Identifying Critical Contextual Design Cues Through a Machine Learning Approach

Abstract

Introduction The development of autonomous technologies that take on safety critical functions, such as driverless cars or surgical robots, can potentially reduce accidents and errors and improve productivity. However, while autonomous systems show promise for enhancing safety and productivity, previous research in human‐automation interaction has demonstrated that adding automation does not necessarily guarantee increased system effectiveness or safety. Often, automating a task within a larger system modifies the task by transferring the operator’s workload from one physical or cognitive resource to another, thereby changing the task rather than improving it (Bainbridge, 1987). Poorly designed automation that is not understood by operators often causes human error and reduces system effectiveness due to “clumsy” implementations (Lee & Morgan, 1994). As these systems proliferate, there is an increasing need to understand how such systems should be designed to promote effective interactions between one or more humans working with or around autonomous systems. This is especially true for safety critical settings like operators of such systems including medical systems, factory workers engaged in tasks with or near automation, or pedestrians and bicyclists operating in the same environment with driverless cars. Given the importance of promoting effective and safe interactions between human users and autonomous systems, designers of these systems need tools that allow them to determine not just which designs are effective, but how such systems fare under different contexts. Indeed, the ability of autonomous systems to account for context and changing environments is a significant hurdle that limit applications (Daily, Medasani, Behringer, & Trivedi, 2017; Marcus, 2018; Srinivasan, 2016). One often overlooked source of potential contextual design cues in autonomous systems is the vast amount of data generated by these systems, including those of user interactions. While machine learning approaches to data analysis are often touted for their importance in the operation of these cars, they can also be harnessed for understanding the impact of context, particularly when attempting to determine the effectiveness and safety of a design choice. This paper will discuss the importance of contextual cues in design and demonstrate how machine learning method can be adapted to determine the effectiveness of design features in an autonomous system.