Trust and reputation models among agents

Nick Bassiliades
Intelligent Systems group,
Software Engineering, Web and Intelligent Systems Lab,
Dept. Informatics, Aristotle University of Thessaloniki, Greece
Invited talk at
6th International Conference on Integrated Information
Sep 19-22, 2016, Athens, Greece

 Introduction
 Review of State-of-the-Art on Trust / Reputation Models
 Centralized approaches
 Distributed approaches
 Hybrid approaches
 Presenter’s work on Trust / Reputation Models
 T-REX
 HARM
 DISARM
 Summary / Conclusions
IC-ININFO, Sep 19-22, 2016N. Bassiliades - Trust and reputation models among agents 2

 Agents are supposed to act in open and risky environments
(e.g. Web) with limited or no human intervention
 Making the appropriate decision about who to trust in order to
interact with is necessary but challenging
 Trust and reputation are key elements in the design and
implementation of multi-agent systems
 Trust is expectation or belief that a party will act benignly and
cooperatively with the trusting party
 Reputation is the opinion of the public towards an agent, based
on past experiences of interacting with the agent
 Reputation is used to quantify trust

 Interaction trust: agent’s own direct experience from past
interactions (aka reliability)
- Requires a long time to reach a satisfying estimation level
 When no history of interactions with other agents in a new environment
 Witness reputation: reports of witnesses about an agent’s
behavior, provided by other agents
- Does not guarantee reliable estimation
 Are self-interested agents willing to share information?
 How much can you trust the informer?

 Centralized approach:
 One or more centralized trust authorities keep agent interaction
references (ratings) and give trust estimations
 Convenient for witness reputation models (e.g. eBay, SPORAS, etc.)
+ Simpler to implement; better and faster trust estimations
- Less reliable; Unrealistic: hard to enforce central controlling
authorities in open environments
 Decentralized (distributed) approach:
 Each agent keeps its own interaction references with other agents
and must estimate on its own the trust upon another agent
 Convenient for interaction trust models
+ Robustness: no single point of failure; more realistic
- Need more complex interaction protocols

 Hybrid models: Combination of Interaction Trust and Witness
Reputation
 Regret / Social Regret
 FIRE
 RRAF / TRR
 CRM
 T-REX / HARM / DISARM
 Certified reputation: third-party references provided by the agent
itself
 Distributed approach for witness reputation
Centralized / Distributed

WITNESS REPUTATION
 Agent A wants a service from agent B
 Agent A asks agent C if agent B can be
trusted
 Agent C trusts agent B and replies yes to
A
 Agent A now trusts B and asks B to
perform the service on A’s behalf
 A = truster / beneficiary,
C = trustor / broker / consultant,
B = trustee
A
Truster /
Beneficiary
B
Truste
e
C
Trustor /
Broker /
Consultan
t

INTERACTION TRUST
 Agent A judges if B is to be trusted from
personal experience
 Agent A trusts B and asks B to perform
the service on A’s behalf
 A = trustor / truster / beneficiary,
B = trustee
A
Trustor /
Truster /
Beneficiary
B
Truste
e

REFERENCES
 Agent A asks agent B for proof of trust
 Agent B provides some agents R that can
guarantee that B can be trusted
 Agent A now trusts B and asks B to
perform the service on A’s behalf
 A = trustor / truster / beneficiary,
B = trustee,
R = referee
A
Truster /
Beneficiary
B
Truste
e
R
Referee
R
Referee

IC-ININFO, Sep 19-22, 2016N. Bassiliades - Trust and reputation models among agents
10

 Centralised rating system
 After an interaction eBay users rate partner (−1, 0,+1)
 positive, neutral, negative
 Ratings stored centrally
 Reputation value is computed as the sum of ratings over 6
months
 Reputation is a global single value
- Too simple for applications in MAS (1-D)
 All ratings count equally
- Cannot adapt well to changes in a user’s performance
 E.g. a user may cheat in a few interactions after obtaining a high
reputation value, but still retains a positive reputation

 Centralized repository and reputation / reliability estimation
 Ratings are not kept – reputation updated after each transaction
 New agents start with a minimum reputation
value and build up reputation
 The reputation value of an agent never falls
below the reputation of a newcomer
 Users with very high reputation values experience much smaller rating
changes after each update
 Ratings are discounted over time so that the most recent ratings have
more weight in the evaluation of an agent’s reputation
+ More adaptable to changes in behavior over time
 Reliability measure = ratings deviation
 High deviation:
 Agent has not been active enough to have a more accurate reputation prediction
 Agent’s behaviour has a high degree of variation.
+ Prevent agents with bad
reputation leaving and
entering with fresh
reputation
- May discourage newcomers

13

 Decentralized reputation model
 Εach agent is able to evaluate the reputation of others by itself
 Keeps own ratings about other partners in a local database
 Direct trust is calculated as the weighed average of all ratings
 Each rating is weighed according to its recency
 More recent ratings are weighted more
- Time granularity control does not actually reflect a rating’s recency
 Reliability
 How many ratings are taken into account?
 Deviation of ratings

 An attempt to locate witnesses’ ratings through a social graph
 Agent groups with frequent interactions among them
 Each group is a single source of reputation values
 Only the most representative agent within each group is asked for
information
 Heuristics are used to find groups and to select the best agent to ask
- Social Regret does not reflect the actual social relations among
agents
 Attempts to heuristically reduce the number of queries to be asked in
order to locate ratings
- Considers the opinion of only one agent of each group
 Most agents are marginalized, distorting reality.

 Decentralized, witness reputation model
 Agents cooperate by giving, pursuing, and evaluating referrals
 Each agent maintains a list of acquaintances (other agents it
knows) and their expertise
 When looking for certain information, an agent queries its
acquaintances
 They will try to answer the query, if possible
 If not, they will send back referrals of other agents they believe are
likely to have the desired information
 Agent’s expertise is used to determine how likely it is to
 have interaction with
 know witnesses of the target agent

 Agents are incentivized to report truthfully interactions’
results
 Broker agents
 Buy and aggregate (average) reports from other agents
 Sell needed information to agents
+ Mechanism guarantees that agents who report incorrectly will
gradually lose money, while honest agents will not
± Brokers are distributed, but each one collects reputation
values centrally
- Reputation values limited to 0, 1

 Decentralized witness reputation model
 Each agent keeps references given to it from other agents
 After every transaction, each agent asks partners to provide their
certified ratings about its performance
 It chooses the best ratings to store in its (local) database
 When an agent A contacts B to use service C, it asks B to provide
references about its past performance with respect to C
 Agent A receives certified ratings of B from B and calculates CR of B
+ Agents do not have costs involved in locating witness reports
- Ratings might be misquoted
 Each agent only provides the best ratings about itself

19

 Distributed, hybrid model
 Integrates 4 types of trust / reputation: interaction trust, role-
based trust, witness reputation and certified reputation
 Role-based trust: defined by various role-based relationships between
the agents
 Provides means for domain-specific customization of trust, using rules
 All values are combined into a single measure by weighted
average
- Weak computation model for the combined reputation estimation
 Does not take into consideration the problem of lying and
inaccuracy

 Hybrid reputation model
 Combines interaction trust and witness reputation into a single value,
using weighted sum
 Dynamically computes weights
 Previous RRAF model used static weights assigned by the user
 Weight depends on:
 Number of interactions between two agents
 Expertise that an agent has in evaluating other agents capabilities
 Reputation is based on the global trust that community has in an
agent
+ More accurate reputation prediction
- Difficult to implement in a distributed environment

 Distributed, hybrid, probabilistic-based reputation model
 On-line trust estimation:
 Direct trust evaluation, using own interaction history
 Consulting reports: indirect trust estimation from consulting
agents
 trustworthy agents
 referee agents (introduced by the trustee agent as recommenders)
 Off-line trust estimation:
 Off-line interaction inspection: After interaction, trustor assigns a
(useful/useless) flag for each consulting agent
 Rating & confidence of provided information
 Number and recency of interactions with the trustee agent
 Maintenance: update information about consulting agents
 Initiated at different intervals of time

23

 Centralized
 T-REX
 Hybrid
 HARM
 Hybrid
 Knowledge-based
 Temporal Defeasible Logic
 Distributed
 DISARM
 Hybrid
 Knowledge-based, Defeasible Logic
 Social relationships

Kravari, K., Malliarakis, C., & Bassiliades, N.: T-REX: A Hybrid Agent
Trust Model Based on Witness Reputation and Personal Experience. Proc.
11th International Conference on Electronic Commerce and Web
Technologies (EC-Web 2010). Springer, LNBIP, Vol. 61, Part 3, pp. 107-
118 (2010)
25

 Centralized, hybrid reputation model
 Combines Witness Reputation (ratings) & Direct Trust
(personal experience)
 Weighted average (user-defined weights)
 Ratings decay over time
 All past ratings are taken into account
 Time-stamps are used as weights
 Past ratings loose importance over time through a linear
extinguishing function
+ Low bandwidth, computational and storage cost

 Central ratings repository: Trustor
 A special agent responsible for collecting, storing, retrieving
ratings and calculating trust values
 Considered certified/reliable
 Interacting agents
 Truster / Beneficiary: an agent that wants to interact with another
agent that offers a service
 Trustee: the agent that offers the service
 Role of Trustor
 Before the interaction, Truster asks from Trustor calculation of
Trustees trust value
 After the interaction, Truster submits rating for Trustee’s
performance to Trustor

Truster Trustee
interact?
Trustor
Calculates reputation
from stored ratings
and agent’s weights
Gives personalized weights
for each rating criteria

Truster
Trusteeinteract
Trustor
Evaluation criteria
• Correctness
• Completeness
• Response time

 Rating Vector
 rx ∈ [0.1, 10] | 0.1 = terrible , 10 = perfect
 Normalized Ratings
 Ratings normalized logarithmically to
cross out extreme values
 log(rx) ∈ [-1, 1] | -1 = terrible, 1 = perfect
 Weighted Normalized Ratings
 Agents customize the importance of
criteria using weights
 a: truster
 b: trustee
 Corr: correctness
 Resp: response time
 Comp: completeness
 t: time stamp
 px: weight of rating
rx
𝑅 𝑎𝑏(𝑡 = 𝐶𝑜𝑟𝑟 𝑎𝑏(𝑡 𝑅𝑒𝑠𝑝 𝑎𝑏(𝑡 𝐶𝑜𝑚𝑝 𝑎𝑏(𝑡 = 𝑟1
𝑎𝑏
𝑡 𝑟2
𝑎𝑏
𝑡 𝑟3
𝑎𝑏
𝑡
rab t = Rab t = x=1
3
px∙ log rx
ab t
x=1
3
px

The final reputation value (TR)
Transactions between a and b
(interaction trust)
Transactions of b with
all other agents
(witness report)
social trust weights (πp, πo)
Time is important 
more recent ratings “weigh” more
𝑇𝑅 𝑎𝑏 𝑡 =
𝜋 𝑝
𝜋 𝑝 + 𝜋 𝑜
∙
∀𝑡𝑖<𝑡 𝑟𝑎𝑏 𝑡𝑖 ∙ 𝑡𝑖
∀𝑡𝑖<𝑡 𝑡𝑖
+
𝜋 𝑜
∙
∀𝑗≠𝑎,𝑗≠𝑏
∀𝑡𝑖<𝑡 𝑟𝑗𝑏 𝑡𝑖 ∙ 𝑡𝑖

 Dynamicity
 Trust depends on time
 Flexibility
 Can be customized from completely witness-based to completely
personal-experience-based
 Reliability
 Centralized repository (Trustor) is assumed honest
 Low bandwidth
 Trusters report in one communication step
 Trustees do not communicate at all
 Low storage cost
 Trusters and trustees do not store anything
 Trustor stores everything
 Computational cost?

Original Formula
Computational complexity O(n*t)
n: number of agents / t: amount of
transactions
Swap order
of sums ti, x
Sums can be calculated incrementally, when a new report arrives
𝜋 𝑝
∙
∀𝑡𝑖<𝑡
𝑥=1
3
𝑝 𝑥∙ log 𝑟𝑥
𝑎𝑏
𝑡
𝑥=1
3
𝑝 𝑥
∙ 𝑡𝑖
+
𝜋 𝑜
∙
∀𝑡𝑖<𝑡
𝑥=1
3
𝑝 𝑥∙ log 𝑟𝑥
𝑗𝑏
𝑡
𝑥=1
3
𝑝 𝑥
∙ 𝑡𝑖
𝜋 𝑝
∙ 𝑥=1
3
𝑝 𝑥∙ ∀𝑡𝑖<𝑡 log 𝑟𝑥
𝑎𝑏
𝑡 ∙ 𝑡𝑖
𝑥=1
3
𝑝 𝑥 ∙ ∀𝑡𝑖<𝑡 𝑡𝑖
+
𝜋 𝑜
∙
𝑥=1
3
𝑝 𝑥∙ ∀𝑡𝑖<𝑡 log 𝑟𝑥
𝑗𝑏
𝑡 ∙ 𝑡𝑖
𝑥=1
3
𝑝 𝑥 ∙ ∀𝑡𝑖<𝑡 𝑡𝑖
𝜎 𝑥
𝑞𝑦
𝑡 =
∀𝑡 𝑖<𝑡
log 𝑟𝑥
𝑞𝑦
𝑡 ∙ 𝑡𝑖 𝑇 𝑡 =
∀𝑡 𝑖<𝑡
𝑡𝑖

Computational complexity O(n)
n: number of agents
Storage complexity O(n2)
Swap order
of sums j, x
This sum can be calculated incrementally
𝜋 𝑝
∙ x=1
3
𝑝 𝑥∙ 𝜎 𝑥
𝑎𝑏
(𝑡
𝑥=1
3
𝑝 𝑥 ∙ 𝑇(𝑡
+
𝜋 𝑜
∙
𝑥=1
3
𝑗𝑏
(𝑡
𝑥=1
3
𝜋 𝑝
∙ 𝑥=1
3
𝑎𝑏
(𝑡
𝑥=1
3
+
𝜋 𝑜
∙ 𝑥=1
3
px∙ ∀𝑗≠𝑎,𝑗≠𝑏 𝜎 𝑥
𝑗𝑏
(𝑡
𝑥=1
3
𝑆 𝑥
𝑎𝑏
𝑡 =
𝜎 𝑥
𝑗𝑏
(𝑡

Computational complexity  reduced to O(1)
Storage complexity  remains O(n2)
𝜋 𝑝
∙ 𝑥=1
3
𝑎𝑏(𝑡
𝑥=1
3
+
𝜋 𝑜
∙ 𝑥=1
3
𝑝 𝑥∙ 𝑆 𝑥
𝑎𝑏(𝑡
𝑥=1
3

Kravari, K., & Bassiliades, N. (2012). HARM: A Hybrid Rule-based Agent
Reputation Model Based on Temporal Defeasible Logic. 6th International
Symposium on Rules: Research Based and Industry Focused (RuleML-
2012). Springer, LNCS 7438: 193-207.
36

 Centralized hybrid reputation model
 Combine Interaction Trust and Witness Reputation
 Rule-based approach
 Temporal defeasible logic
 Non-monotonic reasoning
 Ratings have a time offset
 Indicates when ratings become active to be considered for trust
assessment
 Intuitive method for assessing trust
 Related to traditional human reasoning

 Temporal defeasible logic (TDL) is an extension of defeasible
logic (DL).
 DL is a kind of non-monotonic reasoning
 Why defeasible logic?
 Rule-based, deterministic (without disjunction)
 Enhanced representational capabilities
 Classical negation used in rule heads and bodies
 Negation-as-failure can be emulated
 Rules may support conflicting conclusions
 Skeptical: conflicting rules do not fire
 Priorities on rules resolve conflicts among rules
 Low computational complexity

 Facts: e.g. student(Sofia)
 Strict Rules: e.g. student(X)  person(X)
 Defeasible Rules: e.g. r: person(X)  works(X)
r’: student(X)  ¬works(X)
 Priority Relation between rules, e.g. r’ > r
 Proof theory example:
 A literal q is defeasibly provable if:
 supported by a rule whose premises are all defeasibly provable AND
 q is not definitely provable AND
 each attacking rule is non-applicable or defeated by a superior counter-
attacking rule

 Temporal literals:
 Expiring temporal literals l:t
 Literal l is valid for t time instances
 Persistent temporal literals l@t
 Literal l is active after t time instances have passed and is valid thereafter
 Temporal rules: a1:t1 ... an:tn d b:tb
 d is the delay between the cause and the effect
 Example:
(r1) => a@1 Literal a is created due to r1.
(r2) a@1=>7 b:3 It becomes active at time offset 1.
It causes the head of r2 to be fired at time
8.
The result b lasts only until time 10.
Thereafter, only the fact a remains.

 Validity
 An agent is valid if it is both sincere and credible
 Sincere: believes what it says
 Credible: what it believes is true in the world
 Completeness
 An agent is complete if it is both cooperative and vigilant
 Cooperative: says what it believes
 Vigilant: believes what is true in the world
 Correctness
 An agent is correct if its provided service is correct with respect to
a specification
 Response time
 Time that an agent needs to complete the transaction

 Agent A establishes interaction with agent B:
 (A) Truster is the evaluating agent
 (B) Trustee is the evaluated agent
 Truster’s rating value has 8 coefficients:
 2 IDs: Truster, Trustee
 4 abilities: Validity, Completeness, Correctness, Response time
 2 weights (how much attention agent should pay on each rating?):
 Confidence: how confident the agent is for the rating
 Ratings of confident trusters are more likely to be right
 Transaction value: how important the transaction was for the agent
 Trusters are more likely to report truthful ratings on important transactions
 Example (defeasible RuleML / d-POSL syntax):
rating(id→1,truster→A,trustee→B,validity→5,completeness→6,
correctness→6,resp_time→8,confidence→0.8,transaction_val→0.9).

 Direct Experience (PRAX )
 Indirect Experience
 reports provided by strangers (SRAX)
 reports provided by known agents (e.g. friends) due to previous
interactions (KRAX )
 Final reputation value
 of an agent X, required by an agent A
RAX = {PRAX , KRAX, SRAX}

 One or more rating categories may be missing
 E.g. a newcomer has no personal experience
 A user is much more likely to believe statements from a trusted
acquaintance than from a stranger.
 Personal opinion (AX) is more valuable than strangers’ opinion (SX)
and known partners (KX).
 Superiority relationships among rating categories
KX
AX, KX, SX
AX, KX AX, SX KX, SX
AX SX

 RAX is a function that combines each available category
 personal opinion (AX)
 strangers’ opinion (SX)
 previously trusted partners (KX)
 HARM allows agents to define weights of ratings’ coefficients
 Personal preferences
 , ,AX AX AXAXR PR KR SR 
        
 
4 4 4
1 1 1
log log log
, , ,
, , , _
coefficient coefficient coefficient
i AX i AX i AX
AX
i i ii i i
AVG w pr AVG w kr AVG w sr
R
w w w
coefficient validity completeness correctness response time
  
   
  
 
 

  

r1: count_rating(rating→?idx, truster→?a, trustee→ ?x) :=
confidence_threshold(?conf), transaction_value_threshold(?tran),
rating(id→?idx, confidence→?confx, transaction_val→?tranx),
?confx >= ?conf, ?tranx >= ?tran.
r2: count_rating(…) :=
…
?confx >= ?conf.
r3: count_rating(…) :=
…
?tranx >= ?tran.
r1 > r2 > r3
• if both confidence and transaction importance
are high, then rating will be used for estimation
• if transaction value is lower than the threshold,
but confidence is high, then use rating
• if there are only ratings with high transaction
value, then they should be used
• In any other case, omit the rating

 All the previous rules are conclude positive literals.
 These literals are conflicting each other, for the same pair of
agents (truster and trustee)
 We want in the presence e.g. of personal experience to omit strangers’
ratings.
 That’s why there is also a superiority relationship between the rules.
 The conflict set is formally determined as follows:
C[count_rating(truster→?a, trustee→?x)] =
{ ¬ count_rating(truster→?a, trustee→?x) } 
{ count_rating(truster→?a1, trustee→?x1) | ?a  ?a1 ∧ ?x  ?x1
}

known(agent1→?a, agent2→?y) :-
count_rating(rating → ?id, truster→?a, trustee→?y).
count_pr(agent→?a, truster→?a, trustee→?x, rating→?id) :-
count_rating(rating → ?id, truster→? a, trustee→ ?x).
count_kr(agent→?a, truster→?k, trustee→?x, rating →?id) :-
known(agent1→?a, agent2→?k),
count_rating(rating→?id, truster→?k, trustee→ ?x).
count_sr(agent→?a, truster→?s, trustee→?x, rating→?id) :-
count_rating(rating → ?id, truster →?s, trustee→ ?x),
not(known(agent1→?a, agent2→?s)).
Which agents are considered as
known?
Ratingcategories

 Final step is to decide whose experience will “count”: direct, indirect
(witness), or both.
 The decision for RAX is based on a relationship theory
 e.g. Theory #1: All categories count equally.
r8: participate(agent→?a, trustee→?x, rating→?id_ratingAX) :=
count_pr(agent→?a, trustee→?x, rating→ ?id_ratingAX).
r9: participate(agent→?a, trustee→?x, rating→?id_ratingKX) :=
count_kr(agent→?a, trustee→?x, rating→ ?id_ratingKX).
r10: participate(agent→?a, trustee→?x, rating→?id_ratingSX) :=
count_sr(agent→?a, trustee→?x, rating→ ?id_ratingSX).

KX
AX, KX, SX
AX SX
ALL CATEGORIES COUNT
EQUALLY

THEORY #2
PERSONAL EXPERIENCE IS
PREFERRED TO FRIENDS’
OPINION TO STRANGERS’
OPINIONr8: participate(agent→?a, trustee→?x, rating→?id_ratingAX) :=
r8 > r9 > r10

KX
AX, KX, SX
AX SX
PERSONAL EXPERIENCE IS
PREFERRED TO FRIENDS’
OPINION TO STRANGERS’
OPINION
>>

THEORY #3
PERSONAL EXPERIENCE AND
FRIENDS’ OPINION IS
PREFERRED TO STRANGERS’
OPINIONr8: participate(agent→?a, trustee→?x, rating→?id_ratingAX) :=
r8 > r10, r9 > r10

KX
AX, KX, SX
AX SX
PERSONAL EXPERIENCE AND
FRIENDS’ OPINION IS
PREFERRED TO STRANGERS’
OPINION
>

 Agents may change their behavior / objectives at any time
 Evolution of trust over time should be taken into account
 Only the latest ratings participate in the reputation estimation
 In the temporal extension of HARM:
 each rating is a persistent temporal literal of TDL
 each rule conclusion is an expiring temporal literal of TDL
 Truster’s rating is active after time_offset time instances have
passed and is valid thereafter
rating(id→val1, truster→val2, trustee→ val3, validity→val4,
completeness→val5, correctness→val6, resp_time→val7,
confidence→val8, transaction_val→value9)@time_offset.

 Rules are modified accordingly:
 each rating is active after t time instances have passed
 each conclusion has a duration that it holds
 each rule has a delay between the cause and the effect
count_rating(rating→?idx, truster→?a, trustee→?x):duration :=delay
confidence_threshold(?conf),
transaction_value_threshold(?tran),
rating(id→?idx, confidence→?confx,transaction_value→?tranx)@t,
?confx >= ?conf, ?tranx >= ?tran.

K. Kravari, N. Bassiliades, “DISARM: A Social Distributed Agent
Reputation Model based on Defeasible Logic”, Journal of Systems
and Software, Vol. 117, pp. 130–152, July 2016
57

 Distributed extension of HARM
 Distributed hybrid reputation model
 Combines Interaction Trust and Witness Reputation
 Ratings are located through agent’s social relationships
 Rule-based approach
 Defeasible logic
 Non-monotonic reasoning
 Time is directly used in:
 Decision making rules about recency of ratings
 Calculation of reputation estimation (similar to T-REX)
 Intuitive method for assessing trust
 Related to traditional human reasoning

 Social relationships of trust among agents
 If an agent is satisfied with a partner it is more likely to interact
again in the future
 If dissatisfied it will not interact again
 Each agent maintains 2 relationship lists:
 White-list: Trusted agents
 Black-list: Non-trusted agents
 All other agents are indifferent (neutral zone)
 Each agent decides which agents are added / removed from
each list, using rules
 Personal social network

 Truster’s rating value has 11 coefficients:
 2 IDs: Truster, Trustee
 4 abilities: Validity, Completeness, Correctness, Response time
 2 weights: Confidence, Transaction value
 Timestamp
 Cooperation: willingness to do what is asked for
 Important in distributed social environments
 Outcome feeling: (dis)satisfaction for the transaction outcome
 Degree of request fulfillment
 Example (defeasible RuleML / d-POSL syntax):
rating (id→1, truster→A, trustee→X, t→140630105632, resp_time→9,
validity→7, completeness→6, correctness→6, cooperation→8,
outcome_feeling→7, confidence→0.9, transaction_val→0.8)
3 more than
HARM

...
2. Propagating request
Agent A
Truster
WL agents providing ratings
1. Request ratings
3. Receive ratings
Agent X
Trustee
5. Choose agent x
Service request
6. Receive service
4. Evaluate Reputation
(DISARM rules + ratings)
7. Rate agent X

good_behavior(time → ?t, truster→ ?a, trustee→ ?x, reason → all) :-
resp_time_thrshld(?resp), valid_thrshld(?val), …, trans_val_thrshld(?trval),
rating(id→?idx, time → ?t, truster→ ?a, trustee→ ?x, resp_time→?respx,
validity→?valx, transaction_val→?trvalx, completeness→?comx,
correctness→?corx, cooperation→?coopx, outcome_feeling→?outfx),
?respx<?resp, ?valx>?val, ?comx>?com, ?corx>?cor, ?coopx>?coop, ?outfx>?outf.
bad_behavior(time → ?t, truster→ ?a, trustee→ ?x, reason → response_time) :-
rating(id→?idx, time → ?t, truster→ ?a, trustee→ ?x, resp_time→?respx),
resp_time_thrshld(?resp), ?respx >?resp.
 Any combination of parameters can be used with any defeasible theory.

 Has been good twice for the same reason
add_whitelist(trustee→ ?x, time → ?t2) :=
good_behavior(time→?t1, truster→?self, trustee→?x, reason→?r),
good_behavior(time→?t2, truster→?self, trustee→?x, reason→?r),
?t2 > ?t1.
 Has been bad thrice for the same reason
add_blacklist(trustee→ ?x, time → ?t3) :=
bad_behavior(time→?t1, truster→?self, trustee→?x, reason→?r),
?t2 > ?t1, ?t3 > ?t2.

blacklist(trustee→ ?x, time → ?t) :=
¬whitelist(trustee→ ?x, time → ?t1),
add_blacklist(trustee→ ?x, time → ?t2), ?t2 > ?t1.
¬blacklist(trustee→ ?x, time → ?t2) :=
blacklist(trustee→ ?x, time → ?t1),
add_whitelist(trustee→ ?x, time → ?t2),
?t2 > ?t1.
whitelist(trustee→ ?x, time → ?t) :=
¬blacklist(trustee→ ?x, time → ?t1),
add_whitelist(trustee→ ?x, time → ?t2), ?t2 > ?t1.
…
Add to the blacklist
Remove from the
blacklist
Add to the whitelist

 Ask for ratings about an agent sending request messages
 To whom and how?
 To everybody
 To direct “neighbors” of the agent’s “social network”
 To indirect “neighbors” of the “social network” though message
propagation for a predefined number of hops (Time-to-Live - P2P)
 “Neighbors” are the agents in the whitelist
 Original request:
send_message(sender→?self, receiver→?r,
msg →request_reputation(about→?x,ttl→?t)) :=
ttl_limit(?t), whitelist(?r), locate_ratings(about→?x).

 Upon receiving request, return rating to the sender
send_message(sender→?self, receiver→?s,
msg →rating(id→ idx, truster→ ?self, trustee→ ?x, …)) :=
receive_message(sender→?s, receiver→?self,
msg →request_rating(about→?x)),
rating(id→?idx, truster→ ?self, trustee→ ?x, …).
 If time-to-live has not expired propagate request to all friends
send_message(sender→?s, receiver→?r,
msg →request_reputation(about→?x, ttl→?t1)):=
receive_message(sender→?s, receiver→?self,
msg →request_rating(about→?x,ttl→?t)),
?t >0, WL(?r), ?t1 is ?t - 1.

 Direct Experience (PRX)
 Indirect Experience (reports provided by other agents):
 “Friends” (WRX) – agents in the whitelist
 Known agents from previous interactions (KRX)
 Complete strangers (SRX)
 Final reputation value
 RX = {PRX, WRX, KRX, SRX}
KR
PR, WR, KR, SR
PR, WR PR, SR KR, SR
PR SR
PR, KR
WR
WR, SRWR, KR
PR, WR, SR WR, KR, SRPR, KR, SRPR, WR, KR
New compared to
HARM

 According to user’s preferences
eligible_rating(rating→?idx,truster→?a,trustee→?x,reason→cnf_imp) :=
conf_thrshld(?conf), trans_val_thrshld(?tr),
rating(id→?idx,truster→?a,trustee→?x,conf→?confx,trans_val→?trx),
?confx >= ?conf, ?trx >= ?tr.
 According to temporal restrictions
count_rating(rating→?idx, truster→?a, trustee→?x) :=
time_from_thrshld(?ftime), time_to_thrshld(?ttime),
rating(id→?idx, t→?tx, truster→?a, trustee→ ?x),
?ftime <=?tx <= ?ttime.

count_wr (rating →?idx, trustee→?x) :-
eligible_rating(rating → ?idx, cat→?c, truster→?k, trustee→ ?x),
count_rating(rating→?idx, truster→?k, trustee→ ?x),
known(agent→?k),
whitelist (trustee →?k).
count_kr (rating →?idx, trustee→?x) :-
eligible_rating(rating→?idx, cat→?c, truster→?k, trustee→ ?x),
count_rating(rating→?idx, truster→?k, trustee→ ?x),
known(agent→?k),
not(whitelist(trustee →?k)),
not(blacklist (trustee →?k)).

 When ratings provided by an agent are outside the standard
deviation of all received ratings, the agent might behave dishonestly
bad_assessment (time → ?t, truster→ ?y, trustee→ ?x) :-
standard_deviation_value(?t,?y,?x,?stdevy),
standard_deviation_value (?t,_,?x,?stdev),
?stdevy > ?stdev.
 When two bad assessments for the same agent were given in a
certain time window, trust is lost
remove_whitelist(agent→ ?y, time → ?t2) :=
whitelist(truster→ ?y),
time_window(?wtime),
bad_assessment(time → ?t1, truster→ ?y, trustee→ ?x),
bad_assessment(time → ?t2, truster→ ?y, trustee→ ?x),
?t2 <= ?t1 + ?wtime.

71

 Simulation in the EMERALD* multi-agent
system
 Service provider agents
 All provide the same service
 Service consumer agents
 Choose provider with the higher reputation value
 Performance metric: Utility Gain
*K. Kravari, E. Kontopoulos, N. Bassiliades, “EMERALD:
A Multi-Agent System for Knowledge-based Reasoning
Interoperability in the Semantic Web”, 6th Hellenic
Conference on Artificial Intelligence (SETN 2010),
Springer, LNCS 6040, pp. 173-182, 2010.
Number of simulations:
500
Number of providers:
100
Good providers 10
Ordinary
providers
40
Intermittent
providers
5
Bad providers 45

HARM T-REX No Trust CR SPORAS
5.73 5.57 0.16 5.48 4.65

0
1
2
3
4
5
6
7
MeanUG
Time
DISARM Social Regret Certified Reputation
CRM FIRE HARM
NONE
Mean Utility Gain

Better performance when alone,
due to more social relationships

0
5
10
15
20
25
30
MemorySpacec(%)
Time
DISARM Social Regret Certified Reputation CRM FIRE HARM
Storage Space

79

 Interaction Trust (personal experience) vs. Witness Reputation
(Experience of others)
 Hybrid models
 Centralized (easy to locate ratings) vs. Distributed (more robust)
 Several State-of-the-Art models have been presented
 SPORAS, Regret, Certified Reputation, Referral, FIRE, TRR, CRM, …
 Presenter’s and associates trust / reputation models
 T-REX (centralized, hybrid, time decay, computationally optimized)
 HARM (centralized, hybrid, knowledge-based, temporal defeasible
logic)
 DISARM (distributed, hybrid, knowledge-based, defeasible logic, time
decay, social relationships, manages dishonesty)

 Centralized models
+ Achieve higher performance because they have access to more information
+ Simple interaction protocols, easy to locate ratings
+ Both interaction trust and witness reputation can be easily implemented
- Single-point-of-failure
- Cannot scale well (bottleneck, storage & computational complexity)
- Central authority hard to enforce in open multiagent systems
 Distributed models
- Less accurate trust predictions, due to limited information
- Complex interaction protocols, difficult to locate ratings
- More appropriate for interaction trust
+ Robust – no single-point-of-failure
+ Can scale well (no bottlenecks, less complexity)
+ More realistic in open multiagent systems

 Interaction trust
+ More trustful
- Requires a long time to reach a satisfying estimation level
 Witness reputation
- Does not guarantee reliable estimation
+ Estimation is available from the beginning of entering a community
 Hybrid models
+ Combine interaction trust and witness reputation
- Combined trust metrics are usually only based on arbitrary /
experimentally-optimized weights

 Centralized models
- Cannot scale well (bottleneck, storage & computational complexity)
+ T-REX reduces computational complexity at the expense of storage
complexity
+ HARM reduces computational complexity by reducing considered ratings,
through rating selection based on user’s domain-specific knowledge
 Distributed models
- Less accurate trust predictions, due to limited information
- Complex interaction protocols, difficult to locate ratings
+ DISARM finds ratings through agent social relationships and increases
accuracy by using only known-to-be-trustful agents
 Hybrid models
- Combined trust metrics are usually only based on arbitrary weights
+ HARM & DISARM employ a knowledge-based highly-customizable (both to
user prefs & time) approach, using non-monotonic defeasible reasoning

 Aristotle University of Thessaloniki, Greece
 Largest University in Greece and South-East Europe
 Since 1925, 41 Departments, ~2K faculty, ~80K students
 Dept. of Informatics
 Since 1992, 28 faculty, 5 research labs, ~1100 undergraduate students,
~200 MSc students, ~80 PhD students, ~120 PhD graduates, >3500
pubs
 Software Engineering, Web and Intelligent Systems Lab
 7 faculty, 20 PhD students, 9 Post-doctorate affiliates
 Intelligent Systems group (http://intelligence.csd.auth.gr)
 4 faculty, 9 PhD students, 16 PhD graduates
 Research on Artificial Intelligence, Machine Learning / Data Mining,
Knowledge Representation & Reasoning / Semantic Web, Planning,
Multi-Agent Systems
 425 publications, 35 projects

 The work described in this talk has been performed in
cooperation with Dr. Kalliopi Kravari
 Former PhD student, currently postdoctorate affiliate
 Occasional contributors:
 Dr. Christos Malliarakis (former MSc student, co-author)
 Dr. Efstratios Kontopoulos (former PhD student, co-author)
 Dr. Antonios Bikakis (Lecturer, University College London, PhD
examiner)

 K. Kravari, C. Malliarakis, N. Bassiliades, “T-REX: A hybrid agent trust
model based on witness reputation and personal experience”, Proc. 11th
International Conference on Electronic Commerce and Web Technologies
(EC-Web 2010), Bilbao, Spain, Lecture Notes in Business Information
Processing, Vol. 61, Part 3, Springer, pp. 107-118, 2010.
 K. Kravari, E. Kontopoulos, N. Bassiliades, “EMERALD: A Multi-Agent
System for Knowledge-based Reasoning Interoperability in the Semantic
Web”, 6th Hellenic Conference on Artificial Intelligence (SETN 2010),
Springer, LNCS 6040, pp. 173-182, 2010.
 K. Kravari, N. Bassiliades, “HARM: A Hybrid Rule-based Agent
Reputation Model based on Temporal Defeasible Logic”, 6th
International Symposium on Rules: Research Based and Industry
Focused (RuleML-2012). Springer Berlin/Heidelberg, LNCS, Vol. 7438,
pp. 193-207, 2012.
 K. Kravari, N. Bassiliades, “DISARM: A Social Distributed Agent
Reputation Model based on Defeasible Logic”, Journal of Systems and
Software, Vol. 117, pp. 130–152, July 2016

Trust and reputation models among agents

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (20)

Similar to Trust and reputation models among agents

Similar to Trust and reputation models among agents (20)

Recently uploaded

Recently uploaded (19)

Trust and reputation models among agents