International Journal of Engineering and Technical Research (IJETR)

ISSN: 2321-0869 (O) 2454-4698 (P), Volume-7, Issue-1, January 2017

Repeated stochastic bargaining game to share

queuing network resources
T. B. Ravaliminoarimalalason, H. Z. Andriamanohisoa, F. Randimbindrainibe

When the function ui (ci ) is concave, the optimal solution

Abstract— Resources in a queuing network are very rare exists because the bargaining set is a convex set.
and expensive. Sharing them to the customers of the network is Players bid the resource by proposing payments
a hard problem, and it impacts a lot the performances of the
system. We present in this paper a new optimal and feasible ( p1 ,L , pn ) where pi is the payment proposed by player i.
method to share these resources based on the result of the In [2], we can see that allocation proportional to player’s
game theory, especially the repeated stochastic bargaining proposal is an optimal solution to such game. In spite of this,
game. Our model, which we call “myopic model”, is tougher a problem is found when a player has proposed a very high
even in an instability phenomenon. payment. To overcome this inconvenient, we propose a gain
which is function of the utility and the payment proposal to
Index Terms—Bargaining game, Queue, Repeated game,
have more equilibrium between players.
Stochastic game.
We assume that the system ask an elementary price µ to
each player. Given the payments ( p1 ,L , pn ) proposed by
I. INTRODUCTION the players, and the allocated resources (c1 ,L , cn ) , we must
A processor sharing queue is a particular queue where the have for all i :
word queue is a misnomer. All existing customers ahead the pi
servers are immediately served with a part of its available ci = (2)
resource. These customers are dynamic. Each of them leaves µ
the queue to move to another after receiving the service he If the queue server allocated all available resource, it is easy
has requested from the server, whence the concept of to find that the elementary price is given by (3).
queuing network comes.
The contribution we provide in this paper is how to share the µ= ∑p i
(3)
server resources between those customers. Our method is C
based on the game theory, especially a repeated stochastic By sharing all available resource, and assuming that players
bargaining game. In this way, we propose a model of are identical (the server asks the same price µ to each of
myopic player which optimizes only his current gain. them), a single elementary price µ exists for each proposed
payment ( p1 ,L , pn ) [3].
II. GAME THEORY ELEMENTS
In our model, given a price µ , each player has to
A. A bargaining game maximize his gain function g i on the possible proposal
Let’s briefly review our model using a bargaining game. pi set.
It is a non-cooperative game, among strategic game, which
opposed n players sharing a resource C. Assuming that each ⎛p ⎞
gi ( pi , µ ) = ui ⎜ i ⎟ − pi (4)
player i receive a utility ui (ci ) for an allocated resource ci . ⎝µ⎠
For all i, we suppose that the utility function ui (ci ) is This gain function is defined from the utility compared to
concave, strictly increasing and continuous for ci > 0 and resource allocated to the player in return for the payment he
proposed. We will use this function to limit abuses in term
that the derivative uiʹ (ci ) is finite. of higher price proposal for some players. At a higher price
A solution to that game is the maximization of the social will decrease this gain.
utility on the players set [1]. The equilibrium can be defined as the price proposal
n n ( p1* ,L , pn* ) where players will maximize their gains, and the
Argmax ∑ ui (ci ) such as ∑ ci ≤ C, ci ≥ 0, i = 1,L (1)
i =1 i =1 system will grant the elementary price µ defined in (5).
*

gi ( pi* , µ ) ≥ g i ( pi , µ ), pi ≥ 0, i = 1,L , n, µ =
∑p i
(5)
C
T. B. Ravaliminoarimalalason, PhD School in Sciences and Technical
Engineering and Innovation, University of Antananarivo, Antananarivo, In that case, the solution ci* = pi* µ is an optimal solution
Madagascar, + 261340016433.
H. Z. Andriamanohisoa, PhD School in Sciences and Technical satisfying (1).
Engineering and Innovation, University of Antananarivo, Antananarivo, To prove it, we can use the lagrangian method to establish
Madagascar, + 261341033943. that the conditions in (5) are identical to the conditions in (1)
F. Randimbindrainibe, PhD School in Sciences and Technical
Engineering and Innovation, University of Antananarivo, Antananarivo, with ci* = pi* µ .
Madagascar, + 261340646690.

49 www.erpublication.org
 Repeated stochastic bargaining game to share queuing network resources

B. A repeated game At each step of the game, given the current state e ∈ E ,
With repeated game we can model situations where players choose strategies s = (s1 ,L , sN ) to play. Each player
players interact repetitively playing the same game [4]-[6]. i own a gain gi (e, s) , and then, the system goes from state e
They choose their actions simultaneously without knowing
to state eʹ according to the transition model T who satisfies
the choices of other players. So, a basic game is played in
(7).
each period of a discrete time t = 1, 2, ... , and at the end of
each period, the players observe the performed actions. ∑ T (e, s, eʹ) = 1
eʹ∈E
(7)
We are given a game in normal form
Card Si
(
J = N , ( Si )i∈• , ( gi )i∈• ) where N is the players set, Si the We call a policy π i : E → [0,1] the vector whose

set of possible strategies for player i, and g i his associated elements define a probability distribution over the strategies
of player i, specific to a game in normal form defined by the
gain. state e. for player i, the policy defines a local strategy in
At the step t = 1 , each player i chooses an action si1 ∈ Si each state within the meaning of game theory. The expected
independently of other players. Let’s denote s1 = ( si1 )i∈• the utility refers to the expected gain on the strategies of
opposing players. For a joint policy π = (π1 ,L , π N ) , we
vector joint played actions in step 1. At the end of this step,
define the expected utility of player i for each state e as
s1 is revealed to all players. expressed in (8).
At step t ( t ≥ 2 ), knowing the history
ui (π , e) = E s∈S [ gi (e, s)] (8)
ht = ( s1 , s 2 ,L , s t −1 ) of the played actions during the past,
each player i chooses an action sit ∈ Si independently of where S = S1 ×L × SN , and E denotes the expectation
other players. Let’s denote H t the set of possible history at function.
step t. And therefore, we can also define the utility U i (π , e) of
It remains to define the gain function. For this, we should states for player i, associated to a joint policy π , as the
determine how the players evaluate the result of an infinite expected utility for player i from the state e if all players
length of history. Indeed, if ( s1 , s 2 ,L ) ∈ H ∞ , the player i follow this joint policy.
receive a gain g i ( s t ) at step t. In a model of discounted ∞
U i (π , e) = E s∈S ⎡⎢ ∑ δ t −1ui (π , e) | e0 = e ⎤⎥
game, the player gives more weight to a unit of gain ⎣ t =1 ⎦ (9)
received today compared to a unit of gain received = ui (π , e) + δ ∑ ∑ T (e, s, eʹ).π (e, s).U i (π , eʹ)
tomorrow. For this, we will use a discount factor δ ∈ ]0,1[ . s∈S eʹ∈E

Thus, one unit of gain owned at step 2 is only δ unit of gain where π (e, s) designates the probability of the joint
in step 1, and one unit of gain owned at step t is δ t −1 unit of strategies s on the state e according to the joint policy π ,
gain in step 1. and δ is the discount factor.
In this context, the gain owned by player i during a game In a stochastic game, a Nash equilibrium is a vector
play h = ( s1 , s 2 ,L ) and evaluated at time t = 1 is expressed strategy π * = (π 1* ,L , π N * ) as for all state e ∈ E and for all
by (6) [6]. player i [9] :
U i ( (π i* , π −*i ), e ) ≥ U i ( (π i , π −*i ), e ) , ∀π i ∈ Π i
∞
giδ (h) = (1 − δ )∑ δ t −1 gi (st ) (6) (10)
t =1
where Π i is the set of policies offered to player i. The
In (6), the factor (1 − δ ) is a normalization factor to take the
gain back to the same unit at any step. notation (π i* , π −*i ) means the vector of policies π * where

C. A stochastic game
π i* is the policy of player i and π −*i the joint policy (π *j ) of
j ≠i

Stochastic games are extension of Markov process in case players other than i.
of several agents called players in a common environment
[7]-[9]. These players play a joint action which defines the III. MODEL OF CUSTOMERS AND QUEUING
owned gains and the new state of the environment. NETWORK
A stochastic game can be defined by a quintuplet
A. Principle
( N , E, ( S ) , ( g )
i i∈• i i∈• )
, T where:
On a given queue, customers arrive and others leave. Each
- N is the set of players who act to the game customer needs a total of resource bi to fulfill his
- E is the finite set of states of the game
requirement that he asked to the server. In the following, we
- Si is the set of possible strategies (actions) for player i put that these resources are sampled and shareable in order
- g i is the gain function, which is a function of the state to work in the discrete domain. At the beginning of each
of the game and the strategies played by all players : time interval t , each player i sends his strategy sit for the
gi : E × S1 ×L × S N → R bargaining of the resource C of the server. We restrict to a
- T is the transition model between states, which finite countable strategies set S i . The server will evaluate
depends to joint strategies :
these proposals to compute the resources cit that he will
T : E × S1 ×L × SN × E → [0,1]
assign to the players. Each resource cit has a price pit that

50 www.erpublication.org
 International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869 (O) 2454-4698 (P), Volume-7, Issue-1, January 2017
the server sends with the resource to the player. Once Let git : Bi × S1 ×L × S → R be a function gain for the
Nt
received, each resource will be used by each player and they
will calculate their gains. They also assess the remains of player, function of the local state bit and the joint strategies
their respective requirement function of the consumed s t = (s1t , !, sNt t ). We evaluate this gain from the allocated
resources. We precise that the player requirement at the
beginning of time t was assessed at the end of time resources, which are themselves based on price proposals
t − 1 after using the resource cit −1 allocated at this time. s t done by all players.
The transition model between local states is defined by
As we consider the mobility of the players, each of them
will decide from his requirement whether he will stay in his the function Ti t : Bi × S1 × ! × S N t × Bi → [0,1] as expressed
current queue, or move to the next queue. Initially, when the by (13) :
customer i arrives in a queue, the first requirement is noted t t t +1
bi0 . Over time, depending on the allocated resources, that ∑ T (b , s , b
i i i ) =1 (13)
bit +1∈Bi
t
requirement becomes b as : i
Since the requirements (local states) bit +1 and bit are
t t −1 t −1
b =b −c
i i i (11) dependent, and are also function of the joint strategy s t , we
The decision of the customer is determined by (12). can say that this function can be well defined. We will
⎧ stay if b > 0 t further evaluate this transition function. From these data, it
i
dec(bit ) = ⎨ t
(12) is possible to model the actions and movements of the
⎩move if bi ≤ 0 players with a stochastic game defined by the quintuplet
This principle is illustrated on Fig. 1. ( N t , ( Bi ), ( Si ), ( g i ), (Ti )).
C. Bargaining of the resource of the server
Given a resource C of the server, it will be bargained
through the customers of the queue, here called as players.
At time t , each player must maximize his gain on the
possible proposals set Pi as shown on (14) obtained from (4)

⎛ pt ⎞
g it ( pit , µ t ) = ui ⎜⎜ it ⎟⎟ − pit (14)
⎝µ ⎠
On (14), the function u i (cit ) is the utility function of
player i regarding the resource cit that the server has
allocated after the bargaining computation. This function
must be a concave function as described on paragraph II.
And to better assess the allocated resource, it is necessary
that this utility function u i is also function of the
requirement bit . We can use, for example, a logarithmic or
quadratic valuation given in (15) and (16).
log(cit )
ui (cit , bit ) = A (15)
log(bit )
2
Fig. 1. Games at time t ⎛ ct ⎞
ui (c , b ) = A − B⎜⎜ it
t
i i
t
⎟
⎟ (16)
⎝ bi ⎠
B. The game formulation
where A and B are arbitrary positive constants. The proof of
Let’s model these actions and movements by a stochastic
the concavity of these functions comes from the negativity
game. This is a stochastic game between N t players. The of their second derivatives.
number of players N t varies over time as players arrive to or Players send to server their proposals
depart from the queue according to their requirements. t
pt = ( p1t ,L , pNt t ) ∈ R N . Once received by the server, it
At time t , for player i , the local state of the game is
defined by the requirement bit . By its finite cardinal, let’s computes the resource to allocate c t = (c1t ,!, c t t ).
N

note Bi the set of possible requirements of player i . For On (14), the price µ t is not yet known beforehand, so the
player i , let’s put sit the strategy that he proposes to the players are not able to compute the optimal proposal pit .
server. The set of possible proposals, noted S i = { sit }, is also We suppose that price anticipation described by (3) is used.
The function gain that he must maximize is defined by (17)
the set of possible strategies for that player. This strategy is
developed on the next paragraph. assuming that ∑ j ≠i p tj is constant.

51 www.erpublication.org
 Repeated stochastic bargaining game to share queuing network resources

2) Impact of the game history and its future

⎛ pit ⎞
git ( pit ) = ui ⎜ t
C ⎟ − pit (17) As we are faced to a repeated game, players can use the
⎜ ∑ j pj ⎟ history of the game to bargain the resource of the server. Let
⎝ ⎠
ht = {b1 , d 1 , µ 1 ,L , bt −1 , d t −1 , µ t −1 , bt } ∈ H t where bt indicates
By cancelling the derivative with respect to pit :
the requirements at time t, d t the proposals sent to the
⎛ p t⎞⎛ p t ⎞ server at this time, µ t the elementary price enforced by the
( g )ʹ = u ʹ ⎜⎜
t
i i
C ⎟ . ⎜1 −
i
t t
i
⎟ .C − 1 = 0 (18)
⎝ ∑ j p ⎠ ⎝ ∑ j pj
⎟ ⎜ ⎟
j ⎠ server, and H t the set of possible history for this game. The
game history that the player i can observe is called
pit
Let’s note d it = C . So, we have : observation of the player i that we denote oit . This
∑ p tj
j observation is limited because the player i is not able to
observe some part of the history of his requirements, his
⎛ dt ⎞ 1
uiʹ ( dit ) . ⎜1 − i ⎟ = (19) proposals, and the elementary price sent by the server due to
⎝ C⎠ C lack of memory. So, we have oit ⊂ ht . It is also required that
Let’s call d it the strategy of player i, that he sends to player can observe his current state bit : bit ∈ oit . Let’s denote
server at time t. This strategy will be assessed from (19) Oit the set of possible observations up the time t for player i.
independently of the other players. So, once the joint Considering these observations, the player i can adjust his
strategy st = (d1t ,L , d t t ) is received by the server, it way of calculating his proposal at time t. Let’s call it as
N

pit policy of player i at time t, denoted π it . It differs from a

computes the bargaining. As d it = , the price pit to
µt simple strategy d it by using the observations up the time t.
perform this computation can be derived (by the server) π it : Oit → Ai
from the strategy d it sent by the player i with some factor (23)
oit a ai = π it (oit )
µ t (the elementary price of resource) that he enforces.
In (23), Ai indicates the set of possible proposals.
D. Study of the stochastic game of the player i Equation (23) mentions that policies π it are function of
1) Model of the state transition observation of player i at time t. So, the player i have to
determine a policy which can ensure a best response. The
As we suggested that the player i plays a stochastic game better is to find a time independent policy that we call a
defined by a quintuplet ( N t , ( Bi ), ( S i ), ( g i ), (Ti )). To model stationary policy π i , so as to be usable at any time by
the dynamism of player i at time t, we defined his local state simply basing to observations. We can get this policy
as his requirement bit . The state change process is very clear because the current requirement bit depends only on the
after using the resource cit allocated by the server. The previous requirement bit −1 and the allocated resource at this
t
transition from the state b to another state b means thati i
t +1
time that is function of the proposal d it −1 as given by (21). In
t +1 t
the allocated resource at time t is equal to b − b : i i that case, the policy is markovian.
Let denote π = (π i , π −i ) the joint stationary policy. The
bit +1 = bit − cit (20)
t
gain gik (bik , dik ) owned by the player i at step k is discounted
The probability of transition from the state b to another i
by a factor δ k −t at time t, and the total gain owned by this
state bit +1 can be assessed as the probability that the allocated player at time t with the joint stationary policy π is denoted
resource at time t is cit = bit − bit +1 . The model as Git (bit , π ) . This gain is expressed by the recurrent relation
Ti : Bi × S1 × !× S N t × Bi → [0,1] of the state transition is
t
(24).
expressed by (21). Git (bit , π ) = git (bit , dit ) + δ ∑ T (bit +1 | bit ).Git +1 (bit +1 , π ) (24)
bit +1∈Bi
⎧ t t +1 dit
⎪1 if bi
− bi
= C The policy π i * which ensures the best response for player i is
Ti (bit , s t , bit +1 ) = ⎨ ∑ j d tj (21)
⎪ 0 otherwise given by (25).
⎩
π i* (π −i ) = Arg max Git (bit , (π ï , π −ï )) (25)
where the joint strategy of all player is st = (d1t ,L , d t t ) , πi
N

d it The problem is how the player i can find this optimal policy
and C means the allocated resource to player i at time
∑ j d tj π i* .
t. Equation (25) shows that the policy of player i ensuring
The gain owned by that player at this time is expressed by the best response depends on other players policies, that is a
(22). function of other players states. Player i doesn’t know other
players states, so he doesn’t able to compute his optimal
git (dit ) = ui (dit ) − µ t dit (22) policy. However, he can optimize only his immediate gain
git (bit , d it ) . The expected gain is therefore discounted by a

52 www.erpublication.org
 International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869 (O) 2454-4698 (P), Volume-7, Issue-1, January 2017

factor δ = 0 . In that case, the proposal strategy d it coincides

with the optimal policy π i * . Let’s call myopic policy this
policy which doesn’t consider, or ignore, the impact of the
future. The optimal policy is denoted π i m , which is a time
independent function of a single variable, and depends only
on the current state of player i.
3) Performances of the model
We can use two different measures to evaluate the
performance of our model: the total gain (with discount) that
each player must maximize, and the sojourn time that each
player must minimize. The total gain owned at time t is
given by (26).
Git (bit , π m ) = git (bit , dit ) (26)
Fig. 2. Simulation description
t
where the proposal d and the function π (b ) coincide.
i
m t
i
“Source” sends same packets to “Dest MYOP”, “Dest
If the player i arrives to the queue at time tii , the gain KSBS”, “Dest FIFO” and “Dest PS” through the queues.
owned at a time t ( t ≥ tii ) is discounted by a factor The links between these entities feel no packet propagation
i delay or propagation error.
δ ti −t during the calculation of the total gain. The player i will Simulations are based on the following parameters:
move from this queue at time ti f as (27). - The inter-arrival of packets T (time between
successive generations of packet at “Source”) has an
ti f = min {bit ≤ 0, t > tii } (27) exponential distribution parameter λ = 1 second
t
- The µ packet sizes generated by “Source” as an
The sojourn time for player i is expected by (28).
exponential distribution with parameter 1024 bits.
tiS = ti f − tii (28) - The processing capacity C of the server of each queue
is fixed. This capacity, expressed in bits per second
The total gain for player i is expected by (29).
(bps), is identical for all four queues of the system.
f
ti

GiT = ∑ δ ti −t Git (bit , π m ) A. Stable system

t =tii Let’s consider a stable system, where the capacity of
ti
f
server is greater than the load rate of the queue. We used
= ∑ δ ti −t g it (bit , π m ) (29) C = 1100 bps for the simulation.
t =tii

ti
f During 10 minutes of simulation, we get the results below.
= ∑ δ ti −t ( ui (π m ) − µ t π m )
t =tii

We can also use the expected total gain to penalize the

player in term of time, by making him able to own more
gain if he doesn’t stay longer on the queue. The expected
total gain is expressed by (40).
GiT
GiT =
ti f − tii
f
ti (30)
∑ δ (u (π ti −t
i
m
) − µ tπ m )
t =tii
=
ti f − tii

IV. SIMULATION EVALUATION AND ANALYSIS

Fig. 3. Evolution of the average number of packets on each
To evaluate our model, we tried to implement our model queue on a stable system
on queuing networks who convey packets simulated on
OPNET Modeler software [10]. We compared the We find on Fig. 3 similar properties of the PS queue and
established model to other models to know his performance. our MYOPIC queue. FIFO queue and the egalitarian
As described in Fig. 2, simulations consist of: solution have more packets queuing compared to PS and
- FIFO (First in First Out) queue, MYOPIC.
- Classic PS (Processor Sharing) queue,
- Processor sharing queue using the KSBS
(Kalai-Smorodinsky Bargaining Solution) [11],
- Our myopic players model.

53 www.erpublication.org
 Repeated stochastic bargaining game to share queuing network resources

Fig. 4. Evolution of the average sojourn time on each queue on a Fig. 6. Evolution of the average number of packets on each queue
stable system on an unstable system
As in Fig. 4, the average sojourn time is almost identical
for PS queue and MYOPIC queue. There is a stable
difference around 0.7 second between them, where the
sojourn time for PS queue is greater than the one for
MYOPIC queue.

Fig. 7. Evolution of the average sojourn time on each queue on an

unstable system

Fig. 5. Evolution of the throughput from each queue on a stable

system
In Fig. 5, the system provides identical throughput. The
number of packets per second which come out of each
queue is almost the same after a long time of simulation.
B. Unstable system
Now, let’s consider an unstable system, where the capacity
of the server is lower than the load rate of the queue. For
that, we used C = 900 bps.
In Fig. 6, at the 10th minute, we already find that the
system is unstable; the number of packets on each queue is
increasing but the lowest is shown by the MYOPIC queue. Fig. 8. Evolution of the average throughput from each queue on an
Contrary to a stable system, the difference between PS unstable system
queue and MYOPIC queue performance is highlighted, the
two curves diverge. Table I shows the number of packets received by each
Till the 10th minute, we can read on Fig. 6 that in average: destination after the 10 minutes of simulation. It puts
- 33.45 packets are found processed on the MYOPIC queue, evidence the difference of the average throughput as in Fig.
- 35.12 packets are found processed on the PS queue, 8.
- 36.71 packets are found queuing on the FIFO queue, Table. I. Number of packets at each destination at the end
- 41.64 packets are found queuing on the KSBS queue.
In term of average sojourn time, sojourn on a MYOPIC Node Name [Total]
Dest FIFO 495
queue is lowest compared to the other scheduling method.
Dest KSBS 488
We can interpret it as a low latency in practice.
Dest MYOPIC 501
Dest PS 496

54 www.erpublication.org
 International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869 (O) 2454-4698 (P), Volume-7, Issue-1, January 2017
Source 607
Let’s precise that the Source sends the same number of
packets at the same time with a same distribution, but this
big difference is due to the scheduling and processing on
each queue.
We can say that our MYOPIC system can better manage the
packets in case of instability compared to the classic PS
queue (e.g: in case of temporary congestion).

V. CONCLUSION
Our contribution consists of a new way to manage the
resource of queue. Our methodology is based on a repeated
stochastic bargaining game to share the resources of a
queuing network. We introduced a new principle of a
myopic player who doesn’t optimize his future gain by the
history of the game. The simulation shows the performance
of our model which has a better scheduling during an
instability period.

REFERENCES
[1] J.F. Nash, “The Bargaining Problem”, Econometrica, vol. 18, 1950,
pp. 155-162
[2] T.B. Ravaliminoarimalalason, H.Z. Andriamanohisoa, F.
Randimbindrainibe, “Modélisation de partage de ressources dans un
réseau de files d’attente à processeur partagé par la théorie des jeux de
marchandage. Résolution algorithmique.”, MADA-ETI, vol. 2, 2014,
pp 28-39,
[3] R. Johari, “The price of anarchy and the design of scalable resource
allocation mechanisms”, Algorithmic game theory, Cambridge
University Press, 2007, pp. 543-568
[4] G.J. Mailath, L. Samuelson, Repeated Games and Reputations :
Long-run relationships, Oxford University Press, 2006
[5] J.-F. Mertens, S. Sorin, S. Zamir, “Repeated games”, Core D.P. -
Université Catholique de Louvain, 1994, pp. 20-22
[6] D. Abreu, “Theory of Infinitely Repeated Games with Discounting”,
Econometrica, vol. 56, 1988, pp. 383-396
[7] L.S. Shapley, “Stochastic games”, Proceedings of the National
Academy of Sciences of the U. S. A., vol. 39, 1953, pp. 1095-1100
[8] A. Neyman, S. Sorin, “Stochastic games and applications”, NATO
Science Series, Kluwer Academic Publishers, 2003
[9] A.M. Fink, “Equilibrium in a stochastic n-person game”, Journal of
Sciences of Hiroshima University, vol.28, 1964, pp. 89-93
[10] K.A. Chew, Introduction to OPNET Modeler, Centre for
Communication Systems Research, University of Surrey - Guildford -
UK, 2002
[11] E. Kalai, M. Smorodinsky, “Other Solutions to Nash’s Bargaining
Problem”, Econometrica, vol. 43, 1975, pp. 513-518

T. B. Ravaliminoarimalalason, PhD School in Sciences and Technical

Engineering and Innovation, University of Antananarivo, Antananarivo,
Madagascar, + 261340016433.

H. Z. Andriamanohisoa, PhD School in Sciences and Technical

Engineering and Innovation, University of Antananarivo, Antananarivo,
Madagascar, + 261341033943.

F. Randimbindrainibe, PhD School in Sciences and Technical

Engineering and Innovation, University of Antananarivo, Antananarivo,
Madagascar, + 261340646690.

55 www.erpublication.org