Besides the basic issues of building an interaction trust measure and a reputa- tion measure presented in the previous sections (Sections 2.2 and 2.3), there are a number of generic issues that can affect the performance of a trust model. These include: bootstrapping, dynamism in open MAS, inaccurate reports, and corre- lated evidence. They are going to be discussed in turn in subsequent subsections (Sections 2.4.1 to 2.4.4).
2.4.1
Bootstrapping
Newly joined agents who have no acquaintances will face various difficulties in join- ing the community. Typically, a new agent should have an initial set of contacts to establish its first interactions, as well as to collect reputation information about some initial potential partners. In addition, new agents may find themselves not accepted by some service providers because of their low initial reputation. How- ever, this problem is ignored in most of the research in this area. We believe that solving the bootstrapping issue is necessary so that agents will be able to make use of a trust model in any situations (desideratumR1). Therefore, we add a new requirement for our trust model:
R1a: The trust model should be able to deal with the bootstrapping issue of newly joined agents.
2.4.2
Dynamism in open MAS
As discussed Section 1.1, due to its openness, the environment in an open MAS will change continually. Possible types of changes include:
• The agent population. Existing agents leave the environment and new ones join on a continual basis. This means that an agent might have to repeatedly learn about new agents since its previous interaction partners may no longer be available.
• Agent behaviour. Since agents are owned by different stakeholders, their goals and motivations may change over time. In addition, an agent’s situation may also change. Now all of these will result in agents’ changing their behaviours. For example, an honest agent may become a liar, or an agent may reduce its service quality due to less resources available to it.
• Relationships between agents. Agents may break old relationships and make new ones depending on their situations and needs. For example, virtual organisations can be automatically formed or disbanded according to the participants’ capabilities and goals [Norman et al., 2004]. Hence, the rela- tionships, and thus the trust, between them are also changed over time. Given such a wide range of changes that can happen in an open MAS, a trust model for such an environment should reasonably maintain its normal effective operations under these types of changes. However, none of the existing trust models explicitly take such dynamism into account and none of them has been demonstrated to cope well with it. Therefore, in order for a trust model to be suitable for open MAS (Requirement R2), it should reasonably maintain its normal effective operations in situations where various changes in an open MAS take place (here called Requirement R2c).
2.4.3
Inaccurate reports
As agents in open MAS are self-interested, they may lie when being asked for their observations if they can gain some benefits from so doing (see [Schillo et al., 2000] for an example). In an attempt to solve this problem, the model of Schillo et al. shows how witness information can be reliably used to reason effectively against lying. However, the model greatly simplifies direct interactions (e.g. coop- erate/defect in the disclosed Prisoner’s Dilemma), thus, it is not useful in realistic settings.
To help overcome this problem, Jurca and Faltings [2003] presented a model in which agents pay for reputation information. When an agent needs to find rep- utation information, it contacts an R-agent to buy the information. Agents also receive money when reporting their observations to R-agents, but only after the verification of their reports. In this context, a mechanism is devised to determine the specific amount for each payment so that the agents that report truthfully will not lose money and agents that report falsely will lose money. Thus, lying agents
will gradually lose their money until they do not have enough to buy reputation information. This mechanism makes it rational for an agent to report its obser- vations honestly. However, the idea of side payment may not be feasible in an open MAS. For example, in order to have the rational property mentioned above, the model of Jurca and Faltings requires that the currency for side payment is unexchangeable with the currency used in ordinary transactions. In open MAS, devising a new currency system that is different from the traditional ones, to be accepted by the agents from various origins is not practical.
Regret uses fuzzy rules to classify the reliability of the witnesses based on their relationships with the target agent (see Section 2.3.2 and Section 2.3.3). In this way, they also take into account the possibility of a witness lying based on fuzzy rules. However, this approach is a preventive measure and is based on social information, which is not always available in every situation. In our opinion, the trustworthiness of a witness in reporting its observations should also be taken into account. The reason for this is that the experiences with the witness (i.e. interaction trust) or the relationships between the witness and the collecting agent (i.e. role-based trust) are more reliable than social information.
In order to determine the accuracy of third-party ratings, Whitby et al. [2004] assume that the “true” rating of an agent is defined by the majority’s opinions. In particular, they model the performance of an agent as a beta PDF which is aggregated from all witness ratings received. Then a witness is considered unre- liable and filtered out when the reputation derived from its ratings is judged to be too different from the majority’s (by comparing the reputation value with the PDF). Since this method bases it decisions entirely on PDFs of witness reports, if these reports are scarce and/or too diverse it will not be able to recognise lying witnesses. Moreover, it is possible that a witness can lie in a small proportion of their reports without being filtered out. To rectify this, TRAVOS provides a probabilistic method for filtering out the opinions of inaccurate reputation sources. Reputation is shared in the form of frequencies of successful and unsuccessful con- tracts that the reputation source has had with the trustee, which after interacting with the trustee itself, the truster compares with its own observations. By this means, the truster calculates the probability that the reputation source’s informa- tion supports the true behaviour of the trustee within a reasonable margin of error, and uses this probability to weight the impact of the reputation source’s opinions on future decisions made be the truster. However, TRAVOS’s dependence on its binary rating system again is its weakness.
Yu and Singh propose a similar approach to that of Whitby et al. Specifically, they use a weighted majority algorithm to adjust the weight for each witness over time. Although the weights of the deceitful agents are reduced, these agents are never disregarded completely. Several successful applications of this approach have been demonstrated, but only for agent populations where deceitful agents are in the minority and are balanced between agents that falsely exaggerate their friends’ performance and those that defame other agents.
In summary, all the proposed approaches above are limited in that they require additional domain knowledge or make unrealistic assumptions about the environ- ment. In order to fulfill the RequirementR4of robustness, an effective mechanism is needed to deal with inaccuracy reports (here called RequirementR4a).
2.4.4
Correlated evidence
This problem happens when the opinions of different witnesses are based on the same event or when there is a considerable amount of information shared among a group of agents that make their opinions similar to each other. In both cases, the reliability of the information should not be as high as the number of similar opinions suggests [Sabater and Sierra, 2002]. Sabater and Sierra use graph anal- ysis techniques to address this issue. The process starts with identifying graph components of a domain dependent sociogram. Then an agent in each component will be selected to be the representative agent for all agents in the component. Witnesses will be selected from those representative agents only. However, a node that deems to be representative for a component in a sociogram is not necessarily able to give a full witness’ account for all the agents in the component and, there- fore, choosing only one agents from those in a component may discard possible unique witness reports of the rest. Moreover, the approach is based on heuristics and there is no empirical result presented to show its capabilities. The problem of correlated evidence affects the efficiency and robustness of a reputation model and should be dealt with (here called the requirement R4b).
Requirements
R1 The trust model should be able to provide an effective trust measure that can
R1a deal with the bootstrapping issue of newly joined agents.
R1b make use of role-based trust, interaction trust, and wit- ness reputation when the required information for these dimensions of trust is available.
R2 The trust model should be suitable for open MAS. In particu- lar,
R2a each agent should be able to collect observations and cal- culate the reputation values by itself.
R2b the trust model should be scalable to a large number of agents that might be present in open MAS.
R2c the trust model should reasonably maintain its normal effective operation in situations where there are various changes in its environment.
R3 The trust model should be adaptable to different domains of applications that an open MAS may have.
R4 The trust model should be robust against
R4a possible lying from agents.
R4b the correlated evidence problem.
Table 2.1: The requirements for a trust model in open MAS.