In the literature, the reputation mechanisms are mostly applied to large scale multi-agent sys- tems, which host numerous interactive agents that seek to find the highly reputed agents to interact with. As mentioned before, the trust and reputation are very close concepts in the sense
that in most frameworks both concepts are analyzed and addressed without a clear separation line. In general, the reputation refers to public opinion about a particular agent and the trust is a reliability that is measured by an agent regarding others. In the following, we continue the survey by presenting different models that compute agents’ reputation values. Since the data aggregation processes are very similar, we do not repeat the same analysis in this chapter, but we investigate the system integrity and accuracy in long-term interactive networks. We mainly concentrate on the techniques that guarantee long-term effective reputation management sys- tem.
Unlike trust frameworks, we do not categorize the reputation models against ideal repu- tation system characteristics. The reason is simply because the same set of criteria that we mentioned for trust frameworks could be applied to reputation assessment systems. Our con- cern in reputation systems is mainly establishing a sound and long-term secure mechanism allowing active agents in a dynamic multi-agent environment to achieve their goals. We gener- ally aim at advancing the reputation mechanisms to function accurately and safe in long term interactions. We go beyond the computation problem, which was our main concern in trust frameworks. We would like to maintain a sound and secure reputation system within which the accuracy is achieved via provided incentives to the interacting agents. This point of view is new and does not overlap with the target of the related works that concentrate on addressing the rep- utation mechanism. Consequently, we discuss different reputation frameworks as well as their characteristics. We continue with highlighting the needs of an efficient reputation mechanism that last long enough in a multi-agent system with dynamic behaviors.
2.3.1
Bayesian Network Model
The Bayesian network model [89] is an interactive/witness reputation model in which service provider’s behavior is analyzed in different aspects such as download speed, quality, and type. In this model, a na¨ıve Bayesian network represents conditional dependencies between the relia- bility of the service provider and the analyzed aspects. Therefore, each service consumer agent constructs a na¨ıve Bayesian network regarding each service provider agent. The user might
get access to the na¨ıve Bayesian network constructed by other agents in case it does not have enough information about the particular service provider. In this case, the evaluating agent con- siders the reliability of the recommender in its decision-making process. The reputation values are continuously updated using RL model’s formulation [38].
The Bayesian network model is constructed based on the similarity of the service consumer and provider agents and their preferences in interactions. This approach in unsuitable in the cases where the preferences of the recommender and evaluator agents do not perfectly match and the collected information does not accurately represent the trustworthiness of the service provider agent. In fact, this model fails to accurately combine the obtained data when the recommenders had weak interacting relationship with the particular agent under study.
2.3.2
Weighted Majority Algorithm (WMA)
The weighted majority algorithm is a discrete algebraic method [98] that uses Dempster-Shafer theory [15, 76] to compute the trust value of agents. This approach falls into witness report category. The method is based on three parameters: belief (b), disbelief (d), and uncertainty (u) that sum up to1. An orthogonal sum function is defined in this method that aggregates the parameters to combine impressions of different witness ratings. In WMA, there are assigned values for witness agents which reflect their reliability from evaluator agent’s point of view. These weights are assigned by the evaluator agent and updated due to reliability changes of the witness agents.
In [97], the weights are increased and decreased based on their positive or negative in- fluence in correct reputation evaluation maintained by the evaluator agent. This algorithm is inadequate since the evaluator agent might unfairly decrease the weights when it does not have enough information about the provider agent. There is no methodology defined to fix the mis- taken weight updates. Additionally, the witness agent would not get the chance to increase its associated weight once been decreased. The method does not characterize a realistic multi- agent environment where agents dynamically change their acting strategies.
2.3.3
Cluster Filtering Approach
Cluster filtering approach [20] is another witness trust-based mechanism that mainly concen- trates on discarding inaccurate ratings to compute a general overview regarding a particular agent’s reputation. In this model, a collaborative filtering technique [1] is used to recognize the most trustworthy agents for the evaluating agent. Using cluster filtering approach, a particu- lar evaluating agent divides its surrounding agents into high and low rating clusters. The high rating cluster represents reports with relatively high inaccuracy and therefore are discarded. This approach also considers most recent reports to avoid confusion in clustering. This model addresses the agileness of the reputation computation, but the whole technique does not accu- rately function in some situations where multi-agent network hosts agents with rapid change of behavior.
In [20], the author concludes that by controlling anonymity, the inaccurate witness reports cannot be minimized because of collusion that could be established between the agent to be evaluated and the witness agent. Moreover, this framework works with higher efficiency in large networks where agents’ relocation is not highly considered. The concept of collusion emerges when agents act in an open environment and individuals by default do not have full knowledge about their surrounding environment and agents might expect better outcome by colluding with other agents. This issue of collusion is fully analyzed in Chapter 5.
2.3.4
Robust Reputation System for Mobile Ad-hoc Networks (RRS-
MAN)
In [11], authors propose a robust reputation system for mobile Ad-hoc networks. This system is based on distributed individuals and the objective is to handle false disseminated information. In this model, agents have a belief set regarding the reliability and public reputation of other agents. The reliability is used as the probability that the agent will provide truthful information. The reputation reflects agent’s influence in decision making maintained by the agent that holds this information. This model associates different ratings for the collected data and therefore,
categorizes the surrounding agents in its belief set. To handle false information provided by other agents, RRSMAN updates the reputation ratings with respect to their accuracy in com- parison with other random trusted advisers. The collected data from an agent is considered accurate if the difference with the held reputation is less than a deviation threshold. In the RRSMAN model, the collected data are investigated in the order they are received. This could decrease the accuracy of the model in cases where a high load of non-relevant data needs to be analyzed where an important evidence of environment changes is reported. Furthermore, the time discount factor is not considered in this model and therefore, old information are treated the same as new ones. In open multi-agent environments with dynamic behavior changes, this approach quickly fails since the used approach in collecting data fails in stochastic environ- ments that host agents with different ranges of behaviors.
2.3.5
Discussion on Reputation Frameworks
In this section, we categorized some reputation frameworks that mainly focus on reputation computation and the ways to keep it accurate. The attempts to compute such a value is similar to the ones we discussed in trust frameworks. This is the reason we do not go into further details about comparing different approaches of aggregating some collected data. We would like to highlight the fact that in multi-agent systems there are rational agents that by default follow their goals and they could take actions that benefit themselves whereas others undergo some payoff loss. For example, in an environment where agents exchange services based on some fees, the goal of the service provider agent could be to charge as many agents as possible. To achieve this goal, the agent needs to maintain a high reputation value to absorb other agents’ attentions. In such a system, if there is a way to claim high reputation, the rational service provider agent would resort to any way to maintain it. Consequently, the reputation should be competitive and the reputation mechanism should be robust against malbehavior of agents to mislead the environment with fake reputation values. Moreover, the reputation mechanism should last long and this is fulfilled when such a mechanism can update itself with respect to dynamic change of environment.
Considering the related work, we feel the need of a robust mechanism that imposes in- centives in such a way that malicious behavior of agents are minimized. In general, we need a mechanism to maintain accuracy of the reputation system by preparing a situation within which agents seek maximum payoff by acting truthfully and compete for high reputation rather than colluding for fake reputation values to temporary increase their reputations. Chapter 4 is asso- ciated to propose such a reputation mechanism that its objective is to tackle the incentive-based reputation assessment problem. Chapter 5 and 6 are also about a robust reputation mechanism which is claimed to be sound and secure in long term interactive interval. The combination of these three chapters develop a complete reputation model that could be implemented in multi-agent environments where there is a need to cope with malicious agents and constrain the accuracy of the network and safety of the interacting agents’ transactions. The structure used in these chapters is the network of web services that could be also grouped together as community of web services. The rational behind the use of this structure is the enterprise system of web services and their rational activities while exchanging services between one another. Moreover, web services (that are attached to agents as web service agents) compete to serve their services and obtain utilities that could be in the form of service fee or any other sort of payoff. We explore more details about this structure in the following section.