Designing collective decision-making dynamics for multi-agent systems, with inspiration from honeybees

Abstract

For many multi-agent systems, collective decision-making among alternatives is a crucial task. A group of agents may be required to collectively decide on their next action, and may face limitations on their sensing, communication and computational abilities. A swarm of honeybees choosing a new nest-site faces these challenges, and has been shown to reliably make decisions with accuracy, efficiency and adaptability. The honeybee decision-making dynamics can be modelled by a pitchfork bifurcation, a nonlinear phenomenon that is ubiquitous in animal decision-making.

We describe and analyse a model for collective decision-making that possesses a pitchfork bifurcation. The model allows us to leverage the characteristics of the honeybee dynamics for application in multi-agent network systems and to extend the capabilities of our decision-making dynamics beyond those of the biological system.

Using tools from nonlinear analysis, we show that our model retains some important characteristics of the honeybee decision-making dynamics, and we examine the impact of system and environmental parameters on the behaviour of the model. We derive an extension to an existing centrality measure to describe the relative influence of each agent, and to show how agent preferences can lead to bias in the network. We design decentralised, adaptive feedback dynamics on a parameter of the model, which ensure that a decision is made. We discuss how this system parameter, which quantifies how much each agent is influenced by its neighbours, provides an intuitive mechanism to involve a human operator in the decision-making. We continue this discussion as we implement our model with a simple robotic system.

Throughout this thesis, we discuss the trade-off in the design of decision-making dynamics between systems that are robust to unwanted disturbances, but are also sensitive to the values of important system parameters. We show how dynamics modelled by a pitchfork bifurcation exhibit hypersensitivity close to the bifurcation point, and hyperrobustness far away from it.