context-dependent value of information and cost-sensitive learning for situational awareness on smart phones

智能手机上的上下文信息可用于为特定情况定制应用程序（例如，根据位置、汽油价格和交通情况提供定制的路线建议）。然而，典型的上下文感知智能手机应用程序使用非常有限的上下文信息，如用户身份、位置和时间。随着智能手机越来越多地配备传感器，并且具有越来越强的查询本地环境以获取额外信息的能力，信息的优先级问题变得重要。这个问题涉及决定从哪个传感器收集信息，以便在给定的环境中最好地适应用户的需求和偏好。这些决定对智能手机很重要，因为它们会影响所需的处理量，从而影响设备成本和电池寿命。我们的研究从两个方面来解决这个问题。首先，我们将现有的仅基于信息价值的传感器选择机制应用于智能手机应用。这些现有的方法有很好的研究解决方案，但还没有具体应用于这个领域。其次，我们研究了决策过程中传感器选择的开放问题，其中观测特征的选择是基于对智能决策的长期影响。第二个问题更具挑战性，因为信息收集可能对未来某个时刻做出的决策有用，而对信息的效用进行预测是一个未解决的问题。例如，在用户外出吃午餐时收集晚餐产品的定价信息可能是值得的，因为这一天用户的日程安排几乎没有时间在晚餐时间购买。我们的研究已经建立了马尔可夫决策过程中观察聚合的机制，我们计划通过使用这些聚合观察作为信息潜在效用的指示来扩展这项工作。这项研究将影响许多使用基于大量传感器数据的自动化决策的领域。

Context information on a smart phone can be used to tailor applications for specific situations (e.g. provide tailored routing advice based on location, gas prices and traffic). However, typical context-aware smart phone applications use very limited context information suc as user identity, location and time. As smart phones become increasingly outfitted with sensors, and have increasing ability to query the local environment for additional information, the problem of prioritization of information becomes important. This problem involves deciding which sensors to gather information from in order to best accommodate user needs and preferences in a given context. These decisions are important on smart phones, as they impact the amount of processing necessary, and thus the device cost and battery life. Our research approaches this problem in two ways. First, we apply existing sensor selection mechanisms, based soley on the value of information, to smart phone applications. These existing methods have well-studied solutions, but have not been applied specifically in this domain. Second, we study the open problem of sensor selection within decision-making processes, in which observational features are selected based on long-term impact on the decisions made by the smart. This second problem is more challenging, because information gathering can be useful for decisions made at some point in the future, and making predictions about the utility of information is an unsolved problem. For example, it may be worthwhile to gather pricing information for dinner items while the user is out for lunch, on a day when the user's schedule leaves little time for purchases at dinner time. Our research has established a mechanism for observation aggregation in Markov decision processes, and we plan to extend this work by using these aggregate observations as indications of potential utility of information. This research will impact a number of fields which use automated decision making based on large quantities of sensor data.