Designing an Extended Set of Coordination Mechanisms for Multi-agent Systems
Well coordinated behaviors greatly improve intelligent entities overall performance, while inappropriate coordination results in reduced system efficiency, unfinished core tasks, misuse of key resources, and even system crashes. Thus, coordination is a central research task in Distributed Artificial Intelligence (DAI). This dissertation discusses our way of approaching the coordination problem, which is to develop an extended set of coordination mechanisms to manage the inter-dependencies between the activities among multiple agents. This dissertation presents four steps to tackle this problem. First, formal representation: articulating a domain-independent approach for specifying a set of coordination mechanisms and representing the characteristics of the agents’ tasks and actions with a highly expressive formalism. We present an Extended Hierarchical Task Network (EHTN) that is expressive enough to represent worth-oriented goals, contingencies, and the uncertainties that arise when task plans are in fact distributed over multiple agents by annotating tasks and actions quantitatively. We design and implement an extended set of GPGP (Generalized Partial Global Planning) coordination mechanisms, which are recast using this formalism. Second, mechanism development: constructing a large number of mechanisms to deal with the dependencies between multiple agents’ tasks. We have catalogued seven-teen GPGP coordination mechanisms for the enable relationship and discussed potential mechanisms for the other kinds of relationships, such as facilitate and hinder. Third, architectural support: updating the agents’ internal structure by inserting a novel coordination module between the planner and the scheduler. We strengthen our agents’ previous architecture by introducing a GPGP module, which uses the uncoordinated plans from the planner as input, applies appropriate GPGP coordination mechanisms to the uncoordinated plans, and generates coordinated plans to the scheduler for better schedules. Fourth, experimentation: applying these implemented coordination mechanisms to various domains to analyze their performances. Particularly, we chose an emergency medical service (EMS) domain to demonstrate the effectiveness of the extended set of GPGP coordination mechanisms. Based on the experimental results, qualitative analysis has been carried out and significant conclusions have been presented to guide agents in selecting the best mechanisms in various environments.