Methodologies for scalable and responsive swarms and networked systems

Student thesis: Doctoral Thesis

Abstract

Currently, engineers can only aspire to match nature’s capabilities where the same control mechanism can scale from one hundred starlings to five thousand whilst maintaining properties such as cohesion and rapid response to predation. These aspirations have led to engineered swarms that rely on remote or human control to assemble formations, which therefore do not require leaders within the network that can communicate using inter-agent connections. As a consequence, it is primarily in theory where network leadership is considered, usually through the exploitation of a standard set of centrality metrics - such as PageRank, degree centrality or betweenness - ubiquitous across network science.This dissertation attempts to lay the foundations for large, mobile, multiagents warms through a scalable control approach. This mechanism can achieve similar capabilities to that of a starling flock but with greater control over the swarm’s movement. These criteria are met by defining control rules for the environment, using artificial kinematic fields, rather than for the individual; enabling agents to join or leave the swarm without a breakdown in cohesion or responsiveness - just like starlings. The functionality of this method, and the potential for swarm based applications, is demonstrated through a remote inspection case study. This control approach contributes to our rapidly increasing ability to create networked systems, but our understanding of how to control such complexity is not advancing as fast.The identification of optimal network leaders is a significant step towards achieving a responsive, controllable, swarm. By not constraining the problem to a set number of leaders, the solution space - for anything other than small networks - is too vast to attempt an exhaustive search. The remedy for this leader selection problem has, so far, eluded researchers.Eigenvectors are presented here as key to solving this problem for determining optimal leadership when considering fast convergence to consensus. The semi-analytical algorithms, developed herein, have O(n3) time complexity and are found to perform as well as numerical optimisers with significantly greater complexity, O(n4). Eigenvectors map the dynamic response of a network where they are found to expose the most responsive communities that form in the wake of external perturbations. Comparisons, of eigenvector-based methods, with information flow simulations illustrate that analytical models can, in a computationally efficient manner, cast light on the interplay between leadership and topology. Whilst providing a greater understanding of the effectiveness and function of nature’s networked systems, including the vastly complicated and responsive network of the human brain.
Date of Award1 Apr 2017
LanguageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsEPSRC (Engineering and Physical Sciences Research Council)
SupervisorMalcolm Macdonald (Supervisor)

Cite this