A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust Nathan Griffiths University of Warwick, UK A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.1 Introduction Multiagent systems rely on the interactions of many autonomous (often self-interested) agents Agents have specific individual capabilities, knowledge and resources, and vary in reliability, quality and honesty A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.2 Introduction Multiagent systems rely on the interactions of many autonomous (often self-interested) agents Agents have specific individual capabilities, knowledge and resources, and vary in reliability, quality and honesty When an agent cooperates it is entering into an uncertain interaction, that has an associated risk of failure or reduced performance A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.2 Introduction Multiagent systems rely on the interactions of many autonomous (often self-interested) agents Agents have specific individual capabilities, knowledge and resources, and vary in reliability, quality and honesty When an agent cooperates it is entering into an uncertain interaction, that has an associated risk of failure or reduced performance Agents can use trust to manage this risk This research uses fuzzy logic to allow agents to represent and reason with uncertain and imprecise information regarding others’ trustworthiness. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.2 Trust and Reputation Trust and reputation are related, but distinct, concepts Trust represents an agent’s individual assessment of the reliability, honesty etc. of another Reputation is built from a combination of individual trust assessments (or opinions) A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.3 Trust and Reputation Trust and reputation are related, but distinct, concepts Trust represents an agent’s individual assessment of the reliability, honesty etc. of another Reputation is built from a combination of individual trust assessments (or opinions) Many of the existing applications of trust combine trust and reputation, using a global aggregation of individual trust into a reputation assessment (e.g. Sabater and Sierra; Song et al.; Xiong and Liu, etc.) Existing approaches often do not consider the issues of distrust and lack of trust. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.3 Trust and Reputation To use reputation, we must address issues of: motivating the sharing of private information subjectivity and context of information In some domains it is not clear how to address these. . . Therefore, although both trust and reputation are important, it is useful to treat them as distinct (but related) notions We focus on trust alone — enable agents to make use of what (little?) information they have. (NB: we see trust and reputation as complimentary; ideally we would have both, but sometimes reputation is not practicable.) A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.4 Fuzzy Logic for Trust Although trust is based on previous experience, there is inherent uncertainty and imprecision: assessments are subjective and agents may change unpredictably over time A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.5 Fuzzy Logic for Trust Although trust is based on previous experience, there is inherent uncertainty and imprecision: assessments are subjective and agents may change unpredictably over time Fuzzy logic enables representation of this uncertainty and imprecision Existing models (Manchala; Ramchurn et al.; Song et al.) have successfully used fuzzy logic to represent trust in multiagent systems However, these approaches use trust as a means of establishing reputation, rather than as means to directly support decision making. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.5 Basic Fuzzy Concepts Sets have fuzzy boundaries: objects have a degree of membership µ(x) ∈ [0 : 1] Fuzzy sets are used to define terms with respect to a variable, e.g. the term young ^ on the variable age Can use linguistic hedges to modify terms, e.g. very young ^ Relations between variables can be defined using fuzzy inference rules, e.g. if age is young ^ and income is g g then customerPotential is high very high Rules are applied in parallel and the conclusion membership degrees aggregated, and a crisp value extracted. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.6 Trust Trust is generally taken to be the belief that an agent will act in the best interests of another (i.e. will cooperate), even if given the opportunity to do otherwise (i.e. to defect) A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.7 Trust Trust is generally taken to be the belief that an agent will act in the best interests of another (i.e. will cooperate), even if given the opportunity to do otherwise (i.e. to defect) Distrust is not simply the negation of trust, but rather it is a belief that an agent will act against the best interests of another [Marsh] A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.7 Trust Trust is generally taken to be the belief that an agent will act in the best interests of another (i.e. will cooperate), even if given the opportunity to do otherwise (i.e. to defect) Distrust is not simply the negation of trust, but rather it is a belief that an agent will act against the best interests of another [Marsh] Untrust corresponds to the space between distrust and trust, in which an agent is positively trusted, but not sufficiently to cooperate with [Marsh] Undistrust is a negative trust, but insufficient to make definite conclusions in the reasoning process. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.7 Trust Marsh’s model: D UT −1 T 0 1 The addition of undistrust: D −1 UT UD 0 T 1 A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.8 Dimensions of Trust Trust is based on an agent’s experiences, and so agents must track their interaction histories, and record whether their expectations have been met Trust comprises the combination of the different characteristics of an interaction A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.9 Dimensions of Trust Trust is based on an agent’s experiences, and so agents must track their interaction histories, and record whether their expectations have been met Trust comprises the combination of the different characteristics of an interaction Agents can model such characteristics as dimensions of trust, which taken together give an assessment of an agent’s trustworthiness For example, we use the dimensions of success, cost, quality and timeliness. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.9 Interaction Histories Agents maintain a history of the interactions that they have had with an agent α, and track the number of successful (Iαd+ ) and unsuccessful (Iαd− ) interactions for each dimension d, in terms of whether their expectations were met The experience, edα , in each dimension d is: d+ d− I − I α edα = αd+ Iα + Iαd− This crisp value is fuzzified as Eαd = fuzzySingleton(edα ). A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.10 Interaction Histories Agents track of the outcomes of their interactions by using a window of experience that is maintained for each agent A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.11 Interaction Histories Agents track of the outcomes of their interactions by using a window of experience that is maintained for each agent Over time, stored information may become outdated Outdated experiences are purged: the record of an interaction is removed after a purge lag Small purge lag: records do not persist, quick response to changes, may give keep insufficient information; Long purge lag copes with perturbations, slow to respond A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.11 Interaction Histories Agents track of the outcomes of their interactions by using a window of experience that is maintained for each agent Over time, stored information may become outdated Outdated experiences are purged: the record of an interaction is removed after a purge lag Small purge lag: records do not persist, quick response to changes, may give keep insufficient information; Long purge lag copes with perturbations, slow to respond Confidence level given by confidence dα = Iαd+ + Iαd− Low confidence implies untrust or undistrust. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.11 Fuzzy Trust Define fuzzy terms for experience: NB NM NS Z PS PM PB 1 −1 −0.5 0 0.5 1 0 A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.12 Fuzzy Trust Define fuzzy terms for experience: NB NM NS Z PS PM PB 1 −1 −0.5 0 0.5 1 0 and for trust: HD D UT UD T HT 1 −1 −0.5 0 0.5 1 0 A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.12 Fuzzy Trust For each dimension we define a set of fuzzy inference rules that take the fuzzified experiences as antecedents and make conclusions regarding trust, e.g. (RU T 1) if confidence dα < minConfidence and Eαd is positive ^ then Tα is untrust (RU T 2) if confidence dα < minConfidence and Eαd is negative ^ then Tα is undistrust (etc . . . ) (RT 3) ^ ^ Eαd is negativeBig then Tα is highDistrust d ^ ^ or undistrust ^ if Eα is negativeMedium then Tα is very distrust d is negativeSmall ^ ^ if Eα then Tα is undistrust (RT 4) if (RT 1) (RT 2) if (RT 5) ^ or untrust ^ Eαd is zero g then Tα is undistrust d ^ ^ if Eα is positiveSmall then Tα is untrust (RT 6) if ^ ] or untrust ^ (etc . . . ) Eαd is positiveMedium then Tα is very trust A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.13 Determining Trust Check whether there have been sufficient interactions; if not assign untrust or undistrust (rules RU T n) Otherwise, apply fuzzy inference rules (RT n), e.g. 1 1 PS RT 5 PM RT 6 0 0 0 0.25 0 0.5 0.5 T UT 0 0.25 0.5 0 0 0.22 0.5 0.5 A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.14 Agent Selection Distrusted agents filtered out: reject if ^ ) > maxDistrust similarity(Tα , highDistrust A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.15 Agent Selection Distrusted agents filtered out: reject if ^ ) > maxDistrust similarity(Tα , highDistrust Untrusted and undistrusted agents: if no trusted agents then with some probability (rebootstrap rate) the agent with the highest trust level from the set of untrusted and undistrusted agents will be selected A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.15 Agent Selection Distrusted agents filtered out: reject if ^ ) > maxDistrust similarity(Tα , highDistrust Untrusted and undistrusted agents: if no trusted agents then with some probability (rebootstrap rate) the agent with the highest trust level from the set of untrusted and undistrusted agents will be selected Otherwise, combine trust with other factors (advertised cost, quality, etc.) to determine agent rating using rules of the form: g ^ and Fαc is medium ^ and Fαq is very high (RR n) if Tα is highTrust g then Rα is high g ^ and Fαc is medium ^ and Fαq is high (RR m)if Tα is low distrust g (etc. . . ) then Rα is low A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.15 Bootstrapping Initially agents have insufficient experience for reasoning Each agent goes through a bootstrapping phase in which partners are chosen randomly by way of exploration During bootstrapping undistrusted, untrusted, and trusted agents have an equal chance of being selected. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.16 Experimental Results Implemented using the NRC FuzzyJ Toolkit Evaluate different configurations of the fuzzy decision mechanism for exactly the same set of possible interactions (i.e. use a fixed generated “environment history”) Simulated an agent selecting a partner from 50 others for 2000 interactions (bootstrapping phase = 100 interactions) Reliability of agents reduced after 1000 interactions. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.17 Experimental Results Success rate 1 fuzzy fuzzy - no distrust fuzzy - no untrust fuzzy - no untrust or distrust random 0.8 0.6 0.4 0.2 0 0 500 1000 1500 2000 A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.18 Experimental Results Results averaged across 20 “environments”: Selection mechanism Success rate Cost rate Quality rate Fuzzy trust 0.64 0.73 0.84 No distrust 0.63 0.68 0.81 No untrust, undistrust 0.63 0.58 0.79 No distrust, untrust, undistrust 0.61 0.56 0.78 Random (control) 0.42 0.39 0.64 A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.19 Summary We have shown how fuzzy logic can be used to represent trust We have proposed a new notion of undistrust and incorporated this into the reasoning process Initial experimental results indicate that the approach provides a significant reduction in failure rate, and improvement in cost and quality rates We have shown how distrust, untrust and undistrust improve performance There are many areas of ongoing work, including additional experimentation, and integration with existing models of reputation. A Fuzzy Approach to Reasoning with Trust, Distrust and Insufficient Trust – p.20