UC Irvine

In the fast-paced world of technology, life is transforming into a connected universe of Internet-of-Things (IoT), assisting human needs and creating a better world. United with AI, IoT technologies that once could only be dreamed of have become a reality. This work introduces AutoCoach, an IoT-based intelligent agent categorizing drivers into various driving personalities and providing individualized feedback through machine learning models. AutoCoach is intended for improving automobile drivers' performance by applying persuasive technology. System models like Advanced driver-assistance (ADAS) and some Usage-based-Insurance (UBI) aim to increase road safety. However, most prior models do not consider the differences between driving habits. When feedback is offered through those systems, not everyone may respond to them.

While many of those systems look into driving behavior as a single event (e.g., lane-change), a driving behavior may be judged based on the historical data to create a personalized experience leading to better feedback. We created AutoCoach, Cloud-based Android software that collects, analyzes, and learns from a driver's previous driving data to deliver individualized, effective feedback. The AutoCoach design includes two unique components to build an effective persuasive system. The first component is the personality classification, which recognizes drivers' personalities by analyzing driving behavior patterns and habits. At the same time, the second component is the rewarding system, which determines the current driving behavior's risk score based on some immediate past behavior. Once we have identified the driver's personality and risk scores, when to offer feedback to the driver becomes a critical question. We propose the memory factor model, which decides when to provide feedback to drivers based on their personalities. This memory factor identifies the most critical behaviors within a flexible period. AutoCoach then decides on feedback to maintain safe driving or improve awareness for risky driving habits. By applying means of persuasive technology, friendly feedback and suggestion services can be given to drivers to help them improve their behaviors. The system interacts with the users through its GUI interface that provides real-time user feedback for warnings and rewards through its simple yet effective design. To test the success of the AutoCoach v1.0 design, we have created an event identification model using data collected from 10 drivers driving different car models. Moreover, we created the personality model based on data collected from 50 drivers driving in different cities. Each driver was assigned to a personality group based on his past driving behavior. Then, we conducted a user study to evaluate the feasibility and acceptability of AutoCoach by comparing a personality-based model with a non-personality model. The purpose of the study is to understand the drivers' level of acceptance of personalized personality-based driving behavior management systems compared to non-personalized systems by considering how feedback frequency plays a role in sending more effective feedback to drivers. Evaluations used the within-subject mixed design method where 8 participants (2 from each group) participated. The study found that the frequency and timeliness of feedback play a role in creating more perceived intelligence. Lowering feedback and giving it only when truly needed reduces drivers' stress caused by frequent feedback. When we have assigned different policies to drivers, we found that riskier drivers have noticed AutoCoach more and were more grateful for the feedback offered by AutoCoach. Our study findings show that our approach is feasible and beneficial in improving the user experience when utilizing a personality-based driving agent, with 61% agreement that the personality-based model is more intelligent than the non-personality-based systems. We have upgraded the agent (AutoCoach v1.1) to improve the feedback and rewards engines when driving under different contexts. The improved version detects the road type and conditions and decides other feedback and rewarding strategies. The new improvement allowed assigning a more responsive policy to all drivers, allowing higher sensitivity to risk and more frequent feedback. We have conducted another user study requiring drivers to test different agent versions on local roads and highways to evaluate our system. From the user's point of view, v1.1 is 19% more accepted as an intelligent driving behavior management system than v1.0. Results indicate that AutoCoach successfully executes new AI algorithms and capabilities and is a promising concept that can be extended to other fields.