Achieving accurate opinion consensus in large multi-agent systems

Abstract

Modern communication technologies offer the means to share information within decentralised, large and complex networks of agents. A significant number of examples of peer-to-peer interactions can be found in domains such as sensor networks, social web communities and file-sharing networks. Nevertheless, the development of decentralised systems still presents new challenges for sharing uncertain and conflicting information in large communities of agents. In particular, the problem of forming opinion consensus supported by most of the observations distributed in a large system, is still challenging.
To date, this problem has been approached from two perspectives: (i) on a system-level, by analysing the complex processes of opinion sharing in order to determine which system parameters result in higher performance; and (ii) from the perspective of individual agents, by designing algorithms for interactively reaching agreements on the correct opinion or for reasoning about the accuracy of a received opinion by its additional annotation.
However, both of these approaches have significant weaknesses. The first requires centralised control and perfect knowledge about the configuration of the system in order to simulate it, which are unlikely to be available for large decentralised systems. Whereas, the latter algorithms introduce a significant communication overhead, whilst in many cases the capabilities of the agents are restricted and communication strictly limited. Therefore, there is a need to fill the gap between these two approaches by addressing the problem of improving the accuracy of consensus in a decentralised fashion with minimal communication expenses.
With this motivation, in this thesis we focus on the problem of improving the accuracy of consensus in large, complex networks of agents. We consider challenging settings in which communication is strictly limited to the sharing of opinions, which are subjective statements about the correct state of the subject of common interest. These opinions are dynamically introduced by a small number of sensing agents which have low accuracy, and thus the correct opinion just slightly prevails in the readings. In order to form the accurate consensus, the agents have to aggregate opinions from a number of sensing agents which, however, they are very rarely in direct connection with.
Against this background, we focus on improving the accuracy of consensus and develop a solution for decentralised opinion aggregation. We build our work on recent research which suggests that large networked systems exhibit a mode of collective behaviour in which the accuracy is improved. We extend this research and offer a novel opinion sharing model, which is the first to quantify the impact of collective behaviour on the accuracy of consensus. By investigating the properties of our model, we show that within a narrow range of parameters the accuracy of consensus is significantly improved in comparison to the accuracy of a single sensing agent. However, we show that such critical parameters cannot be predicted since they are highly dependent on the system configuration.
To address this problem, we develop the Autonomous Adaptive Tuning (AAT) algorithm, which controls the parameters of each agent individually and gradually tunes the system into the critical mode of collective behaviour. AAT is the first decentralised algorithm which improves accuracy in settings where communication is strictly limited to opinion sharing. As a result of applying AAT, 80-90% of the agents in a large system form the correct opinion, in contrast to 60-75% for the state-of-the-art message-passing algorithm proposed for these settings, known as DACOR. Additionally, we test other research requirements by evaluating teams with different sizes and network topologies, and thereby demonstrate that AAT is both scalable and adaptive. Finally, we showed that AAT is highly robust since it significantly improves the accuracy of consensus even when only being deployed in 10% of the agents in a large heterogeneous system.
However, AAT is designed for settings in which agents do not differentiate their opinion sources, whilst in many other opinion sharing scenarios agents can learn who their sources are. Therefore, we design the Individual Weights Tuning (IWT) algorithm, which can benefit from such additional information. IWT is the first behavioural algorithm that differentiates between the peers of an agent in solving the problem of improving the accuracy of consensus. Agents running IWT attribute higher weights to opinions from peers which deliver the most surprising opinions. Crucially, by incorporating information about the source of an opinion, IWT outperforms AAT for systems with dense communication networks. Considering that IWT has higher computational cost than AAT, we conclude that IWT is more beneficial to use in dense networks while AAT delivers a similar level of accuracy improvement in sparse networks, but with a lower computational cost.

More information

Identifiers

Catalogue record

Export record