2354
A Roadmap for Ambient E-Service
ambient environments. The ambient e-service
DSSOLFDWLRQVDUHFDWHJRUL]HGUHÀHFWLQJWKHLBS
types (transaction service, information service,
navigation, and tracking service and safety ser-
vice) as well as exhibiting dynamic collected
efforts based on the peer-to-peer design.
:H¶OOWDNHWKH³$PELHQWVKRSSLQJPDOOVFH-
nario” for example. The Ambient shopping mall
VFHQDULRLVFODVVL¿HGDVDGLVWULEXWHGWUDQVDFWLRQ
service. In a shopping mall (fully equipped with
wireless network infrastructures), information
items (e.g., advertisement or sales promotion
information) can be broadcast to passing-by
peers with information broadcast station. Peers in
different locations receive different information
items depending on their preference. This means
the experience and attained information items
of peers are different to their locations and user
context (e.g., user preference or interest). That
is, the attained information items of a peer vary
based on the peer’s interactions with the shopping
mall ambient environments. The customers are
not required to go all over the shopping mall to
receive the broadcast information items, but just
pay little money to acquire a suitable information
items service package based on their preferences.
Alternatively, a mobile agent peer for customer
can inquire with nearby peers for what they want
and proceed a bartering process. This will help

WKHQHZHQWHULQJFXVWRPHUVLQKXUU\WRHI¿FLHQWO\
acquire the shopping mall information.
This scenario delineates not only the case
of new customers with high buyer perishability
(entering the shopping mall and being in a rush
to buy certain items without the knowledge of
where to buy and how to buy cheap given relevant
sales promotion), but also carry out the collective
efforts of mobile users (e.g., collective bargaining,
collective buy, or some collective agreement).
Through a transaction e-service, information
items (e.g., e-coupons) can be distributed not
only to the mobile users falling into the broadcast
range of the distributor companies, but also to the
primary target peers (who really need the certain
e-coupons in the right time and right place).
However, in such an ad-hoc structured
environment, peers might not recognize each
other. Should we trust the entire information
sources? There are no evidences that all the peers
are trustable. What if there is someone trying to
acquire my sensitive information? There is a pos-
sibility of act of swindling; hence, users should
protect themselves from any possible forms of
harm. Considering a mobile user’s willingness of
participation, the safety and privacy issue remain
t h e m a jo r c o n c e r n . I f t h e n u m b e r o f p a r t i c i p a n t s o f
an e-service diminishes, the e-service application
would collapse. Accordingly, a convenient and safe
environment would encourage users to participate

and interact with each other.
Since different e-service applications should
cope with different circumstances and bear
different restrictions, the framework of ASEM
outlines the guidelines for the implementation
of ambient e-service. For example, if the trading
process employs a bartering mechanism (that
does not involve real money), the required level
of trust is comparatively lower than those using
RI¿FLDOFXUUHQF\,QRWKHUZRUGVYDULRXVDPELHQW
e-service applications are of particularly different
concerns of the factors outlined in ASEM and lead
to different ambient e-service implementation.
ASEM also enables diminishing the chance
of fraud and deceit. Mobile users can obtain
necessary decision information of certain assured
quality from nearby sources (e.g. mobile users,
service providers). Different information sources
are exerted to facilitate great utilities derived in
b eh al f of u se rs. On ce t he r isk le vel of t r a ns ac tio ns
can be curtailed, the convenience interactions of
ambient e-services would be more prevalent and
aggrandizing the chances of realizing the power
of the collective efforts between mobile users.
Platform Design Domain
In this section, we bring up some ideas for the
future design of ambient e-service platforms (that
constitute dynamic identity management, ambient
2355
A Roadmap for Ambient E-Service

data access control, seamless unlinkability
management, and convenience data access
control).
Dynamic identity management and ambient
data access control particularly concern the
nature of an ambient e-service’s environment
(e.g., wireless communication distance, handheld
device storage capacity, and temporary identity).
For the example of the shopping mall scenario
application, the communication ranges of the
ad-hoc wireless networks centered on a mobile
user vary from place to place. The mobile user is
required to update the surrounding nearby peer
list at their current location, and check if there are
RWKHUSHHUV¶RIIHULQJV¿WWLQJWKHLUQHHGV7KDWLV
the information update of the surrounding peers
is necessary. However, the desired update type
(update frequency) varies between applications. In
the shopping mall application, it is not necessary
to engage a constant update of the list of the
surrounding peers because the movement of a
user often is not so fast. Accordingly, a periodical
update type is a right choice of the update-type
design for the ambient shopping mall scenario.
This short-term lifetime identity is a unique
property in ambient environments. As men-
tioned in the ASEM section, Dynamic Identity
Management aims to issue different identities
for a mobile peer when the peer leaves the en-
vironment and re-enters the environment again

(even though the mobile device used is the same).
However, existing P2P systems/solutions are
still with long-lived identities. For instance, as
addressed in Resnick, Zeckhauser, Friedman, and
Kuwabara (2000), reputation systems generally
take on three properties: (1) entities are long-lived;
(2) feedback about current interactions is captured
and distributed; (3) past feedback guides buyer
decisions. In other words, the identities in ambi-
ent e-service environments are short-lived and
localized, and thus existing methods/solutions
requiring long-lived identities can not be applied
to our environments.
On the other hand, the nature of ambient
environment (localized/short period lifetime’s
identities) could result in the material change
of the reputation’s basics as well as other is-
sues (trust/traceability/privacy). How to derive
a reputation system coping with the nature of
ambient environments accordingly becomes new
a challenge to straighten out.
Regarding seamless unlinkability management,
the requirements for different ambient e-service
scenarios are also different. In the ambient
shopping mall scenario, if a transaction involves
just information items, the amount of necessary
information required would be less than those
transactions that involve real money. Required
security level for privacy concern (e.g., identity
tracing back concern, transaction records) can be

handled by associating weights with respect to a
user’s unique needs and circumstances. A blind
signature method provides higher untraceable
level than the pseudo identity. Alternatively, a
user may have various role identities for different
transactions, and this then involves both dynamic
identity management and seamless unlinkability
management.
Convenience data access control facilitates
the ambient e-service realization. A single
sign-on authorization is more acceptable than
those complex authorization processes. While
the identity authentication can be achieved by
various techniques (e.g., Strong authentication,
password, etc.), the proper method is based on a
users’ unique needs and preference.
Respecting the heterogeneity of data sources,
since all data sources have their own risk levels
(e.g., risk probabilities), carefree heterogeneous
data sources should draw upon the entire data
sources so as to enable the computation required
for decision making. This computation takes into
account the risk level, heterogeneity, and the
quantity of available data. However, an economic
evaluation method is indispensable due to the
computational limitation of ambient handheld
devices.
2356
A Roadmap for Ambient E-Service
%HQH¿WVRIASEM

ASEM aims to provide the design guidelines
of the platforms/infrastructures for supporting
ambient e-services with a safety and trustworthy
environment as well as congregating the collective
effort of mobile users within the environment.
,QWKLVVHFWLRQWKHEHQH¿WVRIASEM from the
VRFLRHFRQRPLFSHUVSHFWLYHDUHEULHÀ\GLVFXVVHG
LQDGGLWLRQWRWKHMXVWL¿FDWLRQRIASEM rendered
being technologically possible as addressed in
previous sections).
From the economic view for privacy invasion,
anecdotal evidence shows that people are willing
to disclose personal information for potential
monetary savings (Russell, 1989), and people do
join Web sites for free gifts and catalogs. Those
evidence supports that individuals respond to
economic incentives in deciding whether to dis-
close information. On the other hand, in various
organizational and marketing contexts, concern
of privacy invasion have been shown to depend
on information control, outcomes arising from
disclosures, information type and sensitivity, per-
ceived relevancy of information use, and target of
disclosures. Hoffman, Novak, and Peralta (1999)
claimed that nearly 63% of consumers would not
provide information to Web sites owing to lack
of trust.
From the socio-economic view, our method
LV WR EH HYDOXDWHG LQ WHUPV RI D FRVWEHQH¿W
DQDO\VLVDQGH[SHFWWKHPDMRULW\EHQH¿WZRXOG

eliminate privacy invasion. In other words,
with our method privacy of a person’s persona
would be appropriately protected because all
the real personal identities are hidden. Seamless
Unlinkability Management enables users to
control their owned information in accord with
the information type and sensitivity; users are able
to decide whether to disclosure the information
or not, as well as the target of their information
disclosed to.
Furthermore, users within the ambient envi-
ronment may provide various data sources (i.e.,
experience or subjective opinions) for others to
make a strategic decision. These collective efforts
encourage the building of the sense of ambient
trust by engaging the reliability of fraud detection
in ambient e-service environments.
However, some systematical costs are
required. Making decisions with heterogeneous
data source provide a comparative reliability
rather than depending on their own information,
especially in the dynamic environment. When the
number of peers exceeds the limit of computation
capability, the complexity of data management
and computation will become a major problem
especially in a Peer-to-Peer environment. For
preliminary estimates, establishing a trustworthy
ambient environment with privacy protection
PLJKW KDYHWR WUDGHZLWKVRPH HI¿FLHQF\ ORVV
Therefore, the trade-off between the cost and

EHQH¿W LV D PDMRU LVVXH IRU IXUWKHU UHVHDUFK
At the moment, this chapter mainly focuses
on the framework prospect, but we intend to
provide a vision of collective wisdom within
ambient e-service environment. However, the
implementation or systematical evaluations are
not the focus of this chapter.
In summary, with the supporting infrastructures
(delimited by ASEM), ambient e-service
applications are believed to bring us a new
carefree information/service era. Let’s take the
shopping mall scenario for example. Mobile
users can acquire desired information items very
conveniently. That is, even a user communicating
with other unknown mobile users, there is still
measurable data for the user to consider, such as
numerous nearby unknown users’ experience or
opinions gathered to serve as a reference material
for advanced transaction decisions. Users can
determine their actions based on various informa-
tion sources, and decide which one is suitable for
their current needs. Fraud and untruthful activities
can also be diminished with the collective effort
from those participants.
2357
A Roadmap for Ambient E-Service
CONCLUSION
Ambient e-services address dynamic collective
efforts of mobile users dynamically engaging
interactions in the ambient environments, ren-

dering a new paradigm of mobile commerce
promising revolutionary business models. This
chapter presents an ambient e-services framework
characterizing three supporting stacks. The am-
bient value stack describes the value process in
ambient environments. The ambient technology
VWDFNLGHQWL¿HVWKHWHFKQRORJ\SURFHVVWRHQVXUH
connectivity and security in ambient interactions
and cooperation between peers and then realize
powerful collective efforts. The environment
stack then represents the ambient basics for the
collaborations.
Ambient e-services applications can be divided
into two types. One is for the distributed trading;
another is for the distributed collaboration. How-
HYHUVRFLDOFRQWH[WDQGVLJQL¿FDQWUDSLGJURZWK
of connections enabled by P2P are the two major
incentives for applying ambient e-service to such
revolutionary business models. We exemplify
several ambient e-service applications. Those ap-
plications differ from existing mobile e-services
(grounded on client/server design) in terms of the
focus of the dynamic interactions between peers
in dynamic ambient e-service environments.
In this chapter, we also present another
framework called Ambient e-Service Embracing
Model (ASEM) that addresses the core elements
(of relevance to the integrated concern of trust,
reputation and privacy) required for assuring
such desired features as convenience, safety,

fairness and collaboration for mobile users when
they engage ambient e-services. This framework
manifests the relationship between the issues
of dynamic identity management and ambient
data management. The framework abstracts
the trust, reputation, and privacy concerns into
an integrated consideration. Since different e-
service applications are of different circumstances
and bear different restrictions, the framework
of ASEM also outlines the guidelines for the
implementation of ambient e-service applications
and the platforms.
The fruitful future research includes a further
in-depth evaluation of ASEM, a complete
design of the ASEM core elements, and a
¿HOGHG LPSOHPHQWDWLRQ RI FHUWDLQ ambient e-
services deriving economic models of ambient
e-services.
REFERENCES
Aberer, K., & Despotovic, Z. (2001). Managing
trust in a peer-2-peer information system. In Pro-
ceedings of the Conference on Information and
Knowledge Management ( pp. 310 -317). CIK M01.
November 5-10, 2001, Atlanta, Georgia.
$VKUD¿ 1  .XLOERHU - 3  3ULYDF\
protection via technology: Platform for privacy
preferences (P3P). International Journal of E-
Business Research, 1(2), 56-69.
Baida, Z., Gordijn, J., Akkermans, H., Saele,
H., & Morch, A. Z. (2005). Finding e-service

offerings by computer-supported customer need
reasoning, International Journal of E-Business
Research, 1(3), 91-112.
Castelfranchi, C., Falcone, R., & Pezzulo, G.
(2003). Integrating trustfulness and decision using
fuzzy cognitive maps. Trust Management 2003,
LNCS 2692 (pp. 195-210).
Enzmann, M., & Schneider, M. (2004). A privacy-
friendly loyalty system for electronic marketplac-
es. In Proceedings of the 2004 IEEE International
Conference on e-Technology, e-Commerce, and
e-Service (EEE04), Taiwan.
Enzmann, M., & Schneider, M. (2005). Improv-
ing customer retention in e-commerce through
a secure and privacy-enhanced loyalty system.
Information Systems Frontiers, 7(4-5), 359-370,
December 2005.
2358
A Roadmap for Ambient E-Service
Golbeck, J., Parsia, B., & Hendler, J. (2003). Trust
networks on the semantic Web. Proceedings of
Cooperative Intelligent Agents 2003.
Gupta, M., Judge, P., & Ammar, M. (2003). A
reputation system for peer-to-peer networks.
NOSSDAV’03. (Jun. 1-3, 2003) 144-152.
Hertzum, M., Andersen, H., Andersen, V., &
Hansen, C. (2002). Trust in information sources:
Seeking information from people, documents,
and virtual agents. Interacting with Computers,
14(5), 575-599.

Hoffman, D. L., Novak, T. P., & Peralta, M. A.
(1999). Building consumer trust online. Com-
munications of the ACM, 42(4), 80-85.
Hoy, Grubbs, M., & Phelps, J. (2003). Consumer
privacy and security protection on church Web
sites: Reasons for concern. Journal of Public
Policy and Marketing, 22(1), 58-70.
Hwang, Y. C., & Yuan, S. T. (2005). Ambient
e-service embracing model. In Proceedings
of the 7
th
IEEE International Conference on E-
Commerce Technology (pp. 535-538). (CEC2005)
Germany.
Hwang Y. C., & Yuan S. T. (2006). Technical
report: Exploring collective wisdom in ambient
e-service environment: implementation method
and evaluations. Taiwan: National Cheng-Chi
University.
Kinateder, M., & Rothermel, K. (2003). Architec-
ture and algorithms for a distributed reputation
system. Trust Management 2003, LNCS 2692
(pp. 1-16).
Lim, E., & Saiu, K. (2003). Advances in mobile
commerce technologies. Hershey, PA: Idea Group
Publishing.
Lin, K., Lu, H., & Yu, T. (2004). A distributed
trust and reputation management framework for
e-services. IEEE International Conference on
Services Computing, Shanghai, China.

Mui, L., Halberstadt, A., & Mohtashemi, M.
(2003). Evaluating reputation in multi-agents
systems. In R. Falcone, S. Barber, L. Korba, &
M. Singh (Eds), Trust, reputation, and security:
theories and practice (pp. 123-137). Berlin:
Springer-Verlag.
Odlyzko, A. (2003). Privacy, economics, and
price discrimination on the Internet. The 5
th
In-
ternational Conference on Electronic Commerce
(ICEC 2003). ACM Press.
Paternò, F. (2003). Understanding interaction
with mobile devices. Interacting with Computers,
15(4), 473-478.
Resnick, P., Zeckhauser, R., Friedman, E., & Ku-
wabara, K. (2000). Reputation systems: Facilitat-
ing trust in Internet interactions. Communications
of the ACM, 43(12), 45-48.
Roussos, G., Peterson, D., & Patel, U. (2003).
Mobile identity management: An enacted view.
International Journal of E-Commerce, 8(1),
81-100.
Russell, C. (1989). Kiss and tell American Demo-
graphics, 11(12), 2.
Sabater, J., & Sierra, C. (2002). Reputation and
social network analysis in multi-agent systems.
The 1
st
International Joint Conference on Autono-

mous Agents & Multiagent Systems (pp. 475-482).
AAMAS’02, July 15-19, 2002, Bologna, Italy.
Schilit, B., Adams, N., & Want, R. (1994). Context-
aware computing applications.In Proceedings of
the 1
st
I n t e r n a t i o n a l Wo r k s h o p o n M o b i l e C o m p u t -
ing Systems and Applications (pp. 85-90).
Shand, B., Dimmonck, N., & Bacon, J. (2003).
Trust for ubiquitous, transparent collaboration.
In Proceedings of the 1
st
IEEE International
2359
A Roadmap for Ambient E-Service
Conference on Pervasive Computing and Com-
munications (PerCom’03).
Tw i g g , A . (2 0 0 3) . A s u b j e c t i v e a p p r o a c h t o r o u t i n g
in P2P and ad hoc networks. Trust Management
2003 (pp. 225-238), LNCS 2692.
Wilhelm, U. G., Staamann, S. M., & Buttyan, L.
(2000). A pessimistic approach to trust in mobile
agent platforms. IEEE Internet Computing (pp.
40-48), September-October.
This work was previously published in the International Journal of E-Business Research, Vol. 3, Issue 1, edited by I. Lee, pp.
51-73, copyright 2007 by IGI Publishing (an imprint of IGI Global).
2360
Copyright © 2009, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited.
Chapter 8.2
A Survey on Neural Networks in

Automated Negotiations
Ioannis Papaioannou
National Technical University of Athens, Greece
Ioanna Roussaki
National Technical University of Athens, Greece
Miltiades Anagnostou
National Technical University of Athens, Greece
INTRODUCTION
Automated negotiation is a very challenging
UHVHDUFK¿HOGWKDWLVJDLQLQJPRPHQWXPLQWKHH
business domain. There are three main categories
RIDXWRPDWHGQHJRWLDWLRQVFODVVL¿HGDFFRUGLQJWR
the participating agent cardinality and the nature
of their interaction (Jennings, Faratin, Lomuscio,
Parsons, Sierra, & Wooldridge, 2001): the bilat-
eral, where each agent negotiates with a single
opponent, the multi-lateral which involves many
providers and clients in an auction-like framework
and the argumentation/persuasion-based models
where the involving parties use more sophisticated
arguments to establish an agreement. In all these
automated negotiation domains, several research
efforts have focused on predicting the behaviour
of negotiating agents. This work can be classi-
¿HGLQWZRPDLQFDWHJRULHV7KH¿UVWLVEDVHGRQ
techniques that require strong a-priori knowledge
concerning the behaviour of the opponent agent
in previous negotiation threads. The second uses
mechanisms that perform well in single-instance
negotiations, where no historical data about the

past negotiating behaviour of the opponent agent
is available. One quite popular tool that can sup-
port the latter case is Neural Networks (NNs)
(Haykin, 1999).
NNs are often used in various real world ap-
plications where the estimation or modelling of a
function or system is required. In the automated
negotiations domain, their usage aims mainly to
enhance the performance of negotiating agents in
predicting their opponents’ behaviour and thus,
achieve better overall results on their behalf.
This paper provides a survey of the most popular
automated negotiation approaches that are us-
2361
A Survey on Neural Networks in Automated Negotiations
ing NNs to estimate elements of the opponent’s
behaviour.
The rest of this paper is structure as follows.
The second section elaborates on the state of
the art bilateral negotiation frameworks that are
EDVHGRQ11V7KHWKLUGVHFWLRQEULHÀ\SUHVHQWV
the multilateral negotiation solutions that exploit
NNs. Finally, in the last section a brief discussion
on the survey is provided.
NEURAL NETWORKS IN
BILATERAL NEGOTIATIONS
In (Zhang, Ye, Makedon, & Ford, 2004) a hy-
brid bilateral negotiation strategy mechanism is
described that supplies negotiation agents with
PRUHÀH[LELOLW\DQGUREXVWQHVVLQDQDXWRPDWHG

negotiation system. The framework supports a
dynamically assignment of an appropriate ne-
gotiation strategy to an agent according to the
current environment, along with a mechanism
to create new negotiation rules by learning from
past negotiations. These learning capabilities are
based on feedforward back-propagation neural
networks and multidimensional inter-transac-
tion association rules. However, the framework
LVQRWDGHTXDWHO\GHVFULEHGDQGGH¿QHGDQGWKH
QHXUDOQHWZRUNVDUHQRWVSHFL¿FDOO\LQVWDQWLDWHG
Additionally, there are neither quantitative nor
qualitative experimental results for real world
cases. Finally, the format of the input to the ge-
neric network that is presented is ambiguously
described.
In (Zeng, Meng, & Zeng, 2005), the authors
employ a neural network to assist the negotia-
WLRQRYHUYHU\VSHFL¿FLVVXHVIURPDUHDOZRUOG
example. The network is trained online by the
past offers made by the opponent, while both
the buyer and the seller agent have the ability
to employ the proposed network. However, the
experimental data sets are very restrictive and
do not address the diversity of those that can be
arisen in real scenarios. Additionally, the authors
do not present the actual size of the hidden layer,
a parameter that is extremely crucial with regards
to the appropriateness to use such a network in a
real time negotiation procedure by an agent with

limited resources.
Furthermore, in (Rau, Tsai, Chen, & Shiang,
2006), the authors studied the negotiation pro-
cess between a shipper and a forwarded using
a learning-based approach, which employed a
feedforward back-propagation neural network
with two input data models and the negotiation
decision functions. Issues of the negotiation were
the shipping price, delay penalty, due date, and
shipping quantity. The proposed mechanism was
applicable to both parties at the same time and the
network architecture was chosen based on past
similar attempts, following a very restrictive pat-
tern for the number of the hidden layer’s neurons.
The conducted experiments showed an overall
improvement of the results for both negotiating
parties, while the framework was proven stable
and with small deadlock probability. However,
as its authors support, further experimentation is
required especially with regards to a wider variety
of strategies and possibly more suitable network
architectures for the hidden layer.
In (Carbonneau, Kersten, & Vahidov, 2006),
a neural network based model is presented for
predicting the opponent’s offers during the ne-
gotiation process. The framework was tested
RYHUDVSHFL¿FVHWRIH[SHULPHQWDOGDWDFROOHFWHG
from other existent frameworks and it is highly
adjusted to these data. The purpose of this solution
is not only to predict the opponent’s next offer,

EXWDOVRWKHSHUFHSWLRQIRUWKHVSHFL¿FSURFHGXUH
i.e. an overall vision on why everything is hap-
pening and where the procedure is led. Thus, the
prediction of the opponent’s next round offer is
only a part of the network’s output. However, the
chosen experiment set is constrained and doesn’t
examine the effectiveness of the framework on
diverse strategies as those proposed in the very
¿UVWVWHSVRIWKHDUHDDQGDUHQRZPDLQO\XVHG
(Faratin, Sierra, & Jennings, 1998). Additionally,
2362
A Survey on Neural Networks in Automated Negotiations
although the authors support the view that their
framework is proper for real-time environments,
WKHIDFWLVWKDWWKHUHVXOWHGQHWZRUNLVGLI¿FXOWWR
be online trained, mainly because of its size and
the resources that are required for such training.
Thus, this network architecture is probably inap-
propriate for mobile agents’ environments, and
VRPHWKLQJVPDOOHUDQGPRUHVSHFL¿FVKRXOGEH
designed, due to the limitations that these envi-
ronments share.
Moreover, in (Oprea, 2003), the author presents
a shopping agent, which is capable of negotiating
in online bilateral, multi-issue procedures using
DQRIÀLQHFUHDWHGDQGWUDLQHGIHHGIRUZDUGQHXUDO
QHWZRUNLQRUGHUWRLQFUHDVHLWVSUR¿WDELOLW\E\
adapting its behaviour according to its opponent’s.
The purpose of the neural network’s application on
each procedure is to predict the opponent’s next

offer on a round by round basis and thus, model
LWVEHKDYLRXUDQG LQWHQWLRQV LQRUGHU WR¿QDOO\
achieve a better or even the best possible deal.
With the exploitation of the neural network the
shopping agent can decide during the online phase
of negotiation, which is the opponent’s strategy
and estimate its reservation value. Concerning
the experiments conducted, the author uses the
ZHOOMXVWL¿HG QHJRWLDWLRQ WDFWLFV SUHVHQWHG LQ
(Faratin, Sierra, & Jennings, 1998) in order to
test the proposed solution and concludes that the
framework is working well in case of medium or
long term agents’ deadlines. However, the results
SUHVHQWHGDUHQRWWKRURXJKO\MXVWL¿HGDQGPRUH
extreme opponent strategies should be tested in
order to decide on the network’s adequacy for
such environments. Probably, the three hidden
OD\HU QHXURQV PLJKW QRW EH VXI¿FLHQW IRU VXFK
cases and long-term estimations.
Finally, Papaioannou et al. have recently
designed and evaluated several single-issue
bilateral negotiation approaches, where the Cli-
ent agent is enhanced with Neural Networks.
0RUH VSHFL¿FDOO\ LQ 5RXVVDNL 3DSDLRDQQRX
& Anagnostou, 2006), the Client agent uses a
lightweight feedforward back-propagation NN
coupled with a fair relative tit-for-tat imitative
tactic, and attempts to estimate the Provider’s
price offer upon the expiration of the Client’s
deadline. This approach increases the number of

agreements reached by one third in average. In
(Papaioannou, Roussaki, & Anagnostou, 2006),
the performance of MLP and RBF NNs towards
the prediction of the Provider’s offers at the last
round has been compared. The experiments indi-
cate that the number of agreements is increased
by ~38% in average via both the MLP- and the
RBF-assisted strategies. Nevertheless, the overall
time and the number of neurons required by the
MLP are considerably higher than these required
by the RBF. In (Roussaki, Papaioannou, & An-
agnostou, 2007), MLP and GR NNs have been
used by the Client agent in order to identify the
unsuccessful negotiation threads (UNTs) at an
early stage, thus terminating them long before
the deadlines expire. It has been observed that
the MLP NN detects more than 90% of UNTs in
average, outperforming by little the GR NN. Fi-
nally, in (Papaioannou, Roussaki, & Anagnostou,
2007), the performance of MLP and RBF NNs has
been compared with cubic splines, least-square-
based polynomial approximators, exponential
approximators and Gaussian approximators, in
order to predict the future offers of the negotiating
Provider Agent. The wide experimental evalua-
tion conducted indicates that both the MLP- and
the RBF-assisted negotiation strategies perform
almost equally well and outperform the other four
approximator-assisted strategies. In this paper, the
proposed framework is extended to address multi-

LVVXHQHJRWLDWLRQVFRQVLGHULQJWKHVLJQL¿FDQFHRI
the issues under negotiation for the negotiating
party, as well as their degree of interdependency.
A disadvantage in the aforementioned NN-based
negotiation frameworks is that they have only
been evaluated in case the Provider agent adopts
a time-dependent strategy.
2363
A Survey on Neural Networks in Automated Negotiations
NEURAL NETWORKS IN
MULTILATERAL NEGOTIATIONS
In (Oprea, 2001), the use of a small-scaled feed-
forward neural network is attempted in order to
predict the opponent agent’s behaviour. In this
framework the enhanced agent is negotiating
against an opponent that is not equipped with
any learning or other intelligent mechanism. The
neural network is properly constructed and trained
at every round to respond with the opponent’s
next value at each negotiation step using only
the three prior offers issued by the opponent.
This fact makes the step-by-step computation
feasible in real time procedures, but not neces-
sarily reliable. However, the proposed approach
was proved adequate only in cases when either
both agents (or at least the opponent agent) have
long-term deadlines.
A different usage of the neural networks’
potential is presented in (Shibata, & Ito, 1999),
where the authors are mainly concerned with

the communication between agents. In principal,
they divide the agents’ communication into two
FODVVHVZLWKUHVSHFWWRLWVPHDQLQJ7KH¿UVWRQH
incorporates the cases where the agent transmits
the observed information while the second those
where the agent’s intention is transmitted. The
framework exploits an Elman recurrent neural
network with feedback loops, especially for the
latter class of cases. The network assists the agents
to avoid possible negotiation deadlocks, although
nothing is known apriori with regards to their
strategy or resources. The network keeps the past
information and adapts online its corresponding
agent’s behaviour accordingly in order to avoid
collisions. The proposed framework was also
tested with four agents leading to promising re-
sults. However, the authors don’t propose or apply
WHFKQLTXHVIRUKLJKHUSUR¿WDELOLW\RIWKHSDUWLFL-
pating agents but only for collision avoidance by
learning the opponent’s intention. Additionally, a
recurrent neural network is a complex structure
and seems inappropriate for application in low
resources agent environments.
Furthermore, in (Abreu, Canuto, & Santana,
2005) the authors present a comparative analysis
of some negotiation methods used in a multi-neu-
ral agent system, called NeurAge. This system
LVFRPSRVHGRIVHYHUDOQHXUDOFODVVL¿HUVFDOOHG
neural agents, and its main aim is to overcome
VRPHGUDZEDFNVRIPXOWLFODVVL¿HUV\VWHPVDQG

as a consequence, to improve their performance.
These neural agents provide a common output,
which results after negotiation among them and
it is the system’s output. For this purpose, three
different negotiation methods are evaluated: the
game theoretic, the auction based and the con-
¿GHQFHEDVHGRQHV7KHUHVXOWVSURYHWKDWWKH
SURSRVHGDSSURDFKLVYDOXDEOHIRUVXFKFODVVL¿HU
systems and might end up being valuable in cases
ZKHUHWDFWLFFODVVL¿FDWLRQVKRXOGEHFRQGXFWHG
However, the system is inappropriate for online
procedures, requires cooperation between mul-
tiple neural agents and has not been tested on real
negotiation tactics’ numerical data. Therefore, the
UHVXOWV PLJKWEH YDOXDEOHZKHQD FODVVL¿FDWLRQ
scheme is required, but are probably inappropriate
as a future prediction pattern.
On the other hand, (Veit, & Czernohous, 2003)
present the results of enhancing consumer agents
with several machine-learning algorithms in a
properly designed electronic market with one static
VXSSOLHU7KHUHVXOWVSURYHWKDWXQGHUYHU\VSHFL¿F
circumstances the neural network assisted agent
performs worse than a simple Q-learning assisted
DJHQWWKDWPDLQWDLQVDVSHFL¿FVHWRIYDOXHVIRU
the learning procedure in an a-priori instantiated
matrix. However, the scenarios are very restric-
tive and in no case address the characteristics of
real world ones where the application of similar
table based agents would fail mainly due to the

diversity of the potential solution spaces for each
negotiation. Besides, the authors themselves admit
this remark, including it in their future plans.
In (Park, & Yang, 2006), the authors propose a
negotiation agents system based on the incremen-