Effective task allocation frameworks for large-scale multiple agent systems.
| dc.contributor.advisor | Shin, Hyo-Sang | |
| dc.contributor.advisor | Tsourdos, Antonios | |
| dc.contributor.author | Jang, Inmo | |
| dc.date.accessioned | 2023-02-07T19:46:56Z | |
| dc.date.available | 2023-02-07T19:46:56Z | |
| dc.date.issued | 2018-10 | |
| dc.description.abstract | This research aims to develop innovative and transformative decision-making frameworks that enable a large-scale multi-robot system, called robotic swarm, to autonomously address multi-robot task allocation problem: given a set of complicated tasks, requiring cooperation, how to partition themselves into subgroups (or called coalitions) and assign the subgroups to each task while maximising the system performance. The frameworks should be executable based on local information in a decentralised manner, operable for a wide range of the system size (i.e., scalable), predictable in terms of collective behaviours, adaptable to dynamic environments, operable asynchronously, and preferably able to accommodate heterogeneous agents. Firstly, for homogeneous robots, this thesis proposes two frameworks based on biological inspiration and game theories, respectively. The former, called LICA-MC (Markov-Chan-based approach under Local Information Consistency Assumption), is inspired by fish in nature: despite insufficient awareness of the entire group, they are well-coordinated by sensing social distances from neighbours. Analogously, each agent in the framework relies only on local information and requires its local consistency over neighbouring agents to adaptively generate the stochastic policy. This feature offers various advantages such as less inter-agent communication, a shorter timescale for using new information, and the potential to accommodate asynchronous behaviours of agents. We prove that the agents can converge to a desired collective status without resorting to any global information, while maintaining scalability, flexibility, and long-term system efficiency. Numerical experiments show that the framework is robust in a realistic environment where information sharing over agents is partially and temporarily disconnected. Furthermore, we explicitly present the design requirements to have all these advantages, and implementation examples concerning travelling costs minimisation, over-congestion avoidance, and quorum models, respectively. The game-theoretical framework, called GRAPE (GRoup Agent Partitioning and placing Event), regards each robot as a self-interested player attempting to join the most preferred coalition according to its individual preferences regarding the size of each coalition. We prove that selfish agents who have social inhibition can always converge to a Nash stable partition (i.e., a social agreement) within polynomial time under the proposed framework. The framework is executable based on local interactions with neighbour agents under a strongly-connected communication network and even in asynchronous environments. This study analyses an outcome’s minimum-guaranteed suboptimality, and additionally shows that at least 50% is guaranteed if social utilities are non-decreasing functions with respect to the number of co-working agents. Numerical experiments confirm that the framework is scalable, fast adaptable against dynamical environments, and robust even in a realistic situation where some of the agents temporarily halt operation during a mission. The two proposed frameworks are compared in the domain of division of labour. Empirical results show that LICA-MC provides excellent scalability with respect to the number of agents, whereas GRAPE has polynomial complexity but is more efficient in terms of convergence time (especially when accommodating a moderate number of robots) and total travelling costs. It also turns out that GRAPE is sensitive to traffic congestion, meanwhile LICA-MC suffers from slower robot speed. We discuss other implicit advantages of the frameworks such as mission suitability and additionally-builtin decision-making functions. Importantly, it is found that GRAPE has the potential to accommodate heterogeneous agents to some extent, which is not the case for LICA-MC. Accordingly, this study attempts to extend GRAPE to incorporate the heterogeneity of agents. Particularly, we consider the case where each task has its minimum workload requirement to be fulfilled by multiple agents and the agents have different work capacities and costs depending on the tasks. The objective is to find an assignment that minimises the total cost of assigned agents while satisfying the requirements. GRAPE cannot be directly used because of the heterogeneity, so we adopt tabu-learning heuristics where an agent penalises its previously chosen coalition whenever it changes decision: this variant is called T-GRAPE. We prove that, by doing so, a Nash stable partition is always guaranteed to be determined in a decentralised manner. Experi-mental results present the properties of the proposed approach regarding suboptimality and algorithmic complexity. Finally, the thesis addresses a more complex decision-making problem involving team formation, team-to-task assignment, agent-to-working-position selection, fair resource allocation concerning tasks’ minimum requirements for completion, and trajectory optimisation with collision avoidance. We propose an integrated framework that decouples the original problem into three subproblems (i.e., coalition formation, position allocation, and path planning) and deals with them sequentially by three respective modules. The coalition formation module based on T-GRAPE deals with a max-min problem, balancing the work resources of agents in proportion to the task’s requirements. We show that, given reasonable assumptions, the position allocation subproblem can be solved efficiently in terms of computational complexity. For the path planning, we utilise an MPC-SCP (Model Predictive Control and Sequential Convex Programming) approach that enables the agents to produce collision-free trajectories. As a proof of concept, we implement the framework into a cooperative stand-in jamming mission scenario using multiple UAVs. Numerical experiments suggest that the framework could be computationally feasible, fault-tolerant, and near-optimal. Comparison of the proposed frameworks for multi-robot task allocation is discussed in the last chapter regarding the desired features described at first (i.e., decentralisation, scalability, predictability, flexibility, asynchronisation, heterogeneity), along with future work and possible applications in other domains. | en_UK |
| dc.description.coursename | PhD in Aerospace | en_UK |
| dc.identifier.uri | https://dspace.lib.cranfield.ac.uk/handle/1826/19140 | |
| dc.language.iso | en | en_UK |
| dc.rights | © Cranfield University, 2015. All rights reserved. No part of this publication may be reproduced without the written permission of the copyright holder. | |
| dc.subject | multi-robot system | en_UK |
| dc.subject | robotic swarm | en_UK |
| dc.subject | asynchronously | en_UK |
| dc.subject | heterogeneous agents | en_UK |
| dc.subject | homogenous robots | en_UK |
| dc.subject | scaleable | en_UK |
| dc.title | Effective task allocation frameworks for large-scale multiple agent systems. | en_UK |
| dc.type | Thesis | en_UK |