Modeling Assistant's Autonomy Constraints as a Means for Improving Autonomous Assistant-Agent Design

Abstract

In this paper we introduce and experimentally evaluate a new suboptimal decision-making design to be used by autonomous agents acting on behalf of a user in repeated tasks, whenever the agent's autonomy level is continuously controlled by the user. This mode of operation is common and can be found whenever user's perception of the agent's competence is affected by the nature of the outcomes resulting from the agent's decisions rather than the optimality of the decisions made, e.g., in spam filtering, CV filtering, poker agents, and robotic vacuum cleaners as well as in newly arriving systems such as autonomous cars. Our proposed design relies on choosing the action that offers the best tradeoff between decision optimality and the influence over future allowed autonomy, where the latter is predicted using standard machine learning techniques. The design is found to be highly effective compared to following the theoreticoptimal decision rule, over various measures, through extensive experimentation with a virtual investment agent, making virtual investments on behalf of 679 subjects using Amazon Mechanical Turk.