Analysis and Prediction of Decision Making with Social Feedback

Abstract

Robots that work in conjunction with humans are becoming commonplace. Some are autonomous, operating without human input, but many require supervision or direct control. In this work we suggest using mixed teams for decision making when robots are faced with complex tasks and human input is beneficial. Evidence that robots can be effective as peers with humans is plentiful, but new tools are needed for designing such systems.

A formal, model-based analysis of decision making in teams that share social feedback is provided. We focus on the Two-Alternative, Forced-Choice (TAFC) task \cite{MoBe02} which has been widely-studied in experiments with human subjects. Decision makers in the TAFC task choose between two options, sequentially in time, and receive feedback on performance. This task is simple and relatively well-understood, but the predictive tools we develop and the principles that we uncover likely extend to different and more complex tasks.

Deterministic and stochastic decision-making strategies are considered. Our ability to predict behavior in teams with social feedback relies on our analysis of a stochastic soft-max choice model where we reveal dependence of performance on parameters describing the task, the decision makers, and the social feedback. These tools can assist in the design of mixed teams. For example, it is possible to identify scenarios where robotic decision makers designed into a mixed team can significantly improve performance. Experiments are underway to test our hypotheses.

Relevant applications and methods of interaction are also of interest. We describe the development of a robotic testbed that supports real-time operation of multiple, robotic vehicles in a three-dimensional field. We have plans for experiments that study mixed teams working with physical robots in our testbed. By considering real-world constraints in that setting, a comprehensive, and realistic app