Research On Evaluation Techniques Of Reputation Mechanisms In Autonomous Resource Aggregation

TherequirementsofnetworkapplicationsstimulatethedevelopmentofInternettech-nology. During this process, how to facilitate the resource aggregation or sharing in sucha distributed, opening platform plays the key role in the study of the field. Recently, alongwith the growing number of Internet users, the wide deployment of diverse applicationsandtherapidgrowthofrelevantsupportedtechnology,thenewInternetenvironmenturgespeople to find a more efficient way to aggregate distributed, dynamic and autonomous re-sources available in end-users, and then promote the efficient sharing and comprehensiveutilizationoftheseresources. Thetechnologies, includingPeer-to-Peer, GridandInternet-based Virtual Computing Environment (iVCE), which are designed to exploit differenttypes of idle resources such as CPU cycles, memory, storage, access bandwidth and more,are all attempts in this area. A fundamental difference of these applications comparedwith traditional distributed systems is the fact that the resource provision is based on thedecisions of autonomous users, called autonomous resource aggregation.At the same time, the autonomy of users also leads to various bad behaviors, suchas self-interested behavior and providing unreliable or malicious services, which brings agreatchallengetosuchasystem. Soreputationmechanismshaverecentlyattractedsignif-icant attentions as a way for diminishing the influence of untrusted behaviors to maintainthe efficient system operation. However, despite of the success in reputation mechanisms,there are still a lot of questions needed to be solved, say, the lack of efficient models toanalyze users’ sharing behavior and the incentive effect of reputation mechanisms, and thelack of quantitatively analyzing method on different reputation models.This dissertation launches the research on the domain of autonomous resource ag-gregation and the problems of reputation mechanism analysis. And the main work in thisdissertation can be briefly addressed as follows.Enlightened by the inconsistency between the theory results in the analysis on users’sharing behavior and the observation results of real systems, taking the dynamicsof sharing environment and user interaction under consideration, we propose ananalyzing framework of game theory for users’ sharing behavior in autonomous re-source aggregation. This framework extracts the service model from the concrete game process, whichensuresitsuniversality; meanwhile, itadoptstheproportion ofsharing users as the state variable of game equilibrium, which efficiently simplifiesthe process of equilibrium analysis.In order to explore the intrinsic factors of sharing behavior in autonomous resourceaggregation, we introduce a random service model into the dynamics sharing gamein the context of no external incentives. The theoretical research shows that with-out external incentives, there still exists a proportion of sharing users in the systemundercertainconditions,whichprovesthefeasibilityofautonomousresourceaggre-gation; However, the experimental results demonstrate that under the assumption ofbounded rationality, the motivation of sharing may be too small and be eclipsed bythe learning noise, and the system cannot maintain the stable aggregation or shar-ing level, which indicates the necessity of incentive mechanisms for autonomousresource aggregation.In order to verify the incentive effect to reputation mechanisms in resource aggrega-tion, we bring in a special reputation mechanism (phase reputation mechanism) asthe new service model of the game above to display the incentive effect of reputa-tion mechanism. The theoretical results point out that if the ratio of real sharing costand service benefit is less than0.42, all users would contribute their resources to thesystem, and it sufficiently verifies the effect of the reputation mechanism; and thesimulation under the circumstance of bounded rationality and incomplete informa-tion gives the same conclusion as the theoretical research. Moreover, we extend theabove research, and raise a method to incentive effect analysis on general reputationmechanisms. Besides, we identify and categorize the game model according to theequilibria distribution, which provides supports for further studies.As an indirect reciprocity scheme, a reputation mechanism provides suitable incen-tives for resource contribution with the promise of future rewards and the threat offuture punishments. Therefore, users’ sharing behavior is directly influenced by thetrustworthiness of the promise or the threat which is closely related to the amountof available sharing resources in system. So, the influence of the initial amount ofavailable sharing resources on reputation mechanisms is studied in this dissertation.Our theoretical results show that there is a network effect in autonomous resource aggregation with above phase reputation mechanism: a user’s maximum toleranceofsharingcostisincreasingwithrespecttothesystemscaleorproportionofsharingusers. Furthermore, we explore the problem of critical mass or start-up and identifyan important characteristic of autonomous resource aggregation that the incentiveeffect of a reputation mechanism is not only determined by the mechanism but alsothe initial level of sharing resources. If the level is below a certain value; even thereexists a reputation mechanism; the system still cannot survive because of the lackof sharing resources.Aiming to solve the problem of lacking a generally and quantitatively compara-tive analysis on existing reputation models, we propose an ordering-based methodto evaluate reputation models. It ignores the reputation values and translates themodel evaluation to the comparison between two ranking orderings. Moreover, weintroduce two evaluation criteria for a reputation model-precision and robustness,and give two metric-matching degree and hitting rate－－for them. Evaluatingon our reputation model P-REP and relevant models by our ordering-based methodand comparing the evaluation results with the general evaluation framework-QTM,the experimental results suggest that the ordering-based method can evaluate differ-ent reputation models in multi-dimension and provides a possible way to evaluatedifferent reputation models.In general, this dissertation focuses on the analyzing techniques for reputation mech-anisms, and the purposes of this dissertation are twofold: first, it aims at the analysis onincentive effect on reputation mechanisms; secondly, it looks at the evaluation of repu-tation models and proposes an ordering-based approach, which provides a possible con-sistent way to evaluate different reputation models. The contributions above will offer atheoretical basis and technical guidance to the research on reputation mechanisms underthe setting of autonomous resource aggregation over Internet.