Autonomous Agents: Characterization and Requirements

A fresh assessment of what autonomous agents are or may aspire to be is presented in the paper. We dismiss the traditional dichotomy between reactive and symbolic architectures. A classiication in terms of the amount of knowledge embedded in the system is made, instead, running from regulatory to planning to adaptive agents. Regulatory agents are implementable as control systems or automata, and represent the ideal case in which the agent has all the knowledge it needs. We nd that each type of agent is best suited for a particular kind of behavior. The diierent kinds of behavior form a hierarchy in terms of frequency of occurrence and response time, and do not gracefully reduce to each other. This suggests implementing agents as hierarchies of simpler agents, each specialized in the kind of behavior it implements. In this scheme, agents at one level act through agents at the next lower level. We introduce the concept of drive to address the issue of where an agent's goals come from. A discussion is made of common pitfalls and suggestions for the design of planning agents. Finally, a parallel architecture that embodies many of the ideas presented is sketched and discussed. This architecture solves the two major problems present in Georgee's earlier architecture, PRS: strictly sequential execution, and inability to learn.