Algorithmic transparency in path planning

Abstract

Computer algorithms facilitate increased automation in various human-centered work areas to improve operational safety and efficiency. Algorithmic transparency is considered essential for human operators, policy makers and system developers, as it allows them to understand the capabilities and limitations of an algorithm. In this research, we focus on path-planning algorithms and propose a purely visual approach to achieve their transparency. This approach extracts and portrays information directly from the algorithms, aiming to visually reveal their inner workings. Benchmark tests indicate that extracting information from path-planning algorithms may significantly slow them down. For time-constrained operations, it is recommended to store only the necessary data during the pathfinding process and perform information extraction afterwards. Based on theories from cognitive engineering, six transparency levels were designed to chunk meaningful information pertaining path-planning algorithms. A user study among non-experts (N=40) was then conducted to evaluate the impact of visual algorithmic transparency on human understanding. The results suggest that increased transparency levels allow non-experts to more correctly and confidently understand the details of a path-planning algorithm. However, it is also found that certain transparency levels can lead to confusion, especially when the algorithm behaves in a way contrary to human expectations. This study further reveals that, given the same level of transparency, sampling-based algorithms may be easier to comprehend than graph-based algorithms. This research can serve as a reference for how to achieve transparency in path-planning-related applications and how to hierarchically portray and organize transparency information.