Int. J. Communications, Network and System Sciences, 2009, 2, 764-774
doi:10.4236/ijcns.2009.28089 blished Online November 2009 (http://www.SciRP.org/journal/ijcns/).
Pu
A Reputation-Based Multi-Agent Model for Network
Resource Selection
Junfeng TIAN, Juan LI, Lidan YANG
Network Technology Institute, Hebei University, Baoding, China
Received July 7, 2009; revised August 12, 2009; accepted September 27, 2009
Abstract
Because of the anonymity and openness of on-line transactions and the richness of network resources, the
problems of the credibility of the on-line trading and the exact selection of network resources have become
acute. For this reason, a reputation-based multi-agent model for network resource selection (RMNRS) is
presented. The model divides the network into numbers of trust domains. Each domain has one domain-agent
and several entity-agents. The model prevents the inconsistency of information that is maintained by differ-
ent agents through the periodically communication between the agents. The model enables the consumers to
receive responses from agents significantly quicker than that of traditional models, because the global repu-
tation values of service providers and consumers are evaluated and updated dynamically after each transac-
tion. And the model allocates two global reputation values to each entity and takes the recognition value that
how much the service provider knows the service into account. In order to make users choose the best
matching services and give users with trusted services, the model also takes the similarity between services
into account and uses the similarity degree to amend the integration reputation value with harmonic-mean.
Finally, the effectiveness and feasibility of this model is illustrated by the experiment.
Keywords: Trust, Reputation, Trust-Domain, Multi-Agent, Similarity
1. Introduction
The World Wide Web has evolved at an extreme rate due
to its capacity to provide an endless amount of resources
to the public users. Hence, the user finds him lost in a
pool of information, without knowing how to select re-
sources and which resources are credible [1]. A wide
range of mechanisms such as contracts and commercial
laws as well as face to face meetings help reduce the
likelihood of risks to the consumers in the traditional
businesses. However, When doing online trading, users
often have little or no prior knowledge of their potential
business partner(s) and the absence of these mechanisms
and face-to-face encounter make them can not check
goods before paying and can not distinguish cheat from
honest effectively. As a result, there has always produced
a trust deficit in e-commerce [2].
Thus, the trust has become an integral part of tradi-
tional transaction and e-commerce transaction. Recom-
mender system which is based on trust have proven to be
an important method to effectively find those resources
that users are interested in from endless resources in the
network, by providing users with more proactive and
personalized information services. And the recommender
system based on trust collaborative filtering is to rec-
ommend trusted and satisfying information services to
users [3,4].
The exiting recommender systems based on trust fil-
tered recommendation information just through authen-
ticating users’ identity or removing users with low trust
value [5]. They didn’t consider the following four prob-
lems: 1) the role of transaction behavior for users in
e-commerce has two types: buyer and seller, so we can’t
only use one trust value or reputation value to measure
the users’ trust level with different transaction behavior.
Because the malicious users may use honest buy behav-
ior to cover up the dishonest sell behavior, or vice versa
[6]; 2) if recommendations are from different recom-
menders, we should treat them differently not only be-
cause of the recommenders’ various trust levels but also
because the recommenders have different knowledge to
their recommendation [7]; 3) in order to make users
quickly find resources, the time from making a request to
receiving the response should short as much as possible

J. F. TIAN ET AL. 765
[8]; 4) after each transaction, both the participators not
only should update their trust values or reputation values
but also should share their trust information of transac-
tions which can increase the spread of trust information
and raise the performance of network.
2. Related Works
In Peer-to-Peer (P2P) e-commerce transaction, users ex-
change information or transact with others through direct
communication, but those users don’t know each other
before and they also don’t trust each other. And the
openness of P2P system makes the users can’t avoid oth-
ers’ malicious behaviors. Once users transact with one
malicious node, they may incur substantial losses [9]. In
order to solve the problem of security for the network
service, M. Blaze proposed the concept of Trust Man-
agement firstly in 1996 [10], and its basic thought was to
admit the imperfection of security information in open
system. It proposed that making safety decision for sys-
tem needed additional security information. Nowadays
the researches on trust are mainly classified into two
types, identity trust and behavior trust. The identity trust
which is based on code, authentication protocol or digital
signature technology checks entity’s authenticity and
makes the decision whether authorize the entity to access.
But the behavior trust pays more attention to the trusted
problem in broader meaning. According to the past be-
havior experiences, it updates the trust relationship be-
tween users dynamically and timely. International re-
search indicates [11] that the network security is devel-
oping toward the direction of credible network. The fu-
ture network security is the credible network with in-
creasing credible behavior, which is a new consensus
that is agreed by the network security research areas in
recent years. The research on whether the users’ behavior
is credible not only increases the security of network
through decreasing or avoiding transaction with mali-
cious users but also improves the success rate of transac-
tions and decreases the extra spending caused by monitor
or precaution which are caused by distrust. Thereby the
overall performance of the network is improved.
There have been lots of researches about behavior
trust at home and abroad. Based on the transitivity of
trust, the model named EigenREP was based on global
reputation in P2P environment [12]. But its drawback is
its astringency, high communication costs and relative
global reputation, thus it can not evaluate whether the
node is credible just through the value of global reputa-
tion. The model proposed in paper [8] presents that each
grid domain is associated with multiple brokers and each
broker with multiple entities. It eases the network traffic
at the broker sites and makes the service providers’ (SPs)
response to the consumers’ request significantly quicker,
but it might lead to the information inconsistency main-
tained by different brokers and solving the problem has
its costs. A trust model [13] based on behaviors was
proposed to achieve the resource sharing and cooperation
among different domains in grid environment, which
dynamically reflects the entity’s subject characteristic.
But its limitations are that it doesn’t update the trust
value, so it cannot reflect the dynamics of trust comput-
ing. Traditional resource selection method [14–16] al-
ways selects service providers with the highest trust
value. They don’t consider whether the selected services
are the services that consumer expected, that is to say
whether the selected services and the expected services
are accordant. Paper [17] proposed a similarity meas-
urement about ontology-based semantic web services
and paper [18] proposed a method of similarity search
for web services. Both of them measure the similarity
between the services with ontology and find the expected
services to consumers. But the weakness is the high
complexity of algorithm. Paper [3–5] proposed the idea
of trust filtering in recommender systems, which consid-
ered that the recommenders should have similar tastes
and preferences, should be trustworthy in the sense that
they had a history of making reliable recommendations
and should have different trust degree in entity trust and
content trust. The weakness is that they don’t solve the
sparseness of similarity well when they find similar users.
Paper [7] proposed a role-based recommendation and
trust evaluation model which firstly takes the role of re-
commender into account. But it didn’t present how to
organize and storage a rational role hierarchy. A novel
distributed trust model is propose in [6], which itera-
tively calculates for each node a global seller reputation
value and a global buyer reputation value based on
transaction history, and whether a node is credible or not
can be identified from them. But it doesn’t provide the
computation of some coefficients.
To solve the problems of existing distributed trust mo-
del, the paper proposes a novel reputation-based multi-
agent model for network resource selection (RMNRS).
With nodes’ identity and their recognition to services, the
model computes for each node a global buyer reputation
and a global seller reputation. And the model estimates
whether one node is credible or not through the final
Trust-Value which is the harmonic-mean of Trust-Value
and similarity degree between request services and pro-
vided services. The model’s characters are listed bellow.
1) Our model is based on trust domain which adopts
multi-level trust management mechanism to manage
agents belonging to different levels. The periodically
communication between agents prevents the informa-
tion’s inconsistency between the agents.
2) Our model computes the similarity between request
service and provided service by ontology. We want to
find the most similar services to the requestor. Combined
with the computing of Trust-Value, the services that we
C opyright © 2009 SciRes. IJCNS

J. F. TIAN ET AL. 766
supplied will be not only trusty but also matching to the
request.
3) Our model takes the recognition value that how
much the service provider knows the service into account,
which makes the provided services more similar to the
request.
4) Our model doesn’t use one trust value to determine
whether a node is credible or not. It keeps global buyer
reputation value and global seller reputation value for
each node, thus it can reflect node’s different trust level
with different transaction behavior.
5) After each transaction, the participators can share
mutual trust information under certain condition, namely
trust propagation.
6) Our model provides the computation of coefficient
which is the weight when integrating two global reputa-
tion values.
This paper is organized as follows. Section 3 presents
the related definitions, algorithms and fundamental prin-
ciple of the model. Section 4 presents the simulation to
validate our model’s effectiveness and feasibility. And
finally Section 5 concludes our work.
3. The Related Definitions, Algorithms and
Fundamental Principle of the RMNRS
Model
3.1. Related Definitions
Definition 1 (Trust): Trust is the subjective probability
expectation of trustor to trustee’s specific behavior which
is relied on experiences and continuous to modify its
value as the change of trustee’s behavior. Paper [7] pro-
posed that trust is a complex subject relating to an en-
tity’s belief in honesty, trustfulness, competence and
reliability of another entity. Paper [19] proposed that
trust is to believe others, which is established on their
own knowledge and experiences and is a subjective be-
havior between entities. Trust is different from the belief
that person believes in object things, which is a subjec-
tive judgment. The trust itself is not a fact or proof, it is
acknowledge of observed fact. According to the different
achieving trust way when entities interact with each
other, Beth [20] divided the trust into direct trust and
recommendation trust. To trust an entity directly means
to believe in its capabilities with respect to the given
trust class. Recommendation trust expresses the belief in
the capability of an entity to decide whether another en-
tity is reliable in the given trust class and in its honesty
when recommending third entities.
Definition 2 (Trust-Domain): According to the Web-
Based activities and related application, we divide the
virtue network into numbers of self-government domains
and define the self-government domain as Trust-Domain.
Definition 3 (Domain-Entity): Domain-Entity is the
node or object who has some resources in network. It can
be a user, service or resource. The interaction between
entities has two types: the interaction of intra-domain
and inter-domain.
Definition 4 (Transaction): One transaction is one in-
teractive behavior which happens between two nodes
when they need mutual services in the P2P network, such
as one business dealing in e-commerce, one file
download and so on. The buyer is the one who requests
the services and the seller is the one who provides the
requested services.
Definition 5 (Reputation): Reputation is the expecta-
tion of one entity’s future behavior based on the observa-
tion of the entity’s past behavior or evaluation informa-
tion in transactions [9]. There are two types, local repu-
tation and global reputation. The local reputation is de-
fined as the expectation of a node’s future behavior
based on the past evaluation information which is pro-
vided by one of its buyers. The global reputation is de-
fined as the expectation of a node’s future behavior based
on the past evaluation information which is provided by
those nodes who had transacted with node j ago.
In this paper, each node has four types of reputation
value: local buyer reputation, local seller reputation,
global buyer reputation and global seller reputation.
Definition 6 (Trust-Value): The Trust-Value of Trust-
Domain is defined as the mean of the Trust-Value of all
entity-agents. The Trust-Value of entity-agent is the
mean of the global reputation value of all the entities
managed by the entity-agent. The Trust-Value of one
entity is the weighted mean of global buyer reputation
value and global seller reputation value of the entity.
Definition 7 (Identity): The Identity is not the role of
entities in transaction. It refers to the identity symbol of
one entity’s recognition degree to one service in certain
service domain, such as the social positions, social titles
or certificates.
3.2. Trust Domain and Agent
In social network, everyone or group has his or her own
interesting and joins in trusted communication circle,
they have high credibility on the people or group in the
same circle [21]. While in virtue network, because of the
disparate resources, the sharing, cooperation and high
performance of resources has become difficult. The wide
connectivity of network requires to establish public and
effective security mechanism between different nodes
and peoples and to implement consistent security strat-
egy. It also requires doing specific security management
according to the application of multi-network. In this
aper, we import the agent mechanism to abstract the p

J. F. TIAN ET AL. 767
Figure 1. Trust-Domain and agent management architecture.
network, as shown in Figure 1 which implies the archi-
tecture of the model. We use this mechanism to manage
the entities’ trust computing and trust relationship be-
tween entities. Our model is not only manageable but
also permits the managed object to cooperate independ-
ently, which is in accordance with the network comput-
ing mode and development tend [23].
We assume that there is an absolute trusted agent:
Root-Agent (RA) who takes charge of every agent in
each domain. Every domain has a Domain-Agent (DA)
and numbers of Entity-Agents (EA)
1) The Root-Agent is the manager of global trust rela-
tionship of the system. It is an absolute trusted root. It
manages and collects all the Domain-Agents’ informa-
tion and maintains the global trust relationship. The in-
formation that the RA maintains is the ID of DA, the
Trust-Value of domain and semantic base.
2) The Domain-Agent is the manager of part trust re-
lationship of the system and the trust relationship of do-
main. If there are entities request to join in the domain,
the DA has the right that permits them to join in or
judges which EA they belong to and it passes the infor-
mation of entities to RA, so that RA can update the se-
mantic base timely. When one EA’s Trust-Value is less
than threshold, the DA has the authority to retake the
authority and awards the authority to other entities. Ac-
cording to the Trust-Value of the entity who wants to be
the EA, DA makes the decision that whether it can be or
not. The entity has the status to be an EA only when the
Trust-Value of one entity is over the threshold. All the
entities that an EA maintains provide the same or similar
services. The DA collects EAs’ information periodically
(the period is decided by the size of system or the request
of interaction) that it maintains and broadcasts the in-
formation as guide Trust-Value to each EA. The infor-
mation that the DA maintains is the ID of EA and the
direct transaction table (the domain-ID of entity, the en-
tity-agent-ID of the entity, global buyer reputation value,
global seller reputation value, service type, time, cost and
the satisfaction degree of transaction).
3) The entities that maintained by Entity-Agent is or-
ganized as Binary Sorting Tree which has such charac-
ters: if the left sub-tree is not null then all the Trust-
Value of entities in left sub-tree will be smaller than that
of root; if the right sub-tree is not null then all the
Trust-Value of entities in right sub-tree will be bigger
than that of root; both the sub-tree are Binary Sorting
Trees. We can add, delete, update and lookup needed
information from the Binary Sorting Tree. The informa-
tion that maintained by EA is illustrated as follows.
a) The storage structure of entities in Binary Sorting
Tree.
Typedef struct BiTree {
DataType degree; // the Trust-Value of entity;
struct BiTree * lchild,* rchild; // the pointer of the
left sub-tree and right sub-tree;
} BiTnode,*Bitree;
The binary tree is sorted by the trust degree of entities.
The EA checks the Trust-Value of entities periodically
and updates its position in the tree. If the Trust-Value of
one entity is smaller than threshold, the EA can remove it
from the tree. When one entity wants to join in, the EA
puts it in the feat position.
b) The storage structure of entities in Binary Sorting
Tree of entity’s identity.
Typedef struct RoTree {
DataType Re-degree; // the recognition degree of
entity’s identity;
struct BiTree * lchild,* rchild;
}RoTnode,* Rotree;
The Binary Sorting Tree of entity’s identity is estab-
lished based on the recognition level of one entity to the
services that it provides. The tree nodes are ordered by
the recognition degree. The operation that we can do to
the tree is similar to that of a). The reason that why the
EA maintains the role tree is that it helps to select suit-
able entity.
c) Direct transaction table: the domain-ID of transac-
tion entity, the entity-agent-ID of the entity, global buyer
reputation value, global seller reputation value, service
type, time, cost and the satisfaction degree of transaction.
The communication between entities is through
Trusted Communication Agent Interface (TCAI) [22],
which makes the communication between entities more

J. F. TIAN ET AL. 768
reliable. The agents pass information on to each other
periodically. Once one entity’s information was updated,
it can push [23] its updated information to the related
agents, which can avoid the information disaccording to
others.
Our model is based on domain and has three levels, so
we divide the trust relationship between entities into two
types, the trust relationship of intra-domain and inter-
domain.
1) The trust relationship of inter-domain: The evaluation
objects of trust are based on domain. The trusted level of a
domain is evaluated through the behavior that the entities
showed in the transaction. Thus the trust level of the do-
main reflects all the entities’ trust level in the domain.
2) The trust relationship of intra-domain: The trust of
intra-domain is mainly the interaction between entities;
the evaluation objects of trust are the entities in the do-
main. After the initialization, the trust level will be up-
dated according to the behavior of entities in the follow-
ing transaction.
3.3. The Fundamental Principles of RMNRS
This section mainly shows the basic principles of this
model and the specific algorithms are listed in Subsec-
tion 3.4-3.9.
Step 1: The consumer sends the request which con-
tains the service type and the mini-trust threshold to the
agents.
Step 2: The agents select those nodes with higher
global reputation value and calculate their integrated
Trust-Value.
Step 3: The agents firstly calculate the similarity de-
gree between buyer and Candidates, and then the agents
calculate the harmonic-mean of similarity degree and
integrated Trust-Value.
Step 4: If there are entities whose final Trust-Value
(harmonic-mean) is bigger than the mini-trust threshold,
go to Step 5, conversely go to Step 11.
Step 5: If the number of candidates is over one, go to
Step 6, on the other hand, go to Step 7.
Step 6: Under the same credible condition, we select
the nodes with the highest global buyer reputation.
Step 7: The buyer transacts with the selected entity
and both parties give each other the satisfaction degree to
this transaction.
Step 8: Both their entity-agents and domain-agents
will update their reputation value and transaction table
after transaction.
Step 9: According to the satisfaction degree, if both
parties want to share their trust information, go to Step
10, on the other hand, go to Step 13.
Step 10: Both parties share their trust information of
transaction.
Step 11: The agent ask the buyer whether it wants to
modify the mini-trust threshold, so that it can find the
suitable entity to transact with. If the user wants to de-
crease the threshold, go to Step 1, and on the other hand
go to Step 12.
Step 12: The transaction is failure.
Step 13: The transaction is success.
3.4. The Initialization of Reputation
When one entity didn’t have any interactions with other
entities ago, should we trust it? In this paper we show
several methods. According to specific environment you
can select suitable method.
1) You can set the initial global buyer reputation value
and global seller reputation value to be 1[0,1]

 and
2[0,1]


separately.
2) According to the security information (such as the
information of identity [7]) that the user provides, you
can convert the information into the initial reputation
value of entity through the function, where R is the set of
role information and D is the domain that R belongs to.
After the initialization, the global reputation value will
be updated according to the behavior of entities in the
following transaction. To simplify the experiment, we set
both the initial global buyer and seller reputation value of
entity to be 0.5.
3.5. The Computation of Local Reputation Value
According to the definition of the local reputation, the
value of local reputation is calculated in the light of the
historical transaction evaluation that the buyer sent to the
seller. After each transaction, both the participators feed
back the degree of satisfaction to each other. And the
user also should give out a threshold, so if the degree of
satisfaction is higher than the threshold, we think that
this transaction is satisfactory, on the other hand, we
think it is unsatisfactory. We use the percentage of the
number of satisfactory transactions in all transactions to
represent the local reputation. And we let Lbij represent
the local buyer reputation, which mean that the local
reputation given by node i as the buyer to the node j as
the seller and let Lsij represent the local seller reputation,
which mean that the local reputation given by node i as
the seller to node j as the buyer. The formula is illus-
trated as follows.
(, )
(, )(, )
bg
ij
bgbb punish
Nij
Lb NijNij N  (1)
(, )
(, )(,)
sg
ij
s
gsb punish
Nij
Ls NijNij N  . (2)

J. F. TIAN ET AL. 769
where Nbg(i, j), Nbb(i, j), represent the number of satis-
factory transaction and the number of unsatisfactory
transaction between buyer i and seller j. Nsg(i,j), Nsb(i,j)
represent the number of satisfactory transaction and the
number of unsatisfactory transaction between seller i and
buyer j. Npunish is the coefficient which is the punishment
to malicious transaction. Especially when the cost of the
transaction is very high, the punishment coefficient
should be bigger. So the formula is
(, )(, )
11
((1)(, ))(, )
bb bb
Nij Nij
punish mk mk
kk
NCijC 

 

ij,
where Cmk(i,j) is the cost of kth unsatisfactory transaction
between node i and node j. We assume all the cost of
transaction is bigger than 1. On the contrary, we set it to
be 1; The formula not only reflects the higher the cost is,
the bigger the weight is, but also satisfies the monotonic
increasing with the increasing transaction cost and its
result is in the cope of [1,1+α], where α is the regulatory
factor. So that the user can adjust the scope of punish-
ment according to his needs.
On the basis of the analysis above, the properties of
the local reputation formula are shown as follows.
1) The model evaluates whether the node is credible or
not directly. The more the number of satisfactory trans-
action is, the closer to 1.0 the reputation value is. The
more the number of unsatisfactory transaction is, the
closer to 0.0 the reputation value is.
2) The introduction of the punishment coefficient Npu-
bish≥1 makes the reputation value decrease quicker than
rise. It embodies the punishment to the malicious nodes
and especially to those nodes who make use of high cost
transactions to cheat.
3.6. The Computation of Global Reputation
Value
The global reputation value is the integrated evaluation
of one node, which is obtained form those nodes who
had transacted with the node. We let Gbi denote the
global buyer reputation, where node i is the buyer and
Gsi to be the global seller reputation, where node i is the
seller. But there are some factors we should consider
when we calculate the value of global reputation.
1) The global buyer (seller) reputation of node i should
be the integrated evaluation of all those nodes who had
transacted with node i.
2) The evaluation of different nodes with different
Trust-Value should be kept separate. That is because the
evaluation of the nodes with higher Trust-Value is more
important than that of nodes with lower Trust-Value.
3) The more the number of transaction between both
parties is, the more credible the evaluation is.
4) The global reputation is an accumulative process, so
only through persistent credible transaction, the value
will be higher.
The formulas are listed as follows.
(,)(,)
5
(,)(,)
5
() (1)
(1)
() (1)
Nsgj iNsbji
j
ji
jVsi
iNsgj iNsbji
j
jVsi
Gs keLs
Gb k
Gs ke






 




(3)
(,)(,)
5
(,)(,)
5
() (1)
(1)
() (1)
Nbgj iNbbji
j
ji
jVbi
iNbgj iNbbj i
j
jVbi
Gb keLb
Gs k
Gb ke






 




(4)
where the meanings of Nsg(j,i), Nsb(j,i), Nbg(j,i) and Nbb(j,i)
are the same as that of Chapter 3.5. Vbi represents the set
of nodes who had transacted with node i and they are
buyer. Vsi represents the set of nodes who had transacted
with node i and they are seller.
The weighted average method not only can embody
the views of all the nodes, but also keep the meaning of
global reputation unchanged. The character of 1-exp(-
(Ng(j,i)-Nb(j,i))/5) negative exponent increase as Ng(j,i)-
Nb(j,i) in accordance with the feature that credible trans-
action can improve the reputation and incredible transac-
tion can decrease the reputation.
3.7. The Computation of Trust-Value
The Trust-Value of node j represented as Tij is the critical
factor for node i to determine whether to transact with
node j. Because it is the integrated value of Lsji and Gsj,
its computation must consider the trust level of node i to
j’s local reputation and global reputation. If node i has
transacted with node j, the Trust-Value of node j is the
weighted sum of Lsji and Gsj. If node i didn’t have any
transactions with node j before, node i can ask his en-
tity-agent and domain-agent to recommend providers.
Both the agents search those entities that had transacted
with node j and ask them to recommend node j. Those
recommenders give views about node j according to the
history performance of node j. Then the agents feed back
these information to the requestor i. To ensure the valid-
ity of recommendation, the rules of recommendation are
shown as follows.
1) The recommendations have time limitation, the
scope of time is [σ1, σ2] which can be defined as the
transaction needs. That is because the previous transac-
tion information can not exactly reflect the SPs’ credibil-
ity of the current situation.
2) Because the recommendation is finite, the depth
that the agent searches recommenders in the binary sort-
ing tree must be smaller than h.
3) In order to reduce the likelihood of collusion, we set
the number of recommenders must more than certain
C opyright © 2009 SciRes. IJCNS

J. F. TIAN ET AL. 770
threshold

where 0

. If the number of recom-
menders is less than

, we can add some virtue nodes
as recommender whose reputation is the same as the ini-
tial Trust-Value and the number of nodes that had trans-
acted with each recommender is. 
Because of the asymmetry of trust, the recommenda-
tion involves the consumer’s trust to the recommender’s
recommendation. The higher trust level the consumer is
to the recommender, the more trust the consumer is to
the recommendation. If recommender has high recogni-
tion about his services, the trust level of his recommen-
dation also can be increased. We let Tij denote the
Trust-Value of node j as opposite to node i. The formula
is shown as follows.
()
()
ji ji
jDir
ij
ir rjrjr
r Undir
Ls Gs
TDir Undir
TrecLs Gs
Dir Undir







 


 (5)
where Dir is the set of nodes who had transacted with
node i directly and Undir is the set of nodes who are
recommended by recommender r to transact with node i.
The Dir and Undir represent the length of the
corresponding set. recr ∈ [0,1] is the recognition degree
of recommender r to service. α and β is the weighting of
local reputation and global reputation separately and
β=1-α. The value of α is computed as follows.
The local reputation is the evaluation of one node ac-
cording to his history transaction behavior with another
node. So if the number of transactions is more, the
transaction cost and the evaluation of the transaction are
higher, the buyer will more trust the local reputation
which is achieved through his direct experiences.
111 1
()(
mnm m
mbig mgood
jij i
CCEvaE 
 

 
)va
where n is the total number of transaction, m is the num-
ber of transactions with high cost and good evaluation.
The user can defined a threshold to determine the value
of m. Cm is the cost of transaction and Cmbig is the high
cost of transaction. Eva is the evaluation of transaction,
Evagood is the good evaluation of transaction and Eva,
Evagood∈[0,1].
3.8. The Harmonic-Mean of Trust-Value and
Similarity Degree
The traditional recommender systems based on trust be-
lieve that the higher the global reputation of recom-
mender is, the more credible the recommendation is. But
in fact the value of global reputation is not in conformity
with the importance of recommendation. For example,
some malicious nodes may achieve high global reputa-
tion value through fake or some nodes collude in order to
remote their Trust-Value or to destroy other competitor.
Thus the recommender not only has the high trust degree
of recommendation but also should make sure that the
content of recommendation is credible. In this paper, in
order to ensure the content of recommendation is credi-
ble, the model imports the semantic-based service
matching method which calculates the similarity degree
between the request service and the provided service.
The service is reliable only when the provided service
satisfies the user’s needs. The computation of similarity
adopts one simple and efficient method provided by pa-
per [24], denoted as WS.
This paper uses the similarity degree to harmonize the
Trust-Value with weighting. The higher the similarity
degree is, the bigger the harmonic-mean is. The har-
monic-mean is the reciprocal of the arithmetic mean of
the reciprocals, which is mainly used to the situation that
the initial digitals is not the direct initial digitals but its
frequency had been computed. The formula is shown as
follows.
2
(1 )
ij ij
ij
ij ij
WS T
ST TW


S
k
(6)
where δ is the adjustment factor; WS ij is the similarity
degree of the request service and provided service; Tij is
the trust level that how much node i trust in node j; STij is
the harmonic-mean of similarity and Trust-Value which
directly determines whether node i transacts with node j
or not.
3.9. The Updates of Trust-Value and the Sharing
of Trust Information
After each transaction, both parties will feed back the
evaluation of this transaction to each other, namely the
satisfaction degree. Both the entity-agents of two parties
need to respectively update the global buyer reputation
(Gbi) of buyer and the global seller reputation (Gsj) of
seller. The updates in time ensure that the users don’t
need to calculate the reputation value when they send the
request, so that the model can respond the user’s request
quickly. The formulas are listed as follows.
1(1)( ,) kk
ii k
GbGbSijC 


 , (7)
1(1)(, ) kk
j
jkk
GsGsSj iCrec 
 k
(8)
where Sk(i,j)∈[0,1] represents the satisfaction degree of
buyer i to seller j after the kth transaction. Sk(j,i)∈[0,1]
represents the satisfaction degree of seller j to buyer i
after the kth transaction. The satisfaction degree indicates
the level of satisfaction achieved by the consumer on the

J. F. TIAN ET AL. 771
]
service provided by the SP. A normalized value between
0 and 1 is used, with 1.0 indicating 100% satisfaction
and 0.0 indicating the lowest satisfaction. Ck= (Cmk-1)/
Cmk represents the weight of transaction cost in the kth
transaction. reck∈[0,1] is the recognition degree of en-
tity to the service that he provided in kth transaction
which is maintained in the binary sorting tree of entity’s
identity by his entity-agent and is registered by entity
when he joined the domain. ,[0,1

  are the weights
which are assigned according to last transaction time.
In addition to this, we also need to update the trust ta-
bles which are related to the entity and maintained by EA
and DA. The update method is pushing the information
to the needed agent.
Updating the reputation value in time can effectively
reduce the response time that the buyer waits from his
agents and it is done among free time without influenc-
ing the transaction. Thus the user can use the reputation
value directly only with time decay such as linear decline,
exponential decline and so on.
Both the transaction parties can set a threshold to de-
termine whether to share trust information or not. That is
to say, they share their trust information only when their
Trust-Value is more than the threshold. The sharing of
trust information makes the trust propagate quickly and
improves the performance of network.
4. Simulations and Analysis
In this paper, the simulation is based on the PeerSim
which is written in the Java language and is based on
components [25]. It can support the extensibility and
dynamic of the P2P network better. And it adopts the
modular design and uses the configuration file to custom
the modules and parameters. Thus it has high expand-
ability. It also provides the interfaces and statistical
methods to generate the network and makes the simula-
tion and the evaluation of one algorithm more easily.
The principle of organization of the RMNRS model
makes us not use the existing protocols of PeerSim di-
rectly. We are obliged to inherit one existing protocol
and write our own algorithm to simulate. The protocol
that we inherit is: Idle Protocol-Average Function. The
control is cycle-based. In this paper we just inherit the
interconnect relation between nodes and we write the
algorithms about the model of EigenRep and RMNRS
autonomously.
According to the character of the nodes in the network,
we divide the nodes into four types: absolute trust nodes,
trust nodes, critical trust nodes and distrust nodes.
Definition 1: absolute trust is the trust that is estab-
lished on the basis of both parties in the partnership ex-
perience long-term transaction and major event test;
Definition 2: trust is the trust that is established on the
basis of both parties in the partnership experience long-
term transaction and general event test;
Definition 3: critical trust is the trust that both parties
in the partnership don’t have sufficient reason to trust or
distrust each other;
Definition 4: distrust is the trust that after the
long-term transaction, both parties in the partnership
don’t trust one at least.
We define the absolute trust node’s value to be T
where T≥Tmax, trust node’s value to be T where
Tmid<T<Tmax, critical trust node’s value to be T where
Tmin≤T≤Tmid, distrust node’s value to be T where T< Tmin,
and 0≤Tmin≤Tmid≤Tmax≤1.
4.1. Experiment 1
According to the paper, we define the network size is
1000, the simulation cycle is 50, the degree of a node is 6
and the init global reputation value is 0.5. We assume
that Tmin=0.3, Tmid=0.6and Tmax=0.9. We define the per-
cent that the four type nodes respectively are 10%, 30%,
30% and 30%. All of the nodes belong to different trust
domains. Each domain has four types of nodes. Entities’
Trust-Value will be changed timely along with the
transaction in the domain or between domains. Along
with the increase of the number of transactions the varia-
tion of the four types of entities’ average Trust-Value is
showed in Figure 2.
Figure 2. The variation of the average Trust-Value of four types of nodes.

J. F. TIAN ET AL. 772
As shown in Figure 2, the initial reputation value of all
nodes is 0.5 and along with the increasing number of
trades, the model updates the reputation value according
to the Equation (5). Finally we calculate the Trust-Value
of each node. According to the definitions of the four
types of nodes, we can find that the result of RMNRS
model is in accordance with the expected analysis. And
along with the increase of the number of trades, it well
reflects that the variation of the nodes’ Trust-Value.
4.2. Experiment 2
The assumption of nodes is the same as the Experiment 1
in Chapter 4.1. In this simulation, we assume the degree
of each node is 3, that is to say, the direct transaction
table of each node maintains three nodes’ trust informa-
tion in the primary stages and the three nodes are se-
lected randomly by the PeerSim protocols. When we
write the program, we let the nodes’ initial reputation
value be random which is between 0 and 1. The experi-
ment is designed to record the selection number from the
node sending out the request to it receives the request.
The network consumption is based on the number of
selection step when the consumer selects the transaction
partner.
In this model, we allocate two global reputations for
each node. Thus the user can select the nodes with higher
global seller reputation in the first stage which expands
the range of selection. Then the model computes the
harmonic mean of the similarity degree and the integra-
tion Trust-Value. The final Trust-Value sincerely reflects
whether the provided service satisfies the users’ needs,
which can reduce the unnecessary selection. However
the EigenRep model just follows the traditional method
to select the provider and the comparison number is more
than the RMNRS model. Figure 3 realistically reflects
the analysis of this paper and verifies the efficiency of
the RMNRS model.
Figure 3. The comparison between RMNRS and EigenRep
in the network consumption.
4.3. Experiment 3
Whether the transaction is success or not is decided by
the satisfaction degree which is fed back by user. If the
satisfaction degree is larger than 0.6 then this transaction
is success and vice versa. We define the success ratio of
transaction to be the proportion of the success number.
In this simulation, we assume that the absolute trust
nodes provide the credible services with the probability
of 100%. And we also assume that the RMNRS and the
EigenRep select the absolute nodes with the probability
of 80%. As seen in Figure 4, when there are not mali-
cious nodes in the environment, the success ratio of
transaction is 95%. Along with the increase of malicious
nodes, the transaction success ratio of EigenRep declines
obviously. When the ratio of malicious nodes is 50%, the
transaction success ratio of EigenRep is only about 60%.
This is because there is lack of punishment to the mali-
cious nodes in the EigenRep. So its success ratio has
bigger drop. The RMNRS punishes the malicious nodes
and matches the services between the request and the
provided services. And it uses the similarity degree to
amend the Trust-Value with the harmonic mean which
ensures that the provided service is the needed service.
The RMNRS avoids the transaction with the malicious
nodes and improves the success ration of transaction.
Under the condition of existing malicious nodes with the
property of 50%, the transaction success ration of the
RMNRS is about 80%. The experiment verifies the fea-
sibility and efficiency of the RMNRS.
5. Conclusions
In this paper we present a reputation-based multi-agent
model for network resource selection (RMNRS) which
prevents the inconsistency of information maintained by
different agents through the periodically communication
between the agents. The model enables the consumer to
receive the response from the agent significantly more
Figure 4. The comparison between RMNRS and EigenRep
for success ratio of the transaction.

J. F. TIAN ET AL. 773
quickly than that of traditional models because the global
reputation values of both parties are evaluated and up-
dated dynamically after the completion of each transac-
tion. And the model allocates two global reputation val-
ues to each entity and takes the recognition value that
how much the service providers know the service into
account. In order to make users choose the best matching
services to their request and give users with trusted ser-
vices, the model also takes the similarity between ser-
vices into account and uses the similarity degree to
amend the integration reputation value with the har-
monic-mean. The following work is to research how to
avoid the sharing of incredible information with mali-
cious nodes and how to punish those malicious nodes.
6. Acknowledgements
This work was supported by the National Natural Sci-
ence Foundation of China (Grant No. 60873203), the
Natural Science Foundation of Hebei Province (Grant No.
F2008000646) and the Guidance Program of the Depart-
ment of Science and Technology in Hebei Province
(Grant No. 072135192).
7. References
[1] H. Ibrahim, P. K. Atrey, and E. S. Abdulmotaleb, “Se-
mantic similarity based trust computation in websites,”
International Multimedia Conference, New York, ACM,
pp. 65–72, 2007.
[2] S. K. Chong and J. H. Abawajy, “Feedback credibility
issues in trust management systems,” 2007 International
Conference on Multimedia and Ubiquitous Engineering:
proceedings: MUE’07, Los Alamitos, Calif., IEEE Com-
puter Society, pp. 387–391, 2007.
[3] Donovan J O and Smyth B, “Trust in recommender sys-
tems,” in proceedings of the 10th international conference
on Intelligent user interfaces, New York, ACM, pp. 167–
174, 2005.
[4] P. Massa and P. Avesani, “Trust-aware recommender
systems,” in proceedings of the 2007 ACM conference on
Recommender systems. New York, ACM, pp. 17–24,
2007.
[5] Y. Gil and D. Artz, “Towards content trust of web re-
sources,” in proceedings of the 15th international confer-
ence on World Wide Web, New York, ACM, pp. 565–
574, 2006.
[6] D. S. Peng, C. Lin, and W. D. Liu, “A distributed trust
mechanism directly evaluating reputation of nodes,”
Journal of Software, Vol. 19, No. 4, pp. 946–955, April
2008.
[7] Y. Wang and V. Varadharajan, “Role-based recommen-
dation and trust evaluation,” in the 9th IEEE International
Conference on E-Commerce. Technology and the 4th
IEEE International Conference on Enterprise Computing,
E–Commerce and E–Services, Tokyo, IEEE, pp. 278–288,
2007.
[8] P. Varalakshmi, S. Thamarai Selvi and M. Pradeep, “A
multi-broker trust management framework for resource
selection in grid,” in Communication Systems Software
and Middleware, COMSWARE’07, 2nd International
Conference on Bangalore, IEEE, pp. 7–12 January 2007.
[9] S. X. Jiang and J. Z. Li, “A reputation–based trust mecha-
nism for p2p e-commerce systems,” Journal of Software,
Vol. 18, No. 10, pp. 2551–2563, 2007.
[10] M. Blaze, J. Feigenbaum, and J. Lacy, “Decentralized
trust management,” in proceedings of the 17th Sympo-
sium on Security and Privacy, CA, IEEE Computer Soci-
ety Press, pp. 164–173, 1996.
[11] C. Lin, L. Q. Tian, and Y. Z. Wang, “Research on user
behavior trust in trustworthy network,” Journal of Com-
puter Research and Development, Vol. 45, No. 12, pp.
2033–2043, 2008.
[12] S. D. Kamvar and M. T. Schlosser, “EigenRep: Reputa-
tion management in P2P networks,” in Lawrence S, ed.
Proceedings of the 12th International World Wide Web
Conference Budapest, ACM Press, pp. 123−134, 2003.
[13] C. F. Wang and F. C. Sun, “Hierarchical entity self-de-
termined trust model based on behaviors in grid envi-
ronment,” Computer Engineering and Applications, Vol.
43, No. 16, pp. 135–138, 2007.
[14] F. Maheswaran and M. Maheswaran, “Evolving and
managing trust in grid computing systems,” in proceed-
ings of the 2002 IEEE Canadian Conference on Electrical
Computer Engineering, pp. 1424–1429, 2002.
[15] F. Azzedin and M. Maheswaran, “A trust brokering sys-
tem and its application to resource management in public
resource grid,” Parallel and distributed Computing Sym-
posium. pp. 22, 2004.
[16] X. Li and L. Liu, “Peertrust: supporting reputation-based
trust for peer-to-peer electronic communities,” IEEE
Transactions on Knowledge and Data Engineering, Spe-
cial Issue on Peer to Peer Based Data Management, Vol.
16, No. 7, pp. 843–857, 2004.
[17] X. Dong, A. Halevy, J. Madhavan, et al., “Similarity
search for web services,” in Proceedings of the Thirtieth
International Conference on Very Large Data Bases,
VLDB Endowment, pp. 372–383, 2004.
[18] X. Wang, Y. H. Ding, and Y. Zhao, “Similarity meas-
urement about ontology-based semantic web services,” in
Conjunction with 4th European Conference on Web Ser-
vices (ECOWS’06), pp. 4–6, 2006.
[19] X. Y. Li and X. L. Gui, “Research on dynamic trust
model for large scale distributed environment,” Journal of
Software, Vol. 18, No. 6, pp. 1510−1521, 2007.
[20] T. Beth, M. Borcherding and B. Klein, “Valuation of trust
in open networks,” in proceedings of the European Sym-
posium on Research in Computer Security, Brighton UK,
Springer-Verlag, pp. 3–18, 1994.
C opyright © 2009 SciRes. IJCNS

J. F. TIAN ET AL. 774
[21] X. Z. Zhang, “The research of virtual enterprise trust
mechanism—the innovation of trust management in the
network environment,” Hunan, Hunan People’s Publish-
ing House, July 2005.
[22] Steve Hanna, Co-Chair, TNC Work Group, TCG. TNC:
Open Standards for Network Access Control. https://
www.trustedcomputinggroup.org.
[23] J. F. Tian, B. Xiao, X. X. Ma, et al, “The trust model and
its analysis in TDDSS,” Journal of Computer Research
and Development, Vol. 44, No. 4, pp. 598–605, April
2007.
[24] X. Qing, “Semantic-based web service discovering algo-
rithm,” Master’s degree, Jinan, Shandong University, pp.
17–26, 2006.
[25] PeerSim: A simulation environment for P2P protocols in
java. Version 1.0.4. 2008