Optimal Explanation Generation using Attention Distribution Model

Abstract

With highly automated and Autonomous Vehicles (AVs) being one of the most prominent emerging technologies in the automotive industry, efforts to achieve SAE Level 3+ vehicles have skyrocketed in recent years. As new technologies emerge on a daily basis, these systems are becoming increasingly complex. To help people understand - and also accept - these new technologies, there is a need for explanation. There are three essential dimensions to designing explanations, namely content, frequency, and timing. Our goal is to develop an algorithm that optimises explanation in AVs. Most of the existing research focuses on the content of an explanation, whereas the fine-granularity of the frequency and timing of an explanation is relatively unexplored. Previous studies concerning "when to explain" have tended to make broad distinctions between explaining before, during or after an action is performed. For AVs, studies have shown that passengers prefer to receive an explanation before an autonomous action takes place. However, it seems likely that the acclimatisation that occurs through prolonged exposure to and use of a particular AV will reduce the need for explanation. As comprehension of explanations is workload-intensive, it is necessary to optimise both the frequency, i.e. skipping explanations when they are not helpful to reduce workload, and the precise point in time when an explanation is given, i.e. giving an explanation when it provides the maximum workload reduction. Extra mental workload for passengers can be caused by both giving and omitting an explanation. Every explanation that is presented requires cognitive processing in order to be understood, even if its content is considered to be redundant or if it will not be remembered by the addressee. On the other hand, skipping the explanation can cause the passenger to actively scan the environment for potential cues themselves, if necessary. Such an attention strategy would also impose a significant cognitive load on the passenger. In our work, to predict the mental workload of the passenger, we use the state-of-the-art attention model called SEEV (Salience, Effort, Expectancy, and Value). The SEEV model is dynamically used for forecasting the likelihood of the direction of attention. Our work aims to generate an optimally timed strategy for presenting an explanation. Using the SEEV model we build a probabilistic reactive game, i.e., 1.5-player game or Markov Decision Process, and we use reactive synthesis to generate an optimal reactive strategy for presenting an explanation that minimises workload.