Coalition Formation through Motivation and Trust

Cooperation is the fundamental underpinning of multi-agent systems, allowing agents to interact to achieve their goals. Where agents are self-interested, or potentially unreliable, there must be appropriate mechanisms to cope with the uncertainty that arises. In particular, agents must manage the risk associated with interacting with others who have different objectives, or who may fail to fulfil their commitments. Previous work has utilised the notions of motivation and trust in engendering successful cooperation between self-interested agents. Motivations provide a means for representing and reasoning about agents' overall objectives, and trust offers a mechanism for modelling and reasoning about reliability, honesty, veracity and so forth. This paper extends that work to address some of its limitations. In particular, we introduce the concept of a clan: a group of agents who trust each other and have similar objectives. Clan members treat each other favourably when making private decisions about cooperation, in order to gain mutual benefit. We describe mechanisms for agents to form, maintain, and dissolve clans in accordance with their self-interested nature, along with giving details of how clan membership influences individual decision making. Finally, through some simulation experiments we illustrate the effectiveness of clan formation in addressing some of the inherent problems with cooperation among self-interested agents.

Quantified Interference: Information Theory and Information Flow

The paper investigates which of Shannon’s measures (entropy, conditional entropy, mutual information) is the right one for the task of quantifying information flow in a programming language. We examine earlier relevant contributions from Denning, McLean and Gray and we propose and motivate a specific quantitative definition of information flow. We prove results relating equivalence relations, interference of program variables, independence of random variables and the flow of confidential information. Finally, we show how, in our setting, Shannon’s Perfect Secrecy theorem provides a sufficient condition to determine whether a program leaks confidential information.