Scalable Multi-objective Task Allocation, Planning, and Learning in Multiagent Systems

Robotic decision making in multiagent systems is becoming more popular with rapidly advancing technology, where these systems are broadly applicable in many relevant domains including manufacturing, logistics, healthcare, transportation, and space exploration. For instance, multi-robot systems are used in smart manufacturing to automate complex precision tasks. These robots use decision-making algorithms to optimise the assembly process, reduce errors, and improve quality control This thesis focuses on a problem domain where a team of loosely coupled agents cooperate to complete a set of tasks in a stochastic environment. Each task is allocated to an agent who will compute, or learn a scheduler (i.e. policy) that controls its actions in a temporal order, often required to execute the task legally. Agents will also endeavour to meet a set of constraints for possible conflicting multiple-objectives. The well recognised problem of global task allocation and independent task completion manifests the intersection between multiagent planning, learning, task allocation and multi-objective optimisation. This thesis aims at developing novel formal, probabilistic verification and learning approaches. Additionally, this thesis contributes novel implementations, which can efficiently solve the common problem domain of this PhD study. The three key challenges addressed in this thesis are: (1) scaling to accommodate a significant number of agents and tasks, incorporating sufficiently rich models; (2) accuracy and effectiveness of task allocation solutions; and (3) the intuitive (human level) interpretation of solutions.