Présentation sur YouTube.

Computerized decision support systems (DSS) play an increasingly important role across a wide range of applications, from surveillance and reconnaissance for defense to risk assessment for business decisions and medical diagnostics. In all these applications, a human is ultimately responsible for the final decisions and must therefore remain "in the loop." Nevertheless, automation can (hopefully) help reduce the operator's cognitive workload, allowing them to concentrate on the most demanding tasks. The design of these DSS poses significant challenges for shared decision-making between humans and automation. One such challenge is determining when the automation should refer a task or decision to the operator, or instead, handle the task itself to free the operator for more difficult tasks. To illustrate this issue, we consider a model for decision referrals in human-automation teams jointly performing binary classification tasks, where human performance declines as workload increases. We provide an optimal policy that allows the automation to determine which and how many tasks to refer to the human. Implementing this policy requires only a characterization of the human’s performance in terms of error probabilities, rather than a detailed model of their perception and decision-making strategy. This can be measured empirically and connects to the application of signal detection theory in psychology. We also extend the model to account for the fact that the human operator may only follow the DSS recommendations probabilistically, based on their (time-varying) level of trust in the system. For this model, we provide an efficient algorithm to compute the referral policy that minimizes cost at each period and characterize conditions under which this myopic policy is globally optimal.