Stereotype Reputation with Limited Observability

Abstract

Assessing trust and reputation is essential in multi-agent systems where agents must decide who to interact with. Assessment typically relies on the direct experience of a trustor with a trustee agent, or on information from witnesses. Where direct or witness information is unavailable, such as when agent turnover is high, stereotypes learned from common traits and behaviour can provide this information. Such traits may be only partially or subjectively observed, with witnesses not observing traits of some trustees or interpreting their observations differently. Existing stereotype-based techniques are unable to account for such partial observability and subjectivity. In this paper we propose a method for extracting information from witness observations that enables stereotypes to be applied in partially and subjectively observable dynamic environments. Specifically, we present a mechanism for learning translations between observations made by trustor and witness agents with subjective interpretations of traits. We show through simulations that such translation is necessary for reliable reputation assessments in dynamic environments with partial and subjective observability.

Reputation assessment: a review and unifying abstraction

“…To evaluate models in our abstraction we use a simulated marketplace based on that presented by Burnett et al [6,7], and subsequently adopted by Ş ensoy et al [35] and Taylor et al [36,37] [−0.5, −0.4, . .…”

“…To evaluate models in our abstraction we use a simulated marketplace based on that presented by Burnett et al (2010Burnett et al ( , 2013, and subsequently adopted by Şensoy et al (2016) and Taylor et al (2017aTaylor et al ( , 2017b. The simulation and models discussed in this evaluation are available for download as open source software 1 .…”

“…This output can then be combined with more simple aggregations, as in Burnett et al [6,7], Ş ensoy et al [35], and Taylor et al [36,37], or the model used directly as the opinion. Burnett et al [6,7] extend BRS to include stereotypes, and produce an opinion,…”

“…all agents represent context in the same way, but this does not need to be true in general [36]. The results for using known context information are shown in Table 7.…”

“…(2016), and Taylor et al . (2017a, 2017b), or the model used directly as the opinion. Burnett et al .…”