Decentralized control of multiple collaborating agents

Abstract

This thesis investigates the effect of various factors on multi-agent collaboration in a simulated fire-fighting domain. A simulator was written that models fires and fire-fighting agents in an area of a few square city blocks. Different scenarios were constructed to test the effects of limiting the agents' knowledge and to compare the effects of coordinating the agents through a single entity with the results obtained through independent decision making. This research demonstrates that the amount of information available to each agent and the agents' ability to act on this information are typically much more important factors than the use of a complex planning mechanism; as long as the agents are aware of each other and take minimal steps to coordinate their actions they are able to achieve results that are nearly as good as those achieved by much more complicated algorithms.