
to other vehicles with the used utility function.
To sum up, these three utility components reflect
the leaning rational behavior of the vehicles. Al-
though it can be expanded with more components
leading to a more informed behavior, an arising dis-
advantage could be complex interpretation of such re-
sults. As a preliminary approach, we therefore set a
basic utility configuration since our next step would
be integrating a learned behaviour instead of the util-
ity function. Concerning the scalability, a more com-
plex simulation with more trip request samples and a
larger operational area is required.
Table 1: Simulation results containing the sum of all 10 trip
request samples.
config AT D ODR trips lost # trips
neg
3
451.066 1.57% 38 2416
greedy
3
462.064 4.55% 110 2416
neg
7
453.043 0.08% 2 2416
greedy
7
447.881 1.03% 25 2416
5 FUTURE WORK AND
CONCLUSION
In this paper, we have presented a MAS of au-
tonomous vehicles for managing and negotiating trip
assignments in a ride-hailing scenario. An agent
model has been introduced where each vehicle is
represented by a BDI agent. In the ’think’ phase,
the agent makes a decision whether to negotiate on
incoming trip requests based on a utility function,
balancing the customer satisfaction concerning the
pickup time and battery consumption. The negotia-
tion follows the contract net protocol and its process
is evaluated by simulating the agent behavior for ex-
perimental data created from real historical trips. The
experimental results provide a proof of concept for the
application of MAS in a novel MaaS system offering
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
430

Algorithm 1: Pseudocode for the estimated event trace σ
e
(va
j
).
Data: w = (tr
1
,tr
2
, ...,tr
m
) ; // a list of scheduled trip requests
t
0
; // the time when starting to drive
l
0
; // location of agent va
j
at time t
0
θ ; // the threshold for max. tolerable delay
REFILL DURAT ION ; // the average duration of recharge
Result: σ
e
(va
j
) ; // a sequence of estimated events < e
0
, e
1
, ..., e
k
>
e
0
:= (START,t
0
, l
0
,t
0
) σ :=< e
0
> n := 0
for g ← 1 to m do
e type := tr
i
.type
switch e type do
case CUST OMER T RIP do
n++ l := tr
i
.l
start
eta := etd
n−1
+travel time(l
n−1
, l) if eta ≤ tr
i
.VATime + θ then
if eta < tr
i
.VATime then
etd := tr
i
.VATime ; // arrived too early
else
etd := eta
end
e
n
:= (PICKUP, eta, l, etd) σ.append(e
n
) ; // e
n
derived from start of tr
i
n++ l
end
:= tr
i
.l
end
eta := etd +travel time(l, l
end
) e
n
:= (DROP, eta, l
end
, eta) σ.append(e
n
) ;
// e
n
derived from end of tr
i
else
e
n
:= (PASS BY, eta, l, eta) σ.append(e
n
) ; // having missed the customer
end
end
case REFILL T RIP do
n++ l := tr
i
.l
start
eta := etd
n−1
+ travel time(l
n−1
, l) etd := eta + REFILL DURAT ION e
n
:=
(REFILL, eta, l, etd) σ.append(e
n
) // charging completed
end
end
end
Figure 1: Amount of lost trips with 3 processing vehicle
agents.
trip services for customers. Concerning our hypoth-
esis, it can be said that the ODR in the negotiation
approach is less than in the greedy approach. Since
the energy consumption is derived from the AT D,
where it is only less in the case with 7 agents, we
can state that the experiments partially confirmed our
hypothesis. A plausible reason for this circumstance
could be the possibility that the negotiation approach
processes trips with longer distances, which would
be dropped out in the greedy approach and therefore
Figure 2: Amount of lost trips with 7 processing vehicle
agents.
leading to an decrease of its ATD. Our goal for the
future is to use our vehicle agents to control a small
fleet of real autonomous E-Trikes to prove our theory
in real world. We want to extend our agent architec-
ture and add components to make a learning based
BDI agent. During this process, we will add an inter-
mediate layer to interact with a full simulation plat-
form like AMoDeus (Ruch et al., 2018). A complex
simulation will also allow the comparison of our algo-
rithms for the agent behaviour and the utility function
Negotiation in Ride-hailing between Cooperating BDI Agents
431

from other works. The contribution of our work high-
lights the feasibility of an agent-oriented approach
for ride-hailing stipulates multiple lines for future re-
search on the distributed development of MAS.
REFERENCES
Ahadi, R., Ketter, W., Collins, J., and Daina, N. (2021). Sit-
ing and Sizing of Charging Infrastructure for Shared
Autonomous Electric Fleets. AAMAS.
Bazzan, A. L. C. and Kl
¨
ugl, F. (2014). A review on agent-
based technology for traffic and transportation. The
Knowledge Engineering Review, 29(3):375–403.
Bellifemine, F., Caire, G., and Greenwood, D. (2008). De-
veloping multi-agent systems with JADE. Wiley series
in agent technology, Chichester, reprint. edition.
Certicky, M., Jakob, M., Pibil, R., and Moler, Z. (2014).
Agent-based Simulation Testbed for On-demand Mo-
bility Services. Procedia Computer Science, 32:808–
815.
Danassis, P., Filos-Ratsikas, A., and Faltings, B. (2019).
Anytime Heuristic for Weighted Matching Through
Altruism-Inspired Behavior. IJCAI 2019, pages 215–
222.
Deljoo, A. (2017). What Is Going On: Utility-Based Plan
Selection in BDI Agents. The AAAI-17 Workshop
on Knowledge-Based Techniques for Problem Solving
and Reasoning.
DeMers, M. N. (2008). Fundamentals of geographic infor-
mation systems. John Wiley & Sons.
Dlugosch, O., Brandt, T., and Neumann, D. (2020). Com-
bining analytics and simulation methods to assess the
impact of shared, autonomous electric vehicles on sus-
tainable urban mobility. Information & Management,
page 103285.
Erduran, O. I., Minor, M., Hedrich, L., Tarraf, A., Ruehl,
F., and Schroth, H. (2019). Multi-agent Learning
for Energy-Aware Placement of Autonomous Vehi-
cles. In 2019 18th IEEE International Conference On
Machine Learning And Applications (ICMLA), pages
1671–1678, Boca Raton, FL, USA. IEEE.
Fu, K., Meng, F., Ye, J., and Wang, Z. (2020). Com-
pactETA: A Fast Inference System for Travel Time
Prediction. In ACM SIGKDD 2020, pages 3337–3345,
Virtual Event CA USA. ACM.
Jaroslaw Kozlak, S. P. and Zabinska, M. (2013). Multi-
agent models for transportation problems with dif-
ferent strategies of environment information propaga-
tion. PAAMS 2013, Springer Berlin Heidelberg.
Kamargianni, M., Li, W., Matyas, M., and Sch
¨
afer, A.
(2016). A Critical Review of New Mobility Services
for Urban Transport. Transportation Research Proce-
dia, 14:3294–3303.
Kowalski, R. and Sadri, F. (1996). Towards a unified agent
architecture that combines rationality with reactivity.
Logic in Databases, 1154.
Kowalski, R. and Sadri, F. (1999). From Logic Program-
ming towards Multi-agent systems. Annals of Mathe-
matics and Artificial Intelligence, 25(3/4):391–419.
Malas, A., Falou, S. E., and Falou, M. E. (2016). Solv-
ing On-Demand Transport Problem through Negotia-
tion. Proceedings of the Summer Computer Simula-
tion Conference, page 7.
Pavone, M. (2015). Autonomous Mobility-on-Demand Sys-
tems for Future Urban Mobility. In Maurer, M.,
Gerdes, J. C., Lenz, B., and Winner, H., editors, Au-
tonomes Fahren, pages 399–416. Springer Berlin Hei-
delberg.
Qin, Z. T., Tang, X., Jiao, Y., Zhang, F., Xu, Z., Zhu, H.,
and Ye, J. (2020). Ride-Hailing Order Dispatching at
DiDi via Reinforcement Learning. INFORMS Journal
on Applied Analytics, 50(5):272–286.
Rao, A. S. and Georgeff, M. P. (1995). BDI Agents: From
Theory to Practice. ICMAS.
R
¨
ub, I. and Dunin-Keplicz, B. (2020). Basta: Bdi-based
architecture of simulated traffic agents. Journal of In-
formation and Telecommunication, 4(4):440–460.
Ruch, C., Horl, S., and Frazzoli, E. (2018). AMoDeus, a
Simulation-Based Testbed for Autonomous Mobility-
on-Demand Systems. In 2018 ITSC, pages 3639–
3644, Maui, HI. IEEE.
Silva, L. d., Meneguzzi, F., and Logan, B. (2020). BDI
Agent Architectures: A Survey. In IJCAI 2020, pages
4914–4921, Yokohama, Japan.
Smith, R. G. (1980). Communication and Control in a Dis-
tributed Problem Solver. IEEE Transactions On Com-
puters.
Yu, M. and Zhang, Y. (2010). Multi-agent-based Fuzzy Dis-
patching for Trucks at Container Terminal. Interna-
tional Journal of Intelligent Systems and Applications,
2(2):41–47.
Zhang, H., Sheppard, C. J., Lipman, T. E., Zeng, T., and
Moura, S. J. (2020). Charging infrastructure demands
of shared-use autonomous electric vehicles in urban
areas. Transportation Research Part D: Transport and
Environment, 78.
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
432