Large-scale, dynamic and distributed multi-agent coordination for real-time systems

Abstract

The Coalition Formation with Spatial and Temporal constraints Problem (CFSTP) is designed to characterise scenarios at the intersection between task allocation and coalition formation. In this model, tens of heterogeneous agents are deployed over kilometre-wide areas to carry out thousands of tasks, each with its deadline and workload. To maximise the number of tasks completed, the agents need to cooperate by forming, disbanding and reforming coalitions. In this thesis, we start with an in-depth analysis of Coalition Formation with Look-Ahead (CFLA), the state-of-the-art CFSTP algorithm. We outline its main limitations, based on which we derive an extension called CFLA2. We show that we cannot eliminate the limitations of CFLA in CFLA2, hence we also develop a novel algorithm called Cluster-based Task Scheduling (CTS), which is the first to be simultaneously anytime, efficient and with convergence guarantee. We empirically demonstrate the superiority of CTS over CFLA and CFLA2, and propose S-CTS, a simplified but parallel and more efficient variant. In problems generated by the RoboCup Rescue Simulation, S-CTS can compete with the high-performance Binary Max-Sum and DSA algorithms, while being up to two orders of magnitude faster. We then propose a minimal mathematical program of the CFSTP, and reduce it to a Dynamic and Distributed Constraint Optimisation Problem, based on which we design D-CTS, a distributed version of CTS. We create a test framework that simulates the mobilisation of firefighters, which we use to show the effectiveness of D-CTS in large-scale and dynamic environments. Finally, to characterise scenarios in which the faster the tasks are solved, the greater the benefits, we propose the Multi-Agent Routing and Scheduling through Coalition Formation problem (MARSC), a generalisation of both the CFSTP and the important Team Orienteering Problem with Time Windows. We formulate a binary integer program and propose Anytime, exact and parallel Node Traversal (ANT), the first algorithm of its kind for both the MARSC and the CFSTP. Moreover, we define an approximate variant called ANT-ε. Both algorithms are validated in our realistic test framework, using as baselines an extended version of CTS, and an implementation of the Earliest Deadline First technique, which is typically used in real-time systems.

More information

Identifiers

ORCID for Luca Capezzuto: orcid.org/0000-0003-4404-0998

ORCID for Sarvapali Ramchurn: orcid.org/0000-0001-9686-4302

Catalogue record

Export record