EAPSI: Providing Smart User Feedback Based on Bayesian Models

人为气候变化是世纪最大的挑战之一，解决方案的一部分在于人们采取更可持续的行为;这导致了生态反馈设备的发展，用于水消耗，电力使用，最重要的是驾驶行为。然而，尽管有过多的生态驾驶反馈设备，但很少有人根据用户的行为和情况定制反馈。我们需要主动的情况感知反馈，比如“提前30分钟离开以避免交通堵塞”，因为与每天早上堵车时告诉司机“少用刹车”的被动反馈相比，它大大减少了排放。这项研究将建立这样一个情境感知反馈设备的框架，结合车载诊断和全球定位系统（OBD/GPS）的数据，属于新加坡科技设计大学（SUTD）的Lynette Cheah博士，交通和天气数据创建贝叶斯网络（BN）结构学习算法的训练数据。由此产生的BN将作为行为和情况特定的模型，能够确定哪些变量是燃油经济性差的根本原因。表征新加坡驾驶的系统的BN将从多个数据集的组合创建的数据集中学习。多个数据集的合并以及人类决策对数据的参与？的生成意味着训练数据集将是嘈杂的，并且充满了离群值。这项研究的目标是学习BN，它最好地描述了给定这种类型的训练数据的系统的真实性质。为了学习最佳BN，将研究基于不同数学原理的不同BN学习算法。这些不同的学习算法？的结果将与人类专家生成的BN进行比较。该奖项是由美国国家科学基金会（NSF）和新加坡国家研究基金会（NationalResearchFoundation）共同资助的东亚和太平洋夏季研究所（EastAsiaandPacificSummerInstitutes）项目下的一个奖项，旨在资助一名美国研究生的夏季研究。

Anthropogenic climate change is one of the biggest challenges of the century and part of the solution lies with people adopting more sustainable behavior; this has led to the development of eco-feedback devices for water consumption, power use, and most importantly driving behavior. However, despite the plethora of eco-driving feedback devices, very few of them tailor their feedback based on user behavior and situation. There is a need for proactive situation-aware feedback like "leave 30 minutes early to avoid traffic" because it drastically reduces emissions as compared to a reactive feedback that tells a driver "use your brakes less" every morning when they are stuck in traffic. This research will build the framework for such a situationally aware feedback device by combining on-board diagnostic and global positioning system (OBD/GPS) data, belonging to Dr. Lynette Cheah of the Singapore University of Technology and Design (SUTD), with traffic and weather data to create training data for a Bayesian Network (BN) structure learning algorithm. The resulting BN will act as a behavior and situation specific model capable of determining what variables are the root cause of poor fuel economy.The BN of the system that characterizes driving in Singapore will be learned from a dataset created from the combination of multiple datasets. The amalgamation of multiple datasets and the involvement of human decisions in that data?s generation mean that the training dataset will be noisy and filled with outliers. The goal of this research is to learn the BN that best describes the true nature of the system given this type of training data. To learn the best BN, different BN learning algorithms drawing upon different mathematical principles will be investigated. These different learning algorithm?s results will be compared to a human expert generated BN. The BN learning algorithm that most faithfully recreates the human expert generated network will be identified and the future eco-feedback device will be built off this algorithm.This award, under the East Asia and Pacific Summer Institutes program, supports summer research by a U.S. graduate student is jointly funded by NSF and the National Research Foundation of Singapore.