An Agent-based Electronic Market to Help Airlines to Recover from
Delays
Lu
´
ıs Reis
1
, Ana Paula Rocha
2
and Antonio J. M. Castro
2
1
FEUP, University of Porto, Portugal
2
LIACC, FEUP, DEI, University of Porto, Portugal
Keywords:
Electronic Market, Multi-agent Systems, Negotiation, Case-Based Reasoning, Air Transport, Disruption
Management, Flight Disruption, Irregular Operations, Recovery Process.
Abstract:
The Airline Operations Control Center (AOCC) has the responsibility to ensure that flights meet their planned
schedule or, if any problem arises, to find a viable solution that minimizes both the impact in the operational
plan and its cost. The high cost of resources involved in this process (aircraft and crew members) leads to a lack
of additional resources from the airline companies, implying a restricted solution space. Here, we propose an
electronic market modeled as a multi-agent system where airline companies can negotiate and lease each other
the required resources when solving a disruption problem, thus expanding their solution space. The proposed
negotiation occurs in several rounds, where qualitative comments made by the buyer agent on proposals sent by
the sellers enables these to learn how to calculate new proposals, using a case-based reasoning methodology.
1 INTRODUCTION
According to Kohl (Kohl et al., 2004), ”research on
the recovery operation to this date only deals with
a single airline. Cooperation between airlines is not
supported”. Nowadays, each airline tries to solve
the operations recovery problems with their own re-
sources (Castro et al., 2014). If they have an open
position for a specific type of crew in a flight, they try
to find a suitable one from their own staff. The same
happens with aircraft.
Sometimes, the airlines have to rent aircraft and
crew members when needed (known in the industry as
ACMI - Aircraft, Crew, Maintenance and Insurance),
but through a direct contact with charter airlines. It is
not a usual practice to use only crew members (with-
out being part of the aircraft) from other companies.
The electronic market (EM) that we propose in
this paper, is a permanently open virtual market-
place where registered airlines (represented by soft-
ware agents) can meet each other to purchase services
and has the possibility to be integrated with systems
or tools for airline operations control, like the one we
use in this paper. It has the following advantages:
• Airlines that participate in this EM will have more
resources available to solve their problems.
• Airlines may take advantage of exceeding re-
sources in specific dates and times and sell ser-
vices performed by these resources to other air-
lines.
• Can reduce costs and time for the airline that has
a specific problem.
When compared with the work of Malucelli
(Malucelli et al., 2006) we complete the work by
proposing a negotiation algorithm for the EM.
Airlines have an organization called Airline Oper-
ations Control Center (AOCC) that has the responsi-
bility to ensure that flights meet their planned sched-
ule or, if any problem arises, to find a viable solution
that minimizes both the impact in the operational plan
and its cost. Research in the air transportation domain
has shown that airline companies lose between 2% to
3% of their annual revenue as consequence of disrup-
tions and, that, the impact caused by small disruptions
in companies’ profits can be reduced by at least 20%,
through a better recovery process (Chen et al., 2010).
Currently, operations management is essentially
a manual process, supported by tools, that among
other functions include monitoring, event detection
and problems resolution and, strongly depends on the
tactical knowledge of the AOCC’s members (Castro
et al., 2014).
Every time an irregular event that has an impact
on the scheduled plan is detected, the AOCC’s team
176
Reis, L., Rocha, A. and Castro, A.
An Agent-based Electronic Market to Help Airlines to Recover from Delays.
DOI: 10.5220/0006582401760183
In Proceedings of the 10th International Conference on Agents and Artificial Intelligence (ICAART 2018) - Volume 1, pages 176-183
ISBN: 978-989-758-275-2
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

has to plan carefully an alternative schedule, to ensure
that it minimizes at most the disruption cost. A dis-
ruption can be view as composed by four dimensions
(Castro et al., 2014): aircraft, crew member, passen-
ger and flight. In disruption management, the AOCC
commits to recover all the dimensions affected.
To test our proposed EM, we use MASDIMA
(Multi-Agent System for Disruption Management)
(Castro et al., 2014) that addresses the Aircraft, Crew
and Passenger recovery problem using an approach
that is able to recover all problem dimensions simul-
taneously.
In Multi-Agent Systems it is required for an agent
to interact with other agents whom may not share
common goals. This leads to the need to reach agree-
ments (Wooldridge, 2009) through an automated ne-
gotiation process. The existent automated negotiation
systems are composed of three major groups (Oliveira
and Rocha, 2001):
• Auctions;
• Game Theory;
• Negotiation.
Although it only considers a single attribute, due to its
simplicity and well predefined rules, auctions (Vulkan
and Jennings, 2000) are a very popular negotiation
mechanism.
Game Theory (Rosenschein and Zlotkin, 1994) is
a mechanism that can only be applied to perfect infor-
mation and rationality contexts.
Negotiation is the generic name given to other
techniques where agents must reach agreements on
matters of mutual interest (Wooldridge, 2009). These
techniques are more flexible than auctions and game
theory in terms of preexistent protocols and rules, thus
more suitable for open and dynamic environments
(Oliveira and Rocha, 2001).
From the Negotiation group, stands out the multi-
attribute negotiation, which is useful in the situation
where the negotiation decision does not consider only
one attribute but multiple attributes as it is in the case
of AOCC. For instance, when buying any product, the
buyer considers the price as an important attribute in
its decision but the delivery time or the product qual-
ity may also be (usually are) factors to be considered
in the decision of buying or not a certain product.
Giving different utility values to the different
attributes under negotiation solves the problem of
multi-attribute evaluation. The most common pro-
posal evaluation formula is a linear combination of
the attribute correspondent values, weighted by the
respective utilities. Therefore, a multi-attribute ne-
gotiation is converted to a single-attribute one, to be
made over the evaluation value (examples of this are
the work in (Oliveira et al., 1999), (Vulkan and Jen-
nings, 2000), (Matos et al., 1998) and (Cardoso and
Oliveira, 2000)). This is also the approach followed
in our work, although we agree that, in some cases, it
can be difficult to give an exact numeric value to an
attribute utility. A solution which leads to a more intu-
itive situation, can be just to impose a preferential or-
der over the domain values for the different attributes
or on the attributes itself.
The multi-agent system based Electronic Market,
presented in this paper, allows companies to negoti-
ate among themselves the missing resources. The ne-
gotiation algorithm includes case-based reasoning to
learn how to make a counter-proposal. According to
Riesbeck and Schank (Riesbeck and Schank, 2013),
”A case-based reasoner solves problems by using or
adapting solutions to old problems.”, i.e. case-based
reasoning (CBR) focuses on the reuse of knowledge
acquired from previous experiences in order to solve
new problems. Like humans do, CBR is a problem
solving paradigm that uses incremental and sustained
learning since new experiences are retained each time
a problem is solved making those available for future
problems. A negotiation algorithm with a CBR ap-
proach has never been considered in the works men-
tioned.
The rest of this paper is as follows: section 2 is the
main section and presents the proposed multi-agent
system EM. In section 3 it is presented the scenar-
ios but only one of the experiments done, as well as
the results obtained. Finally, section 4 concludes the
work presented.
2 AIRLINE ELECTRONIC
MARKET SOLUTION
When a disrupted flight is detected, the AOCC’s team
should find an alternative trying to minimize both the
delay and the disruption cost. The airline electronic
market proposed here intends to help the airline com-
pany in this disruption management process, by al-
lowing to find external resources, possibly less costly
or available sooner than the company’s own. In this
market there are two types of entities:
• The buyer, that represents an injured airline com-
pany. This is the airline company that has a dis-
rupted flight (an unexpected event causing a delay
in the flight).
• The seller, that represents a service provider air-
line company
Being the object under negotiation, a Need is iden-
tified by the resource(s) needed: a list of crew mem-
An Agent-based Electronic Market to Help Airlines to Recover from Delays
177

bers and an aircraft fleet as well as relevant informa-
tion related to the disrupted flight (scheduled depar-
ture time, trip time, delay, origin and destination), as
depicted in equation (1).
Need =< Res, ST D, TripD, Del, Orig, Dest > (1)
with Res =< CrewList, Aircra f t Fleet >
where:
• Res is the resource(s) needed
• ST D is the trip duration
• Del is the delay of the disrupted flight
• Orig is the airport origin
• Dest is the airport destination
Buyer and sellers will negotiate the resource that
buyer identifies as its need. This resource can be a
set of crew members, an aircraft or both. When the
resource under negotiation is an aircraft, it is required
that a crew to handle it should also be provided. For
the negotiation to take place, buyers and sellers need
to know each other. Sellers should register first, other-
wise will be no one in the market to be asked for some
resource(s). So, the first step is to have multiple sell-
ers registered and wait for some buyer to register too.
When a buyer registers in the market, it retrieves a list
containing all registered sellers and starts a negotia-
tion with them. The negotiation is a process where
proposals are exchanged between buyer and sellers
until an agreement is reached between the buyer and
one of the sellers, and the negotiation ends success-
fully, or no agreement is reached and the negotiation
fails.
2.1 An Adaptive Negotiation
The negotiation protocol proposed for the airline elec-
tronic market is based in the FIPA Iterated Contract
Net and was chosen because it allows multi-round it-
erative bidding. This way, it is ensured that a wide
space of solutions is subject to discussion and refine-
ment, as it is the case of humans’ negotiations. This
protocol works with an initiator (buyer in this case)
and multiple responders (sellers ).
The negotiation protocol is depicted on figure 1.
The buyer initiates the negotiation, by sending to all
sellers an Invitation message, (CFP - Call For Pro-
posals), containing relevant information about the dis-
rupted flight and resource needs (as indicated in (1)).
After sending the Invitation message, the buyer gives
a timeout for sellers to respond. If a seller did not
respond after that timeout, it is removed from the ne-
gotiation.
When a seller receives the Invitation message, it
processes the message verifying if it is able to pro-
vide the required resources or not. If yes, the seller
replies with a Proposal message, containing the price
and availability of its proposal (equation (2)).
Proposal =< α, ρ > (2)
with α ∈ [0, delay o f disrupted f light]
where:
• α is the proposed availability
• ρ is the proposed price
If the seller is not able to provide the resources re-
quired by the buyer, it replies with a Refusal message.
The first round is now concluded and until the end
of the negotiation, all rounds are processed the same
way, explained as follows.
Buyer Role: Buyer receives one proposal from
each interested seller, evaluates all proposals and se-
lects the best one of the current round according to its
utility (agents’ utility is explained in section 2.2). If
the best proposal of the current round is better than
the best one found in previous rounds (if any), it is
considered the new best proposal. If not, the best pro-
posal remains unchanged. Buyer creates then a reply
for each received proposal, issuing a qualitative feed-
back over the availability and price in it, by comparing
these values with the ones in the best proposal. This
reply or Feedback message is send to all sellers that
are currently in the negotiation (equation (3)). The
qualitative comment included in the Feedback mes-
sage (QlEv in equation (3)) can assume one of the
three options:
• OK: means there is no need to improve the at-
tribute that received this feedback
• LOWER: means the attribute that received this
feedback has a high value, should be reduced
• MUCH LOWER: means the attribute that re-
ceived this feedback has a very high value, should
be greatly reduced
Feedback =< QlEv
α
, QlEv
ρ
> (3)
where:
• QlEv
α
is the qualitative evaluation of the pro-
posed availability
• QlEv
ρ
is the qualitative evaluation of the pro-
posed price
Seller Role: When a seller receives the feedback
for the proposal sent, it updates its experience history,
by recording and reasoning the concerned feedback.
Sellers will use its experience history (similar to what
humans do) to formulate new proposals during the
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
178

current and future negotiations. If a seller does not
have any more proposals to propose, it sends a Refuse
message. This process is explained in more detail in
section 2.3.
The negotiation is over when all sellers have sent
a refusal message or a deadline is reached. In the last
round of the negotiation, buyer sends an Accept mes-
sage to the best proposal’s owner, with the accepted
proposal data and a Reject message to all others. The
seller that received the accept message sends back to
buyer a Termination message with all relevant data
about the Need. Upon receiving the termination mes-
sage, buyer unregisters himself from the market as the
negotiation is over.
Note that messages exchanged during all the nego-
tiation ensure that agents’ (buyers or sellers) informa-
tion is kept private. Agents never reveal their costs or
utility. For instance, if sellers would know the buyer’s
disruption cost, their strategy would be to ask for a
price slightly lower than that cost, making the market
an unpractical alternative for buyer.
Figure 1: Negotiation Protocol.
2.2 Agents’ Utility
Agent’s utility is a way of representing its preferences
over a set of possible alternatives, in this case, a set of
possible Need, which will constitute the negotiation
set. The higher the utility, more preferred is the cor-
respondent Need.
Buyer Utility. Buyer uses the utility to evaluate the
proposals sent by the different sellers in response to
its Invitation or Feedback messages. Remember that a
proposal tells the buyer what are the conditions under
which a seller provides a Need, and contains two val-
ues: the proposed availability for the disrupted flight,
and the rental price. In buyers perspective, the utility
of a proposal must measure its availability and price,
where it tries to minimize both. Its value ranges from
0 to 1, and its calculation follows equation (4).
µ = µ
α
× β + µ
ρ
× (1 − β) (4)
with β ∈ [0, 1] where:
• µ is the utility of the proposal
• µ
α
is the utility of the availability parameter [0, 1]
• µ
ρ
is the utility of the price parameter [0, 1]
• β is the importance factor of the availability pa-
rameter
As shown in equation (4), to the buyer, the utility
of a proposal is composed of two parcels, the avail-
ability utility (µ
α
) and the price utility (µ
p
). The main
purpose of the availability utility (µ
α
) is to relate the
availability proposed by the seller with the delay of
the disrupted flight.
Seller Utility. In the case of sellers, the measure of
how preferable a proposal is, is given exclusively by
its price. This is materialized in the fact that sellers
do not need to fulfill any disrupted schedule but it just
has to compensate the leasing associated cost. From
the seller point of view, a proposal is more profitable
the higher is the price. The seller utility is given by
equation (5), and its value ranges from 0 to 1.
µ =
ρ − γ × ζ
(σ × ζ) − (γ × ζ)
(5)
where:
• µ is the utility of the proposal
• ρ is the proposed price
• γ is the minimum price multiplier
• σ is the maximum price multiplier
• ζ is the leasing associated cost
The price multiplier is a static interval generated to
ensure that seller is not impaired, what would happen
if the price was only the leasing associated cost. This
way, it is ensured that seller gets some profit even with
a low utility deal.
An Agent-based Electronic Market to Help Airlines to Recover from Delays
179

2.3 Agents Learning through
Case-based Reasoning
Sellers use CBR (Case-based Reasoning) to decide
what to do upon receiving the buyer feedback over the
proposal they have sent, consulting a record of previ-
ous experiences classified according to its usefulness.
The object that represents an experience, along with
its usefulness, is called case and is represented by a
set of parameters, grouped into three types (Features,
Solution and Evaluation), as shown in figure 2.
The parameters in Features identify the situation of
the current case, regarding the feedback buyer gave,
the number of sellers in the negotiation and the iden-
tification of the resource under negotiation (aircraft or
crew). The parameters in Solution identify the actions
performed (price changing, availability changing) in
that specific situation. The parameter in Evaluation
assigns an evaluation value to the case, that measures
its usefulness.
Case
Features
Price Availability Number
Sellers
Resource
Asked
Solution
Price
Action
Availability
Action
Evaluation
Evaluation
Value
Figure 2: Case Composition.
Find Similar Cases
CBR starts by retrieving similar cases to the one re-
ceived. For this purpose, only parameters identified
as Features in figure 2 are used to compare cases and
to identify equal ones. Although all features are used,
they do not have the same preponderance on the task,
because the feedback over a proposal is more relevant
to decide the action to be made than the others pa-
rameters. So, to each parameter in Features is given a
weight. Similar cases are found through the euclidean
distance between them, a distance of 0 means that the
case is identical, a distance greater than 0 means a
different case. To ensure that features’ weight has
relevance in the distance calculation, the weight was
added to the well known euclidean distance formula.
So, the distance between two cases is the weighted
sum of their features distances, as presented in equa-
tion (6).
d(κ, χ) =
∑
η
q
(κ
η
− χ
η
)
2
× ε
η
(6)
with η ∈ {Features}
where:
• κ is the case received
• χ is one of the case in the data set
• η is the current feature
• ε
η
is the weight of feature η
Select a Case
After identify the set of similar cases, the seller has
to select one of them to apply at the current situation,
what it does using a softmax algorithm (Sutton and
Barto, 1998). This algorithm applies a probability to
each similar case retrieved from CBR, where a case’s
probability is greater the higher its evaluation value.
The probability of a case being selected is given by
equation (7).
P(k) =
e
ϒ
k
∑
n
i=1
e
ϒ
i
(7)
where:
• ϒ
i
is the evaluation of case i
• P(k) is probability of the item k being selected
• n is the number of similar cases
After being assigned a probability to each case,
a random value is generated and the first case with
cumulative probability greater than that random value
is selected. The new proposal to be sent by the seller
is generated by applying the actions enumerated in the
selected case.
If no similar cases exist in the history set, the seller
generates a new proposal by following the qualitative
comments in the feedback received.
Update History
In order to prevent obsolete cases, every time an expe-
rience is reproduced, its evaluation is updated, where
the latter the experience, the more important its eval-
uation is. The evaluation of an experience is updated
as equation (8) shows.
ϒ = ϒ
n−1
∗ (1 − I ) + ϒ
n
∗ I (8)
where:
• ϒ is the updated evaluation value
• ϒ
prev
is the evaluation value of an equal experi-
ence found in the history set
• ϒ
curr
is the evaluation value for the current expe-
rience
• I is the weight given to the most recent experi-
ment
If there is no previous experience equal to the
current one in the history, the evaluation is simply:
ϒ = ϒ
curr
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
180

Evaluate a Case
The evaluation of the current experience ϒ
curr
is cal-
culated as the difference between the feedback over
previous round proposal and the current round pro-
posal, as presented in equation (9).
ϒ
curr
= ∆ρ
f eedback
+ ∆α
f eedback
(9)
where:
• ∆ρ
f eedback
is the price feedback variation;
• ∆α
f eedback
is the availability feedback variation.
If the feedback variation is greater than 0, the eval-
uation is incremented by 0.5 for each attribute. This
means that in the worst scenario, where feedback re-
mains unchanged, evaluation is 0. If only one of the
feedback values changed, evaluation is set to 0.5 and
if both changed, best scenario, evaluation is set to 1.
3 EXPERIMENTS AND RESULTS
To validate our proposal, we have used data provided
by a TAP Air Portugal expert in disruption manage-
ment, regarding disruptions and solutions found for
real problems. Each test reflected a disruption and as-
sorted solution possibilities.
The data provided to test the electronic market is
composed by 12 disruptions where each disruption
contained a considerable amount of fields of which
stand out the ID, delay, cost, disrupted resource, as
well as the estimated departure time and number of
passengers. The number of crew members of each
category (captain, first office, senior cabin crew and
flight attendant) was also included.
The following metrics will be used to measure the
benefit of the solution’s found with the electronic mar-
ket:
• Buyer utility;
• Seller utility;
• Delay reduction;
• Price reduction.
To evaluate price and delay reductions and both
agents utility variations, three different experiments
were executed. The first experiment considered equal
weights for the attributes price and availability. The
second experiment valued the availability with a
weight of 80% and the price with a weight of 20%
in the utility calculation. The third and last experi-
ment showed an inversion regarding the values of the
second one, i.e. availability with a weight of 20% and
the price with a weight of 80% in the utility calcula-
tion. In all experiments, disruption number 12 has no
results to present because does not exists in seller’s
data set any resource similar to the one required, so
seller gives up the negotiation. Due to paper space
limitations, we will only describe the first experiment
in section 3.1. However, in section 4, the results pre-
sented consider the three experiments.
3.1 Experiment 1 (50/50 Experiment)
In this experiment, it was expected to find a similarity
between cost and delay reductions, given that both are
equally valued in the buyer’s utility calculation. The
chart in figure 3 presents the results obtained.
-1
-0,8
-0,6
-0,4
-0,2
0
0,2
0,4
0,6
0,8
1
0 1 2 3 4 5 6 7 8 9 10 11 12
Parameter Value
Disruption Number
Experiment 1 - Results
Seller Utility Buyer Utility Delay Reduction(%) Cost Reduction(%)
Figure 3: Experiment 1 Results.
The points below zero are solutions that had a
price greater than the disruption cost, or in the case
of agents’ utilities, unfeasible solutions, like disrup-
tion number 6 that has a negative value for buyer’s
utility and for this reason it is not a viable solution.
The fact that there is no delay reduction or the seller’s
utility being zero supports that statement, so this is a
case that would never result in a leasing contract.
One interesting fact that can be extracted is that
seller’s utility and delay reduction lines are over-
lapped, which means that have the same values for
every disruption. This is explained by both agents’
utility formulas, detailed in section 2.2.
The results obtained for this experiment show that
with equal weights in the utility calculation, all but
one disruption had its delay minimized by at least
68%. Relatively to the cost reduction the results are
not as good as for availability, which is explained by
the need to minimize a flight’s delay. If the delay is
largely reduced, as showed, then the buyer is willing
to pay more than the disruption cost. Regarding this
experiment results, the average values for each of the
metrics used are presented in table 1.
Although the average delay reduction is good
(74.91%), there is a cost incresase (6.55%) intead of
a cost reduction. Concerning to utilities, the aver-
An Agent-based Electronic Market to Help Airlines to Recover from Delays
181

Table 1: Experiment 1 - Average values.
Seller
Utility
Buyer
Utility
Delay
Reduction (%)
Cost
Reduction(%)
0.75 0.35 74.91 -6.55
age seller utility reveals that the electronic market is
highly useful, at least for this experiment. Regarding
the average buyer utility, it shows some improvement
but in some cases at a great cost, which explains the
considerable difference between seller and buyer util-
ities.
3.2 Results
The electronic market does not consider any costs un-
related to the disrupted resources. However, in order
to choose the most cost-effective solution, passenger
related costs must be considered after the market re-
turns its solutions. For instance, the number of pas-
sengers that will miss a flight connection due to the
delay carries an extra cost to the injured company
(passenger cost) and will affect the passenger satis-
faction, which also carries an extra cost to the com-
pany (passenger goodwill cost). These costs will be
added to the aircraft and crew costs, being distributed
as follows:
• Direct Costs: Aircraft cost plus crew cost plus
passenger cost;
• Integrated Solution Costs: Passenger goodwill
cost times passenger goodwill weight plus direct
costs.
All these costs are considered by the human specialist
(at the AOCC) when it must choose a solution to a dis-
ruption in its daily operation. This section intends to
compare the solutions found by the electronic market
to the ones chosen by a human specialist, by present-
ing the electronic market solutions to the human for
him to analyze and validate. The passenger goodwill
weight is 5, by default, according to the specialist.
The first step in the comparison between the solu-
tions found by the electronic market and the ones cho-
sen by a human specialist is to see how the three so-
lutions (one of each experiment) obtained in the elec-
tronic market impacts in the flight delay and in the
number of passengers missing the flight connections.
The second step is to see the disruption and each
solution cost and its influence on the passenger and
passenger goodwill costs.
The third step is to see the costs without consid-
ering the electronic market solutions: original direct
costs and original integrated solution cost, i.e. the
original integrated solution cost is the sum of aircraft,
crew and passenger costs to which is added the re-
sult of the multiplication between passengers good-
will cost and its weight, as shown in equation (10).
IC
orig
= c
a
+ c
cr
+ c
pax
+ (c
paxgw
× w
gw
) (10)
where:
• c
a
is the aircraft cost;
• c
cr
is the crew cost;
• c
pax
is the passenger cost;
• c
paxgw
is the passenger good will cost;
• w
gw
is the weight of good will.
The new integrated costs represent the integrated
costs of the electronic market solutions while the orig-
inal integrated costs represent the company solution
integrated costs. The final step is to see if there are
savings provided by each one of the electronic mar-
ket solutions because the specialist always chooses
the solution with a higher value of integrated savings.
After being introduced the methodology used to cal-
culate the Integrated Savings, the results over all dis-
ruptions are presented in figure 4.
0
10000
20000
30000
40000
50000
60000
70000
80000
90000
100000
110000
120000
130000
140000
150000
160000
170000
180000
50-50 -CSTTJ 80-20 - CSTTJ 20-80 - CSTTJ 50-50 - CSTQD 80-20 - CSTQD 20-80 - CSTQD 50-50 - CSTTP 80-20 - CSTTP 20-80 - CSTTP 50-50 - CSTJG 80-20 - CSTJG 20-80 - CSTJG 50-50 - CSTTK 80-20 - CSTTK 20-80 - CSTTK 50-50 - CSTNL 80-20 - CSTNL 20-80 - CSTNL 50-50 - CSTNJ 80-20 - CSTNJ 20-80 - CSTNJ 50-50 - CSTNN 80-20 - CSTNN 20-80 - CSTNN 50-50 -CSTJF1 80-20 - CSTJF1 20-80 - CSTJF1 50-50 - CSTTU 80-20 - CSTTU 20-80 - CSTTU 50-50 - CSTJF2 80-20 - CSTJF2 20-80 - CSTJF2 50-50 - CSTNM 80-20 - CSTNM 20-80 - CSTNM
Monetary Units (€)
Disruption - EM Scenario
Costs and Savings
Original Integrated Solution Cost New Integrated Solution Cost
Figure 4: Costs and Savings.
As shown, the solutions obtained through the elec-
tronic market are more cost-effective than the com-
pany’s solutions, except in the disruption which is
identified by CSTJF that has no similar resources in
the electronic market. When comparing the chosen
solution from the electronic market with the disrup-
tive solution, the electronic market solutions present
an average delay reduction of 66.85% and an average
cost reduction of 63.51%. Disregarding the CSTJF
disruption, there is at least one solution obtained (con-
sidering the three experiments made) in the electronic
market that is more cost-effective than the disruptive
solution for each disruption, having a total of seven
disruptions minimized in each experiment.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
182

4 CONCLUSIONS AND FUTURE
WORK
In this paper, it is proposed an electronic market
modeled as a multi-agent system to expand a com-
pany’s solution space regarding disruptions manage-
ment. This electronic market provides alternative
solutions to companies affected by disruptions, us-
ing resources from other companies (and, as such,
contributing to increase collaboration between air-
lines), which is achieved through automated negotia-
tion, where agents negotiate the resource’s availability
and price for a disrupted flight. Human validation (at
the AOCC) is also included to compare the solutions
obtained through the EM with the ones obtained with
the company’s own resources. The Seller agent in the
EM uses case-based reasoning to reuse or adapt pre-
vious experiences, to the current negotiation, which is
also a contribution of our work.
Three different scenarios were tested to validate
the concept, as described in section 3. As there were
no available resources for only one disruption in the
electronic market, the success rate is 91.7% consider-
ing the cost reduction parameter and 67.7% consider-
ing both cost and delay minimized.
Possible future directions to improve this work,
could include firstly, different approaches in the
whole process of identifying previous similar expe-
riences (by the seller), like machine learning and q-
learning in order to understand how the agent learn-
ing process influences the negotiation, either in terms
of proposals’ price and availability or in terms of util-
ity for each agent. The methodology used (CBR) can
also be improved by creating better evaluation scenar-
ios and benefiting the accepted proposal (or the tree of
the proposals that lead to the accepted one).
Secondly, the usage of heuristics to combine re-
sources instead of doing all possible combinations,
would be an interesting feature to include. The usage
of clustering algorithms to classify resources (where
the parameters would be availability and/or price) in
order to have a better and more efficient resource
combination is also something to explore.
Finally, it would be worthy to use trust mod-
els to evaluate the electronic market outcome when
considering the relations established between agents
and whether that trust measure would influence the
agents’ behaviour.
REFERENCES
Cardoso, H. L. and Oliveira, E. (2000). A platform for
electronic commerce with adaptive agents. In Interna-
tional Workshop on Agent-Mediated Electronic Com-
merce, pages 96–107. Springer.
Castro, A. J., Rocha, A. P., and Oliveira, E. (2014). A New
Approach for Disruption Management in Airline Op-
erations Control, volume 562. Springer-Verlag Berlin
Heidelberg.
Chen, X., Chen, X., and Zhang, X. (2010). Crew schedul-
ing models in airline disruption management. In 2010
IEEE 17Th International Conference on Industrial
Engineering and Engineering Management (IE&EM),
pages 1032–1037. Conference Publications.
Kohl, N., Larsen, A., Larsen, J., Ross, A., and Tiourline,
S. (2004). Airline disruption management - perspec-
tives, experiences and outlook. Carmen Research and
Technology Report CRTR-0407.
Malucelli, A., Castro, A. J. M., and Oliveira, E. (2006).
Crew and aircraft recovery through a multi-agent elec-
tronic market. In Krishnamurthy, S. and Isaias, P., ed-
itors, Proceeding of IADIS International Conference
on e-Commerce 2006, pages 51–58, Barcelona Spain.
IADIS Press.
Matos, N., Sierra, C., and Jennings, N. R. (1998). Determin-
ing successful negotiation strategies: An evolutionary
approach. In Multi Agent Systems, 1998. Proceedings.
International Conference on, pages 182–189. IEEE.
Oliveira, E., Fonseca, J. M., and Steiger-Garc¸
˜
ao, A.
(1999). Multi-criteria negotiation in multi-agent sys-
tems. CEEMAS’99, page 190.
Oliveira, E. and Rocha, A. (2001). Agents advanced fea-
tures for negotiation in electronic commerce and vir-
tual organisations formation process. Agent Mediated
Electronic Commerce, pages 78–97.
Riesbeck, C. K. and Schank, R. C. (2013). Inside case-
based reasoning. Psychology Press.
Rosenschein, J. S. and Zlotkin, G. (1994). Rules of en-
counter: designing conventions for automated nego-
tiation among computers. MIT press.
Sutton, R. S. and Barto, A. G. (1998). Reinforcement learn-
ing: An introduction, volume 1. MIT press Cam-
bridge.
Vulkan, N. and Jennings, N. R. (2000). Efficient mecha-
nisms for the supply of services in multi-agent envi-
ronments. Decision Support Systems, 28(1):5–19.
Wooldridge, M. (2009). An introduction to multiagent sys-
tems, pages 15–23,105–111,129–148. John Wiley &
Sons.
An Agent-based Electronic Market to Help Airlines to Recover from Delays
183