Planning and Consensus among Autonomous Agents

Preface This dissertation describes research in the eld of Distributed Artiicial Intelligence (DAI). DAI is the subbeld of Artiicial Intelligence that investigates the behavior of societies of artiicial agents (high-level, interacting computer programs). The speciic research of the dissertation focuses on multi-agent planning under various models of interaction. Within the eld of DAI there is a distinction between two research agendas, Distributed Problem Solving and Multi-Agent Systems. Within the rst paradigm, agents can be assumed to be created by the same designer or body, and thus, work together to solve common goals. Using economic terms, the agents have a common preference proole that yields an identical utility function. The common goal(s) may be decomposed into subgoals that are allocated to diierent agents in the group. Coordination is needed to allow eecient distributed activity towards the achievement of the global goal. While pursuing their tasks, agents may communicate, share their knowledge, and benevolently help one another. In the Multi-Agent Systems paradigm, on the other hand, there is not assumed to be any one designer who can impose any speciic kind of behavior directly on all the agents. Agents, designed by diierent people, may have their own private goals and act sellshly towards the achievement of these goals. Therefore each agent may have its private proole of preferences, and a distinct individual utility function. This dissertation deals with both Distributed Problem Solving and Multi-Agent Systems issues. The subject of multi-agent planning, a major focus of the dissertation, has been of continuing concern in DAI. Some of the earliest work in the eld directly addressed the question of how to get groups of agents to carry out coordinated, coherent activity. These questions have occupied researchers both in the Distributed Problem Solving sub-area of DAI as well as those working on Multi-Agent Systems. The term multi-agent planning has been used to describe both \planning for multiple agents" (where the planning process itself may be centralized), and \planning by multiple agents" (where the planning process is distributed among the agents). Chapter 1 discuss models of interaction, presents our approach, and discusses related work. Chapter 2 introduces a novel combination of the two distinct multi-agent planning approaches mentioned above. We analyze a particular model of control among intelligent agents, that of iv non-absolute control. Non-absolute control involves a \supervisor" agent that issues orders to a \subordinate" agent. Both agents operate with essentially the same goals. The …