Trust assessment in the context of unrepresentative information

Preview

PDF WRAP_Theses_Player_2019.pdf - Submitted Version - Requires a PDF viewer. Download (8MB) | Preview

Request Changes to record.

Abstract

Trust and reputation algorithms are social methods, complementary to security protocols, that guide agents in multi-agent systems (MAS) in identifying trustworthy partners to communicate with. Agents need to interact to complete tasks, which requires delegating to an agent who has the time, resources or information to achieve it. Existing trust and reputation assessment methods can be accurate when they are learning from representative information, however, representative information rarely exists for all agents at all times. Improving trust mechanisms can benefit many open and distributed multi-agent applications. For example, distributing subtasks to trustworthy agents in pervasive computing or choosing who to share safe and high quality files with in a peer-to-peer network.

Trust and reputation algorithms use the outcomes from past interaction experiences with agents to assess their behaviour. Stereotype models supplement trust and reputation methods when there is a lack of direct interaction experiences by inferring the target will behave the same as agents who are observably similar. These mechanisms can be effective in MAS where behaviours and agents do not change, or change in a simplistic way, for example, if agents changed their behaviour at the same rate. In real-world networks, agents experience fluctuations in their location, resources, knowledge, availability, time and priorities. Existing work does not account for the resulting dynamic dynamic populations and dynamic agent behaviours. Additionally, trust, reputation and stereotype models encourage repeat interactions with the same subset of agents which increase the uncertainty about the behaviour of the rest of the agent population. In the long term, having a biased view of the population hinders the discovery of new and better interaction partners. The diversity of agents and environments across MAS means that rigid approaches of maintaining and using data keep outdated information in some situations and not enough data in others. A logical improvement is for agents to manage information flexibly and adapt to their situation.

In this thesis we present the following contributions. We propose a method to improve partner selection by making agents aware of a lack of diversity in their own knowledge and how to then make alternative behavioural assessments. We present methods for detecting dynamic behaviour in groups of agents, and give agents the statistical tools to decide which data are relevant. We introduce a data-free stereotype method to be used when there are no representative data for a data-driven behaviour assessment. Finally, we consider how agents can summarise agent behaviours to learn and exploit in depth behavioural patterns.

The work presented in this thesis is evaluated in a synthetic environment designed to mimic characteristics of real-world networks and are comparable to evaluation environments from prominent trust and stereotype literature. The results show our work improves agents’ average reward from interactions by selecting better partners. We show that the efficacy of our work is most noticeable in environments where agents have sparse data, because it improve agents’ trust assessments under uncertainty.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item