University of Southampton Research Repository

FACULTY OF ENGINEERING AND APPLIED SCIENCE SCHOOL OF ELECTRONICS AND COMPUTER SCIENCE Doctor of Philosophy by Sarvapali Dyanand Ramchurn In this thesis, we propose a panoply of tools and techniques to manage inter-agent dependencies in open, distributed multi-agent systems that have significant degrees of uncertainty. In particular, we focus on situations in which agents are involved in repeated interactions where they need to negotiate to resolve conflicts that may arise between them. To this end, we endow agents with decision making models that exploit the notion of trust and use persuasive techniques during the negotiation process to reduce the level of uncertainty and achieve better deals in the long run. Firstly, we develop and evaluate a new trust model (called CREDIT) that allows agents to measure the degree of trust they should place in their opponents. This model reduces the uncertainty that agents have about their opponents’ reliability. Thus, over repeated interactions, CREDIT enables agents to model their opponents’ reliability using probabilistic techniques and a fuzzy reasoning mechanism that allows the combination of measures based on reputation (indirect interactions) and confidence (direct interactions). In so doing, CREDIT takes a wider range of behaviour-influencing factors into account than existing models, including the norms of the agents and the institution within which transactions occur. We then explore a novel application of trust models by showing how the measures developed in CREDIT ca be applied negotiations in multiple encounters. Specifically we show that agents that use CREDIT are able to avoid unreliable agents, both during the selection of interaction partners and during the negotiation process itself by using trust to adjust their negotiation stance. Also, we empirically show that agents are able to reach good deals with agents that are unreliable to some degree (rather than completely unreliable) and with those that try to strategically exploit their opponent. Secondly, having applied CREDIT to negotiations, we further extend the application of trust to reduce uncertainty about the reliability of agents in mechanism design (where the honesty of agents is elicited by the protocol). Thus, we develop Trust-Based Mechanism Design (TBMD) that allows agents using a trust model (such as CREDIT) to reach efficient agreements that choose the most reliable agents in the long run. In particular, we show that our mechanism enforces truth-telling from the agents (i.e. it is incentive