Designing agents for systems with adjustable autonomy

摘要

Intelligent agents are starting to assume more responsibility in more critical tasks. Agents do not always work as required, however in some systems it is infeasible to simply \\stop " an agent when its behavior is unsatisfactory. Instead it is desirable to reduce the agent's autonomy and give control of the unsatisfactory aspects of the agents task to a more competent entity. Adjustable Autonomy (AA) is a recent idea that means that the autonomy of agents and humans in a system can vary dynamically. In this paper we look at the relationship between the design of agents and the design of AA for an intelligent system. 1 Intelligent Agents Many of the recent, exciting developments in arti cial intelligence (AI) have centered around the concept of an agent. An agent is an autonomous entity that senses its environment and acts intelligently and pro-actively towards its goals[17, 2]. An important characteristic of an agent is that it has the ability to take actions that a ect its environment[7]. Some agents sense and act in purely software environments (e.g. an operating system monitoring agent[16]), while others have aphysical embodiment and inhabit a physical environment (e.g. a museum tour guide robot[3]). Because an agent can act, it can be assigned tasks that can potentially be done more quickly, e ciently, cheaply or safely than a human can do them. Thus humans are freed from menial, dangerous and/or boring tasks. Despite the long list of successful agent applications, it is likely we have only scratched the surface of the possibilities intelligent agents have to change the way we live our lives[6, 8]. In the future, autonomous agents will take on more complex tasks and act more intelligently and more decisively. An intelligent system is one consisting of intelligent agents and, possibly, humans and/or other conventional software. Generally in an intelligent system the assignment of responsibility and authority, i.e. autonomy, is either xed or switches between a small number of xed con gurations. In an intelligent system with AA the system can exibly con gure the assignment of autonomy between the people and agents to best t the situation. 1