Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation
Takuma Oishi and Koji Hasebe
Department of Computer Science, University of Tsukuba, Japan
Keywords:
Negotiation Model, Bilateral Negotiation, Competitive Agent.
Abstract:
In most existing negotiation models, each agent aims only to maximize its own utility, regardless of the utility
of the opponent. However, in reality, there are many negotiations in which the goal is to maximize the relative
difference between one’s own utility and that of the opponent, which can be regarded as a kind of zero-sum
game. The objective of this study is to present a model of competitive bilateral multi-issue negotiation and to
analyze strategies for negotiations of this type. The strategy we propose is that the agent makes predictions
both about the opponent’s preference and how the opponent is currently predicting its own preference. Based
on these predictions, the offer that the opponent is most likely to accept is proposed. To demonstrate the
usefulness of this strategy, we conducted experiments in which agents with several strategies, including ours,
negotiated with one another. The results demonstrated that our proposed strategy had the highest average
utility and winning rate regardless of the error rate of the preference prediction.
1 INTRODUCTION
Automated negotiation is a process in which an au-
tonomous agent interacts with another agent (or hu-
man) to form an agreement that is desirable for both
parties. Several studies have been conducted on au-
tomated negotiations over a period of decades, inter-
acting with fields such as artificial intelligence and e-
commerce. (A comprehensive survey of research in
this area is provided by (Baarslag et al., 2016) and
(Kiruthika et al., 2020).)
Various negotiation models have been proposed
regarding the number of agents and incomplete infor-
mation. However, in most of these, each agent’s ob-
jective is only to maximize its own utility, regardless
of the utility of its opponent. Therefore, the objective
of all agents in a negotiation is to achieve a Pareto-
optimal outcome. However, in the real world, many
negotiations exist in which agents are required not
only to increase their own utilities, but also to max-
imize the difference between their own utilities and
those of their opponents. Typical examples of this in-
clude various tradings such as foreign exchange trans-
actions, and negotiations between parties in compet-
itive relationships, as modeled in some board games
such as CATAN (CATAN GmbH, nd). The negotia-
tion of this type requires more skillful tactics that have
not been considered in existing negotiation models.
The objective of this study is to provide a model of
bilateral multi-issue competitive negotiation, and to
propose a strategy for negotiations of this type. Com-
petitive negotiation is a special type of negotiation in
which the relative difference between the utilities ob-
tained by an agent and its opponent is evaluated as the
actual utility. In this sense, this problem is a kind of
zero-sum game. Another difference between conven-
tional negotiations and competitive ones is that, in the
former, the utility obtained as a result of an agreement
is known in advance, whereas in competitive negoti-
ation, the actual utility is not known in advance be-
cause it depends in part on the opponent’s utility.
Generally, in negotiations, agents make decisions
such as the choice of the contents of offers and
whether to agree to the opponent’s offer, while mak-
ing predictions regarding the opponent’s preferences.
In this study, we focused on decision-making strate-
gies rather than the prediction technique. Specifically,
our proposed strategy comprises two tactics. The first
tactic applies to setting a target value (i.e., the min-
imum acceptable relative utility) for an agent, based
on the prediction of the opponent’s preference as well
as the prediction of how the opponent will predict the
agent’s own preference. The second tactic is used
when choosing the content of the offer to which the
opponent is most likely to agree.
We evaluated the effectiveness of our strategy by
conducting bilateral negotiations on three competitive
issues. The experimental results showed that the pro-
404
Oishi, T. and Hasebe, K.
Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation.
DOI: 10.5220/0011800800003393
In Proceedings of the 15th International Conference on Agents and Artificial Intelligence (ICAART 2023) - Volume 1, pages 404-411
ISBN: 978-989-758-623-1; ISSN: 2184-433X
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
posed strategy had the highest average utility and win-
ning rate regardless of the error rate of the preference
prediction.
The remainder of this paper is organized as fol-
lows. Section 2 presents related work. Section 3 de-
scribes the model of competitive negotiation. Sec-
tion 4 introduces strategies for competitive negotia-
tion. Section 5 presents experimental results. Finally,
Section 6 concludes the paper and presents plans for
future work.
2 RELATED WORK
A representative early study on the automated ne-
gotiation was conducted by Faratin et al. (Faratin
et al., 1998). This study disseminated a technique
for searching for an agreement point by concessions
and described the basic idea of strategic negotiation.
Fatima et al. (Fatima et al., 2006) analyzed negoti-
ation strategies and equilibrium in bilateral negotia-
tions, where in which there is a negotiation deadline
and the opponent’s preferences are unknown. Jen-
nings et al. (Jennings et al., 2001) classified the exist-
ing negotiation strategies by approaches, and studies
(Cao et al., 2015; Zheng et al., 2014) proposed negoti-
ation approaches that combine multiple strategies and
dynamically change an agent’s behavior according to
that of the opponent.
Our study and those described above focus on bi-
lateral negotiations, but there have also been stud-
ies on negotiations involving more than two agents.
For example, (Aydo
˘
gan et al., 2014) investigated the
strategies and protocols in such negotiations. There
are also studies, such as (Mansour and Kowalczyk,
2015; An et al., 2011) regarding situations in which
single-issue negotiations between two agents to ob-
tain multiple resources are performed in parallel in
e-commerce.
The negotiation problem targeted by the above
studies aims at maximizing each agent’s own utility
through negotiation. Conversely, our study focuses on
the competitive negotiation problem that aims to max-
imize the difference between one’s own utility and the
opponent’s utility, which has not been thoroughly in-
vestigated.
There is a study (Keizer et al., 2017) targeting ne-
gotiation in CATAN, which is a specific example of
competitive negotiations. Although that study did not
present a negotiation model, it implemented negotia-
tion strategies by a rule-based technique and machine
learning.
Figure 1: Utility spaces of non-competitive (left) and com-
petitive (right) bilateral negotiation.
3 COMPETITIVE BILATERAL
NEGOTIATION MODEL
In this section, we define competitive bilateral multi-
issue negotiation model. In addition, we show that
this negotiation with perfect information always leads
to an agreement in which the utility for both agents
are zero.
3.1 Basic Concept
As explained earlier, the objective of a rational agent
in conventional bilateral negotiations is to reach an
agreement that maximizes its own utility regardless
of that of the opponent. Conversely, the objective of
the agent in the competitive bilateral negotiation pro-
posed in this study is to maximize the difference in
utility between itself and the opponent. In this sense,
this negotiation can be regarded as a zero-sum game.
Therefore, even if high utility is obtained as a result
of negotiation, it is undesirable to bring high utility
to the opponent as well. Rather, it is desirable that the
difference in utility relative to the other party is larger,
even if one’s own utility is lower.
Figure 1 illustrates the difference between non-
competitive (left) and competitive (right) negotia-
tions. Here, the utility spaces of agents i
1
and i
2
in
each negotiation are shown, and the green and blue
arrows represent the vector of utility aimed at by i
1
and i
2
, respectively. In non-competitive negotiations,
the combined vector of both agents is indicated by a
red arrow. This means that the goal of the agents is to
reach agreement on the Pareto frontier represented by
the curved line. On the contrary, in competitive ne-
gotiations, each agent aims both to maximize its own
utility and minimize that of the opponent. Therefore,
the combined vectors of each agent are opposite to
each other as indicated by the red arrows. This im-
plies that each agent aims to reach an agreement on
its own side of the diagonal line that indicates a rela-
tive utility of zero for both agents.
Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation
405
3.2 Formal Model
The formal model of negotiation is defined as a tu-
ple of the following components: a set of negotiation
agents, negotiation domain (often called the outcome
space), negotiation protocol, and preference profiles.
(The paper (Baarslag et al., 2016) presents a survey
that provides more details.)
Let I be the set of negotiation agents. Because
this study deals only with bilateral negotiations, we
fix I = {i
1
,i
2
}. We also use the notation i to indicate
the opponent of agent i I. The negotiation domain,
or often called the outcome space (denoted by O) rep-
resents the set of all possible negotiation outcomes.
The negotiation protocol determines the rules of ne-
gotiation, such as the order of offers and the condi-
tions of agreement. The preference profile is a binary
relation over the negotiation domain for each agent,
that determines which of any two outcomes is more
preferable. Here, we follow the game-theoretic con-
vention and define the relation by a utility function
U
i
: O R (for i I). However, in competitive nego-
tiations, the goal is to maximize the difference in util-
ities between the two agents. Therefore, in addition to
the usual utility function, we introduce RU
i
: O R,
which represents the relative utility. The formal defi-
nitions of these components are given below.
3.2.1 Negotiation Domain
The negotiation domain O is represented as a product
of one or more sets (called issues) of possible out-
comes. The set of indices for the issues is represented
by J = {1,..., j}, where j is the number of issues.
The set of outcomes for each issue is represented by
O
k
(k J), and thus O is defined to be O
1
× ·· · × O
j
.
3.2.2 Negotiation Protocol
As with many previous studies on bilateral negotia-
tions, we follow the alternating-offers protocol (Ru-
binstein, 1982), in which agents take turns making
suggestions while searching for a mutually acceptable
outcome. More specifically, an agent with turn pro-
poses one of the elements of the negotiation domain
as the content of offer. The other agent who received
the proposal selects one of the following three actions.
Accept: Agree to the proposal. Both agents obtain
utility based on the agreed outcome.
Offer (Counter-offer): Reject the proposal and
make a new offer to the other party.
EndNegotiation: Reject the proposal and end the ne-
gotiation. Both agents receive a utility of zero.
Times (steps) in the progress of negotiation are rep-
resented by discrete values t = 1,2,... T . Negotia-
tions may have a time limit (denoted by t
max
) and, if
no agreement is reached within it, both agents gain a
utility value of zero.
3.2.3 Preference Profile
The utility of agent i I for outcome o O is defined
by the following equation:
U
i
(o) = Σ
j
k=1
w
i
k
V
i
k
(o
k
),
where w
i
k
denotes the weight assigned to agent i for
issue k, satisfying Σ
j
k=1
w
i
k
= 1. V
i
k
: O
k
R is the
evaluation function for issue k for agent i. Accord-
ing to convention, in this study, the range of utility is
defined as [0,1] R.
3.2.4 Relative Utility Function
A relative utility function RU
i
: O R is introduced
to represent the difference between the utility of the
two parties in a competitive negotiation. This is de-
fined as
RU
i
(o) = U
i
(o) U
i
(o).
3.3 Analysis of Perfect Information
Case
In competitive negotiations, if the agents’ preference
profiles are common knowledge, the utility for both
agents will be always zero. We demonstrate this fact
in game theory.
Here, we assume that the negotiation has a time
limit t
max
and that the utility function of the agents is
common knowledge. First, let i be the agent having
a turn at time t
max
and consider the optimal action of
i at this time. At t
max
, i can choose only Accept or
EndNegotiation, and if offer o made by i at t
max
1 satisfies RU
i
(o) > RU
i
(o), i can obtain a positive
relative utility by choosing Accept. Otherwise (i.e.,
RU
i
(o) RU
i
(o)), and i obtains relative utility zero
by choosing EndNegotiation. Thus, at t
max
1, i
should offer o
0
with U
i
(o
0
) U
i
(o
0
), which results
in a final relative utility of zero for both agents. Using
the same argument in the reverse direction, we find
that at any time t < t
max
, both agents make offers with
positive relative utility, leading to the same result.
4 NEGOTIATION STRATEGIES
This section presents some possible strategies for
competitive negotiations, the effectiveness of which
is analyzed empirically in the next section.
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
406
4.1 Tactics
In general, negotiation comprises two decisions:
choosing an outcome from the negotiation domain as
one’s own offer and deciding whether to accept the
opponent’s offer. To decide what to offer, a plausible
approach is to set a target value (i.e., the minimum
acceptable relative utility) in advance, and make an
offer that exceeds this target. Furthermore, there are
two possible ways to select an outcome to offer: one
is to select an outcome randomly that exceeds the tar-
get value, and the other is to select an outcome that the
opponent is more likely to accept. To decide whether
to accept an offer, a plausible approach is to set a tar-
get value in advance, and agree to the opponent’s offer
if it exceeds the target value. Here, these two target
values may generally differ and may vary with time.
Below, we discuss possible tactics for each of
these decisions.
4.1.1 Tactics to Decide Target Value for Offer
Algorithm 1: Evaluation of T RU with prediction-dependent
tactic.
1: Update U
i
est(i)
and U
i
est(i)
2: for all o O do
3: if RU
i
(o) = T RU
i
t
then
4: if U
i
est(i)
(o) U
i
est(i)
(o) > U
i
est(i)
(o
t2
)
U
i
est(i)
(o
t2
) then
5: O
i
t
.add(o)
6: end if
7: end if
8: end for
9: if O
i
t
6= ø then
10: T RU
i
t
= T RU
i
t
11: else
12: T RU
i
t
= T RU
i
t
c
13: end if
There are two main approaches to determine the target
relative utility for an agent’s offer. The first fixes the
value through negotiation (called fixed-value tactic)
and the second allows it to vary over time. The lat-
ter can be classified further into two types: the time-
dependent tactic, in which the target value is deter-
mined depending only on time, and the prediction-
dependent tactic, in which the target value is deter-
mined depending both on time and on predictions
about preferences.
The definitions of these three tactics are given be-
low, where T RU
i
is used to denote the target value of
agent i.
Fixed-Value Tactic: This tactic always sets the target
value to any value greater than zero.
Time-Dependent Tactic: The target value of agent i
in this tactic is defined as follows:
T RU
i
t
= min
i
+ (1 α
i
(t)) × (max
i
min
i
).
This tactic is based on the idea of (Faratin et al.,
1998). Here, min
i
(called the reservation value) and
max
i
denote the minimum and maximum relative util-
ities that agent i desires in negotiation regardless of
time, respectively.
The function α
i
: T R is called a time-
dependent function, and satisfies both α
i
(0) = κ
i
and
α
i
(t
i
max
) = 1 (where κ
i
is a constant). Various time-
dependent tactics can be defined according to the def-
inition of this function. Some well-known ways of
defining this function include polynomials and expo-
nential functions. In this study, the following polyno-
mial function is used:
α
i
(t) = κ
i
+ (1 κ
i
) × (min(t,t
i
max
)/t
i
max
)
1/β
.
This function changes the speed of concession de-
pending on the value of β. When β = 1, yielding pro-
gresses at a constant rate (linear); when β > 1, yield-
ing progresses early; and when β < 1, few conces-
sions are made while time t is small, whereas conces-
sions are suddenly made near t
i
max
.
Prediction-Dependent Tactic: This is introduced in
this study for competitive negotiation. This tactic
makes the following types of predictions:
Prediction of the opponent’s preference.
Prediction about how the opponent will predict
one’s own preference.
Based on these predictions, the target value is set tak-
ing into consideration whether the offer will be ac-
cepted by the opponent and the current time.
The detailed algorithm to determine the target
value is as follows. Let U
i
est(i)
be the preference func-
tion of i predicted by agent i, and let U
i
est(i)
be the
prediction of U
i
est(i)
by agent i.
The algorithm for the target value T RU
i
t
at time
t for agent i is described in Algorithm 1. Here, o
t2
denotes the previous offer by agent i, O
i
t
denotes the
set of candidates of offers(i.e., the set of possible of-
fers expected to yield a relative utility that matches
the target value TRU for agent i at time t), and c de-
notes the amount of change in the target value when
agent i concedes.
As shown in Algorithm 1 below, this tactic does
not change the target value if there is an offer that
the opponent is more likely to agree with that has the
same target value as the previous offer. Otherwise, the
target value is decreased.
Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation
407
Table 1: Strategies for competitive bilateral negotiation.
Strategy name Target value for offer Offer choice Target value for agreement
PMT Prediction-dependent Mislead Time-dependent
TRT Time-dependent Random offer Time-dependent
PMF Prediction-dependent Mislead Fixed
TRF Time-dependent Random offer Fixed
FMF Fixed Mislead Fixed
FRF Fixed Random offer Fixed
4.1.2 Tactics for Offer Choice
There are two main approaches to choose an offer.
The first, called the random offer tactic, chooses an
offer randomly from the set of offers that might bring
a relative utility above the target value. The second
approach, called mislead tactics, is proposed in this
study for competitive negotiation. This uses the pre-
diction of the opponent’s preference as well as the
prediction of how the opponent will predict one’s own
preference. Based on these predictions, the offer to be
selected is the one that the opponent is most likely to
accept from among the possible outcomes exceeding
the current target value. The formal definition of this
tactic is as follows: Let O
i
t
be the set of candidates for
the target value of agent i at time t. Offer o
t
at time t
by mislead tactics is then expressed as follows:
o
t
= argmax
oO
i
t
(U
i
est(i)
(o) U
i
est(i)
(o)).
4.1.3 Tactics for Determining Agreement
Tactics on agreement can be divided into two cate-
gories, depending on whether the target value is fixed
or variable. In both cases, agreement is chosen when
the predetermined target value is exceeded. Similar to
the target value used to determine the offer, tactics can
be time-dependent or prediction-dependent. Only the
simple time-dependent tactic is analyzed in this study.
4.2 Negotiation Strategies
Negotiation strategies are realized by combining the
tactics described above. Therefore there are 12 strate-
gies, consisting of three tactics to determine the target
value for the offer, two to choose the content of the
offer, and two to determine the target value as the cri-
terion for agreement. Of particular interest are the six
strategies shown in Table 1.
5 EXPERIMENTS
In order to evaluate the usefulness of the proposed
negotiation strategy, we conducted experiments in
which the strategies defined in Section 4 were negoti-
ated with each other.
5.1 Parameter Settings
5.1.1 Negotiation Domain
The negotiation domain used in the experiments con-
sisted of three issues, each with ten options. The ne-
gotiation domain therefore consisted of a set of 1,000
outcomes.
5.1.2 Negotiation Strategies
For the experiments, we developed negotiation agents
for each of the six strategies defined in the previous
section, whose parameter settings were presented be-
low.
For the time-dependent tactic, the values of the
four parameters are as follows.
min = 0.05.
max = 1.0.
κ = 0.1.
β = 1.0.
For the prediction-dependent tactic, the parameter
c for concession is set to 0.05. However, if this tactic
makes a concession resulting in T RU 0, to prevent
the agent from making an offer that would hurt itself,
the value of T RU was updated to 0.05 and chose an
offer satisfying RU > 0. In the experiment, the offers
based on the random-offer tactic and mislead tactic
were selected within the range of 0.02 round the value
of T RU . This is because the agent may not find an
offer that yields utility exactly equal to T RU . Thus,
the agent set the value of T RU at time t to RU (o
t2
)
(where o
t2
denotes the previous offer), and decided
whether to make concession by Algorithm 1.
5.1.3 Preferences
The preferences defined by the weights and utility
functions for i
1
and i
2
were set as follows, respec-
tively. Here, the same utility functions were used for
all issues.
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
408
Figure 2: Average utility for each strategy when the cosine
similarities of predictions were 0.99, 0.95, and 0.9.
Figure 3: Win rate (denoted by red) and draw rate (denoted
by orange) for each strategy when the cosine similarities of
predictions were 0.99, 0.95, and 0.9.
w
i
1
= (0.2,0.3,0.5).
V
i
1
= (0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0).
w
i
2
= (0.5,0.3,0.2).
V
i
2
= (1.0,0.9,0.8,0.7,0.6,0.5,0.4,0.3,0.2,0.1).
Each agent i was not informed about the exact
preferences of its opponent. Instead, it was initially
informed of its own and its opponent’s preferences
with a certain range of error, which mimicked the pre-
dicted value U
i
est(i)
and U
i
est(i)
. Specifically, for each
of the weights of the agents’ preferences and the eval-
uation functions, we gave predictions with errors such
that the cosine similarity is 0.99,0.95 and 0.9. Here,
since the difference between the preferences of it’s
own and the opponent can be large or small depend-
ing on the predicted values, in the experiment all of
the cases were evaluated.
5.2 Experimental Results
Under the settings described in the previous section,
the results of round-robin matches with six negotia-
tion strategies are shown below.
Figure 2 shows the average utility obtained by
each strategy. Here, the horizontal axis represents
the results of each strategy, and the three graphs show
the results when the cosine similarities of the prefer-
ence predictions are 0.99, 0.95, and 0.90, respectively,
from the left. The vertical axis shows average utilities
with their maximum and minimum values.
Figure 3 shows the sum of the win and draw rates
for each strategy. Here, the horizontal axis represents
the results of each strategy, and the three graphs show
the results when the preference predictions have co-
sine similarities of 0.99, 0.95, and 0.90, respectively,
from the left. The vertical red and orange graphs rep-
resent win and draw rates, respectively.
As shown in Figure 2, PMT and TRT strategies
were found to be superior to the other strategies, with
a positive average utility regardless of the error rate
of prediction. Specifically, when the cosine similar-
ity of the preference predictions was 0.99, the aver-
age utilities of PMT and TRT strategies were about
0.043 and 0.049, respectively. Also, when the cosine
similarity was 0.9, the average utilities of these two
strategies were about 0.168 and 0.187, respectively.
On the other hand, the average utilities of the other
four strategies always had negative average utility.
From the above results, it can be seen that the
utility of both PMT and TRT strategies increases as
the difference in expectations increases. A possible
reason for this is that both the PMT and TRT strate-
gies have time-dependent target values for agreement.
The larger the prediction error, the greater the proba-
bility of making a wrong decision about one’s rela-
tive gain for a proposal, and the greater the probabil-
ity of a larger error in the value of that relative gain.
While other strategies choose to agree when their rel-
ative gains are large (relative utilities greater than 0),
the PMT and TRT strategies have stricter criteria for
agreement, so they are less likely to agree to an agree-
ment that will actually be to their detriment. As a re-
sult, the larger the forecast error, the higher the rela-
tive utility of the PMT and TRT strategies, suggesting
that having a time-dependent agreement target value
is important for competitive negotiation.
Figure 3 shows that PMT strategy had a lower win
rate than TRT strategy, but the sum of the draw rate
and the win rate was always the highest for the PMT
strategy. For example, when the cosine similarity of
preference predictions was 0.99, the win rates of PMT
and TRT strategies were about 0.452 and 0.482, re-
spectively. The draw rates for these strategies were
about 0.365 and 0.293 respectively. Thus, the sum
of the win and draw rates for PMT and TRT were
about 0.816 and 0.775, respectively, with 4.2 percent-
age point higher win or draw rate for the PMT strat-
egy.
Examples of the negotiation process for PMT and
Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation
409
Figure 4: The negotiation process in which PMT won over
TRT by the largest utility margin.
Figure 5: The negotiation process in which PMT lost to
TRT by the largest utility margin.
TRT strategies are shown in Figures 4 and 5. Here,
the orange and blue plots indicate the history of the
offers by PMT and TRT agents, respectively, and the
red star indicates the agreed outcome. Figures 4 and
5 both show the results when the Cosine similarity of
the preference prediction was 0.99.
Figure 4 shows the negotiation process when PMT
wins TRT by the largest margin of utility. In Figure
4, neither agent agreed with the other’s offer and con-
tinued to make offers that were profitable for them,
but in the end, TRT made a mistake and made an of-
fer that was more profitable for PMT. As a result, the
utility of PMT became positive.
Figure 5 shows the negotiation process when PMT
lost by the largest margin of utility to TRT. In Figure
5, both agents made offers that were advantageous to
them until the end. However, in the end, PMT mispre-
dicted the utility of the offer received from TRT and
chose to agree, resulting in a negative utility for PMT.
In competitive negotiations, winning is important,
but not losing is also important. PMT strategy is also
an excellent strategy in terms of stability. Based on
the above results, the PMT strategy is considered to
be the optimal negotiation strategy in competitive ne-
gotiations.
6 CONCLUSIONS AND FUTURE
WORK
In this study, we proposed a model of competitive
bilateral multi-issue negotiation, in which an agent’s
utility and that of the opponent are evaluated relative
to each other and the actual utility can be regarded
as a zero-sum game. We also proposed a strategy for
the negotiations of this type, in which the basic idea
is to choose the offer that the opponent is most likely
to accept, based on the prediction of the opponent’s
preference and the prediction of one’s own preference
from the opponent’s perspective.
To demonstrate the effectiveness of the proposed
strategy, we conducted experiments in which agents
with various strategies negotiate competitive three-
issue bilateral negotiations. The results show that
the proposed strategy achieves the highest utility and
winning rate, regardless of the prediction error rate.
We also showed that time-dependent target value for
agreement is important for gaining relative utility in
competitive negotiations.
In future work, we will develop a prediction
method required in our negotiation setting, based on
some existing methods such as those using Bayesian
estimation (Lin et al., 2006) or heuristics (Jonker and
Robu, 2004). We are also interested in applying our
strategy to the development of agents that play board
games involving competitive negotiation.
REFERENCES
An, B., Lesser, V., and Sim, K. M. (2011). Strategic agents
for multi-resource negotiation. Autonomous Agents
and Multi-Agent Systems, 23(1):114–153.
Aydo
˘
gan, R., Hindriks, K. V., and Jonker, C. M. (2014).
Multilateral mediated negotiation protocols with feed-
back. In Novel insights in agent-based complex auto-
mated negotiation, pages 43–59. Springer.
Baarslag, T., Hendrikx, M. J., Hindriks, K. V., and Jonker,
C. M. (2016). Learning about the opponent in au-
tomated bilateral negotiation: a comprehensive sur-
vey of opponent modeling techniques. Autonomous
Agents and Multi-Agent Systems, 30(5):849–898.
Cao, M., Luo, X., Luo, X. R., and Dai, X. (2015). Auto-
mated negotiation for e-commerce decision making: a
goal deliberated agent architecture for multi-strategy
selection. Decision Support Systems, 73:1–14.
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
410
CATAN GmbH (n.d.). Welcome to the World of CATAN.
Retrieved January 6, 2023 from https://www.catan.
com/.
Faratin, P., Sierra, C., and Jennings, N. R. (1998). Ne-
gotiation decision functions for autonomous agents.
Robotics and Autonomous Systems, 24(3-4):159–182.
Fatima, S. S., Wooldridge, M. J., and Jennings, N. R.
(2006). Multi-issue negotiation with deadlines. Jour-
nal of Artificial Intelligence Research, 27:381–417.
Jennings, N. R., Faratin, P., Lomuscio, A. R., Parsons, S.,
Sierra, C., and Wooldridge, M. (2001). Automated
negotiation: prospects, methods and challenges. Inter-
national Journal of Group Decision and Negotiation,
10(2):199–215.
Jonker, C. and Robu, V. (2004). Automated multi-attribute
negotiation with efficient use of incomplete preference
information. Available at SSRN 744047.
Keizer, S., Guhe, M., Cuay
´
ahuitl, H., Efstathiou, I., En-
gelbrecht, K.-P., Dobre, M., Lascarides, A., Lemon,
O., et al. (2017). Evaluating persuasion strategies and
deep reinforcement learning methods for negotiation
dialogue agents. ACL.
Kiruthika, U., Somasundaram, T. S., and Raja, S. (2020).
Lifecycle model of a negotiation agent: A survey of
automated negotiation techniques. Group Decision
and Negotiation, 29(6):1239–1262.
Lin, R., Kraus, S., Wilkenfeld, J., and Barry, J. (2006).
An automated agent for bilateral negotiation with
bounded rational agents with incomplete information.
Frontiers in Artificial Intelligence and Applications,
pages 270–274.
Mansour, K. and Kowalczyk, R. (2015). An approach to
one-to-many concurrent negotiation. Group Decision
and Negotiation, 24(1):45–66.
Rubinstein, A. (1982). Perfect equilibrium in a bargaining
model. Econometrica: Journal of the Econometric
Society, 50(1):97–109.
Zheng, X., Martin, P., Brohman, K., and Da Xu, L. (2014).
Cloud service negotiation in internet of things envi-
ronment: A mixed approach. IEEE Transactions on
Industrial Informatics, 10(2):1506–1515.
Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation
411