HARM: A Hybrid Rule-based Agent Reputation Model based on Temporal Defeasible Logic
1. HARM
A Hybrid Rule-based Agent Reputation
Model based on Temporal Defeasible Logic
Kalliopi Kravari, Nick Bassiliades
Department of Informatics
Aristotle University of Thessaloniki
Thessaloniki, Greece
2. OVERVIEW
Agents in the SW interact under uncertain and risky situations.
Whenever they have to interact with partners of whom they know nothing..
they have to make decision involving risk.
Thus:
Their success may depend on their ability to choose reliable partners.
Solution:
Reliable trust and/or reputation models.
SW Intelligent
SW Trust Layer
evolution Agents
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
2
3. OVERVIEW
Trust is the degree of trust that can be invested in a certain agent.
Reputation is the opinion of the public towards an agent.
Reputation (trust) models provide the means to quantify
reputation and trust
help agents to decide who to trust
encourage trustworthy behavior
deter dishonest participation
Current computational reputation models are usually built either on
interaction trust or witness reputation.
an agent’s direct experience
reports provided by
others
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
3
4. APPROACHES’ LIMITATIONS
If the reputation estimation is based only on direct experience, it would
require a long time for an agent to reach a satisfying estimation level.
Why?
because when an agent enters an environment for the first time,
it has no history of interactions with the other agents in the environment.
If the reputation estimation is based only on witness reports, it could
not guarantee reliable estimation.
Why?
because self-interested agents could be unwilling or unable
to sacrifice their resources in order to provide reports.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
4
5. HYBRID MODELS
Hybrid models combine both interaction trust and witness reputation.
We propose HARM:
an incremental reputation model that combines
the advantages of the hybrid reputation models
the benefits of temporal defeasible logic (rule-based approach)
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
5
6. (Temporal) Defeasible Logic
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
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
6
7. Defeasible Logic
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
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
7
8. Temporal Defeasible Logic
Two types of temporal literals:
expiring temporal literals l:t (a literal l is valid for t time instances)
persistent temporal literals l@t
(a literal l is active after t time instances have passed and is valid
thereafter)
temporal rules: a1:d1 ... an:dn d b:db
delay between the cause a1:d1 ... an:dn and the effect b:db
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.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
8
9. HARM – Evaluated Abilities
Evaluates four agent abilities:
validity, completeness, correctness and response time.
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
An agent is complete if it is both cooperative and vigilant.
Cooperative: says what it believes
Vigilant: believes what is true in the world
An agent is correct if its provided service is correct with
respect to a specification.
Response time is the time that an agent needs to complete
the transaction.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
9
10. HARM - Ratings
Agent A establishes interaction with agent B:
(A)Truster is the evaluating agent
(B) Trustee is the evaluated agent
Truster’s rating value (r) in HARM has 8 coefficients:
2 IDs: Truster, Trustee
4 abilities:
Validity, Completeness, Correctness, Response time
2 weights: Confidence, Transaction value
Confidence: how confident the agent is for the rating
Transaction value: how important the transaction was for the
agent
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
10
11. HARM – Experience Types
Direct Experience (PRAX )
Indirect Experience
reports provided by strangers (SRAX)
reports provided by known agents (e.g friends) due to
previous interactions (KRAX )
Both
Final reputation value of an agent X,
required by an agent A:
RAX = {PRAX , KRAX, SRAX}
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
11
12. HARM – Experience Types
Sometimes one or more rating categories are 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.
▫ Thus, personal opinion (AX) is more valuable than strangers’ opinion
(SX), as well as it is more valuable even from previously trusted partners
(KX).
Superiority relationship among rating categories
AX, KX, SX
AX, KX AX, SX KX, SX
AX KX SX
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
12
13. HARM – Final reputation value
RAX is a function that combines each available category
▫ personal opinion (AX)
▫ strangers’ opinion (SX)
▫ previously trusted partners (KX)
RAX PR AX , KR AX , SR AX
HARM allows agents to define weights of ratings’
coefficients
▫ Personal preferences
coefficient coefficient coefficient
AVG wi log prAX AVG wi log krAX AVG wi log srAX
RAX 4
, 4
, 4
,
i 1
wi i 1
wi i 1
wi
coefficient validity, completeness, correctness, response _ time 2
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
13
15. HARM
Rule-based Decision Making Mechanism
Confidence and transaction value allow us to decide
how much attention we should pay on each rating.
It is important to take into account ratings that were
made by confident trusters, since their ratings are more
likely to be right.
Confident trusters, that were interacting in an
important for them transaction, are even more likely to
report truthful ratings.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
15
16. HARM
Which ratings “count”?
r1: count_rating(rating→?idx, truster→?a, trustee→ ?x) :=
confidence_threshold(?conf),
• if both truster’s confidence and
transaction_value_threshold(?tran), transaction importance are high,
rating(id→?idx, confidence→?confx, then that rating will be counted
transaction_value→?tranx), during the estimation process
?confx >= ?conf, ?tranx >= ?tran.
r2: count_rating(…) := • if the transaction value is lower than
the threshold, it doesn’t matter so much
… if the truster’s confidence is high
?confx >= ?conf.
• if there are only ratings with high
r3: count_rating(…) :=
transaction value,
… then they should be taken into account
?tranx >= ?tran.
r1 > r2 > r3 • In any other case, the rating
should be omitted.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
16
17. HARM
Conflicting Literals
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 }
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
17
18. HARM
Determining Experience Types
r4: known(agent1→?a, agent2→?y) :- Which agents
count_rating(rating → ?id, truster→?a, trustee→?y). are considered
as known?
r5: count_prAX(agent→?a, truster→?a, trustee→?x, rating→?id) :-
count_rating(rating → ?id, truster→? a, trustee→ ?x).
r6: count_krAX(agent→?a, truster→?k, trustee→?x, rating →?id) :-
known(agent1→?a, agent2→?k), Categori
zation of
count_rating(rating→?id, truster→?k, trustee→ ?x).
ratings
r7: count_srAX(agent→?a, truster→?s, trustee→?x, rating→?id) :-
count_rating(rating → ?id, truster →?s, trustee→ ?x),
not(known(agent1→?a, agent2→?s)).
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
18
19. HARM
Rule-based Decision Making Mechanism
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).
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
19
20. HARM
Theory #1: All categories count equally
AX, KX, SX
AX, KX AX, SX KX, SX
AX KX SX
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
20
21. HARM
Rule-based Decision Making Mechanism
e.g. Theory #2:
An agent relies on its own experience if it believes it is
sufficient.
If not it acquires the opinions of others.
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).
r8>r9>r10
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
21
22. HARM
Theory #2: Personal experience is preferred to
friends’ opinion to strangers’ opinion
AX, KX, SX
AX, KX AX, SX KX, SX
AX > KX > SX
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
22
23. HARM
Rule-based Decision Making Mechanism
e.g. Theory #3:
If direct experience is available (PRAX), then it is preferred
to be combined with ratings from known agents (KRAX).
If not, HARM acts as a pure witness system.
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).
r8> r10, r9>r10
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
23
24. HARM
Theory #3: Personal experience and friends’
opinion is preferred to strangers’ opinion
AX, KX, SX
AX, KX AX, SX KX, SX
>
AX KX SX
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
24
25. HARM
Temporal Defeasible Logic Extension
Agents may change their 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
The truster’s rating (r) is active after time_offset time
instances have passed and is valid thereafter
rating(id→value1, truster→value2, trustee→ value3, validity→value4,
completeness→value5, correctness→value6,
response_time →value7, confidence→value8,
transaction_value→value9)@time_offset.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
25
26. HARM
Temporal Defeasible Logic Extension
Rules are modified accordingly:
each rating is active after t time instances have passed (“@t”)
each conclusion has a duration (“:duration”)
each rule has a delay, which models the delay between the cause and
the effect.
e.g.
r1: 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.
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
26
27. HARM Evaluation
We implemented the model in EMERALD
Framework for interoperating knowledge-based
intelligent agents in the SW.
Built on JADE multi-agent platform
EMERALD uses Reasoners (agents offering
reasoning services)
Supports the DR-Device defeasible logic system
Used temporal predicates to simulate the temporal
semantics.
o No available temporal defeasible logic reasoner
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
27
28. HARM - Evaluation
All agents provide the same service
Performance is service – independent
Consumer agent selects the provider with the highest
reputation value
5. Report rating
1. Request reputations of the provider agents
2. Inform about the provider with the highest reputation
Consumer agent HARMAgent
4. Service providing
3. Service request
Provider agenty
Provider agentx
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
28
29. HARM - Evaluation Number of simulation: 500
Performance of providers (e.g. Number of providers: 100
quality of service) is the utility Good providers 10
that a consumer gains from each Ordinary providers 40
interaction
Intermittent
Utility Gain (UG), UG [-10, 10] providers
5
Four models are used: Bad providers 45
HARM (rule-based / temporal)
T-REX (temporal degradation) Average UG per interaction
SERM (uses all history) HARM 5.73
NONE (no trust mechanism). T-REX 5.57
SERM 2.41
From previous study:
NONE 0.16
CR, SPORAS (literature famous, CR 5.48
distributed)
SPORAS 4.65
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
29
30. Conclusions
We proposed HARM that combines:
the hybrid approach (interaction trust and witness reputation)
the benefits of temporal defeasible logic (rule-based approach)
Overcomes the difficulty to locate witness reports
(centralized administration authority)
It is the first reputation model that uses explicitly
knowledge, in the form of defeasible logic, to predict
agent’s future behavior
Easy to simulate human decision making
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
30
31. Future Work
Fully implement HARM with temporal defeasible logic
Compare HARM’s performance with other centralized
and decentralized models from the literature
Combine HARM and T-REX
Develop a distributed version of HARM
Verify its performance in real-world e-commerce
applications
Combining it with Semantic Web metadata for trust
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
31
32. Thank you!
Any Questions?
Nick Bassiliades RuleML 2012, Montpellier, Aug 27-29
32
Editor's Notes
..deter dishonest participation by providing the mean through which reputation and ultimately trust can be quantified
Hence, models based only on one or the other approach typically cannot guarantee stable and reliable estimations.
The testbed environment for evaluating HARM is a multi-agent system consisting ofagents providing services and agents using these services in an on-line community.We assume that the performance of a provider (and effectively its trustworthiness) isindependent from the service that is provided. In order to reduce the complexity of the testbed’s environment, it isassumed that there is only one type of service in the testbed.The value of UG variesfrom −10 to 10 and depends on the level of performance of the provider in the interaction.Τώρα όσον αφορά τα άλλα μοντέλα που χρησιμοποιούμε ούτε στο paper λέμε κάτι. Μπορείτε να πείτε αν θέλετε ότι το μόνο το HARM έχει temporal, ενώ το t-rexλαμβάνει υπόψη τον χρόνο θεωρώντας τα πιο πρόσφατα rating πιο σημαντικά με την χρήση ουσιαστικά του clock (Timer) ενώ το setmδεν λαμβάνει υπόψη τον χρόνο.
The testbed environment for evaluating HARM is a multi-agent system consisting ofagents providing services and agents using these services in an on-line community.We assume that the performance of a provider (and effectively its trustworthiness) isindependent from the service that is provided. In order to reduce the complexity of the testbed’s environment, it isassumed that there is only one type of service in the testbed.The value of UG variesfrom −10 to 10 and depends on the level of performance of the provider in the interaction.Τώρα όσον αφορά τα άλλα μοντέλα που χρησιμοποιούμε ούτε στο paper λέμε κάτι. Μπορείτε να πείτε αν θέλετε ότι το μόνο το HARM έχει temporal, ενώ το t-rexλαμβάνει υπόψη τον χρόνο θεωρώντας τα πιο πρόσφατα rating πιο σημαντικά με την χρήση ουσιαστικά του clock (Timer) ενώ το setmδεν λαμβάνει υπόψη τον χρόνο.