Adjustable Autonomy: Controling Influences on Decision Making

Abstract

Due to technological developments we foresee future systems where groups of actors coordinate their actions in a dynamic manner to reach their goals. Our aim is to develop a reasoning model for artificial actors in such systems. Starting point is the relation between autonomy of individuals and coordination of group behavior. We adopt the agent paradigm as basis for the actors. Although autonomy is a key concept in agent research, there is no common definition of agent autonomy and adjustable autonomy. Our agents should be autonomous such that they can be held responsible for their actions. Furthermore, they should be able to work together following different types of coordination in order to achieve dynamic coordination. In this research, we define being autonomous as having control over external influences on the decision-making process. Adjustable autonomy in our context means dynamically dealing with external influences on the decision-making process based on internal motivations. An agent controls to what extend it is being influenced by the environment and by other agents. Agents achieve coordination by allowing influences of others on their decision-making process. We present an abstract reasoning model that facilitates agent autonomy. The two main components of the reasoning model are influence control and decision-making. In the component for influence control the agent deals with new observations and messages. We propose to use reasoning rules to process those external events. The reasoning rules specify the effects of external events on the beliefs and goals of the agent. The beliefs and goals, then, are used for the decision-making process. With this reasoning model we do a simulation experiment of a firefighter organization extinguishing several fires. The organizational goal can be reached via different types of coordination ranging from emergent to centralized coordination. The experiment shows that the perspective on autonomy as influence control provides a promising way to facilitate dynamic coordination. Our reasoning model separates event-processing and decision-making. This allows the development of domain-independent heuristics for event-processing. In this thesis we mention three heuristics: information relevance, organizational knowledge and trust. The first two are discussed in further detail. We propose an algorithm for relevance determination. We argue that potential benefits of using information relevance for influence control can be found in controling the decision quality and preventing information overload. Furthermore, we work out the use of organizational knowledge to process external events. We show how organizational norms can be translated into event-processing rules. We describe how our approach facilitates organizational dynamics. The modularity in our reasoning model ensures that the event-processing rules are explicitly defined. This allows for metareasoning about the event-processing rules. We present a metareasoning model, with which the agent can select and take up the desired attitude with respect to the environment and to other agents. This process gives the agent control over external influences, and therewith it meets the requirements for autonomy and adjustable autonomy. In a simulation experiment we implement a firefighter organization that exhibits dynamic coordination via self-adjustable autonomy of the firefighters.