Enhancing Swarm Efficiency: Exploration-Exploitation Tuning for Fast Target Tracking

Abstract

Is it possible for multi-robot systems to efficiently search and track one or more fast-moving targets? While individual robots on their own are simply unable to mount a proper pursuit of a target moving faster than they can, leveraging swarm intelligence is key to achieving this goal. In this thesis, the question of how to optimize a swarm’s design to achieve such a task efficiently is raised and answered through the lens of balancing exploration— i.e., the gathering of information—and exploitation—i.e., the use of said information. Through the combination of simple movement rules, namely repulsion between agents, promoting exploration, and attraction towards a target, which generates exploitation, we generated the decentralized operation of swarms of agents. To communicate, these agents rely on a k-nearest neighbors network, which can be adjusted to calibrate the swarm’s exploration-exploitation balance. Other parameters, such as implementing short-term memory and the introduction of fast agents—i.e., heterogeneity—are also used to affect this balance with the ultimate goal of optimizing swarm performance. By adjusting the size of the search-space the swarm operates in, we vary its agent density, which is introduced as a novel method to study swarm behaviors and improve the optimization process. The identification of density phases characterized by a swarm’s performance reveals the mechanisms behind the previously discovered critical minimum and maximum densities. By exploring how to shift the ’transition’ phase, where performance is maximized for the number of agents present, a framework is provided to optimize swarm performance and harness swarm intelligence. Furthermore, using metrics to describe the swarm’s exploration-exploitation balance allows an optimum level of connectivity k to be found for each specific tracking scenario, revealing a path towards the design of swarms that perform over wider ranges of conditions. The replacement of agents by faster ones was also shown to be universally beneficial, although extra care must be given as the exploration-exploitation balance is affected. Thus, the development of truly adaptive strategies able to tune this balance according to local conditions, as was manually done throughout this work, is discussed as the next step toward the development of successful swarms.