The human in the loop of a delegated agent: the theory of adjustable social autonomy

Abstract

In this paper, we refer to social autonomy in a collaborative relationship among agents based on delegation and help. We address the problem of adjustable autonomy, i.e., we discuss when modifying the assigned/received delegation entails a corresponding enlargement or restriction of autonomy and at which level. We stress in particular the role played in autonomy by: 1) the degree of "openness" of delegation (execution autonomy), 2) the allowed initiative in (re)starting negotiation (meta-autonomy), 3) the degree and kind of control (feedback+intervention), and 4) the strength of delegation with respect to interaction. We show how the adjustability of delegation and autonomy is actually "bilateral," because not only the user (delegator, trustor, client) can adjust the autonomy of the agent (delegee, trustee, contractor), but the agent can also have (cooperative) reasons for and the ability to change the received delegation and modify its own autonomy in it. Adjustment is also "bidirectional" (from more autonomy to less autonomy, or vice versa), and multidimensional. Finally, we analyze some reasons for modifying the assigned autonomy and show how the adjustment of autonomy depends-on the delegator's side-on a crisis of trust ; vice versa, the delegee's adjustment of its own autonomy depends on some disagreement about the trust received from the delegator, and, in particular, either a higher or lower confidence in itself or in external circumstances. Some preliminary hints about necessary protocols for adjusting the interaction with agents are provided. This work is neither directly prescriptive nor simply descriptive. It is aimed at providing a theoretical framework, i.e., the conceptual instruments necessary for analyzing and understanding interaction with autonomous entities.

Examining Trust, Forgiveness and Regret as Computational Concepts

“…While to our knowledge there is no work combining the three phenomena to further the development of truly social agents, the effect obtained, at least when we consider autonomous agents working in specific environments for humans, is similar to the concept of Adjustable Autonomy, in which humans retain a meta-level control over their more-or-less autonomous agents [28,41,29]. In this work, however, the control is given back to the agents in order to allow them to adjust the amount of leeway (or autonomy) other agents have with their resources, as well as the human.…”

Influence-Based Autonomy Levels in Agent Decision-Making

“…It is widely recognized that the ability of an agent to make decisions autonomously is a strong feature [3]. In some settings it is inevitable to allow autonomous decisions, for example if communication with others fails, or if there is no time to negotiate actions.…”

“…An agent can not fully control its own level of autonomy. An agent's autonomy is bilateral adjustable [3], which means that the level of autonomy can be adjusted by the agent itself as well as by other agents. For example, an agent asking for help instructions chooses to consider options generated by others and therefore it becomes dependent of other agents and lowers its degree of autonomy while pursuing its goal.…”

“…There are three types of influence between agents to be distinguished, [3], [8]: influence by environmental modification, influence by belief alteration and influence on goal/task determination.…”

“…In Sect. 3 we present a deliberation loop for autonomy-aware BDI agents. We show how different levels of autonomy can be defined and how they relate to the agent decision model.…”