I'm interested in making intelligent agents usable.
Currently, we have agents that can filter spam, recommend movies, and help to us navigate home. In the future, agents will be more sophisticated and will deliver the information we need when we need it. Imagine a scenario in which you were to become forgetful due to age, and you found yourself standing in the kitchen. Your agent could infer that you want to make coffee and also infer that you have forgotten the next step. It could then unobtrusively remind you what to do. Such an agent would allow you to stay independent longer.
The vision I am working towards is that an intelligent agent would be more like a dog than a computer. You would train it, and over time it would learn your preferences. You and your agent could develop a mini-culture where certain phrases take on particular meanings based on shared experience.
To achieve this goal, my research focuses on a branch of machine learning that we refer to as user-oriented machine learning (UOML). Most research in machine learning focuses on improving learning accuracy. Research in UOML focuses on learning a policy for agent behavior that is understandable by a human user. And since UOML involves human users, that policy must be learned without burdening the user.
My research uses the location tracking system called Locaccino. Locaccino is a smartphone application that allows users to share their location subject to privacy constraints that depend on the requester, the user's location, and the time of day. Since it is hard for a user to predict what his or her privacy preferences will be, the goal is to use UOML to help the user specify this policy.
Two approaches we are taking are, (1) learn a set of default policies for users, and (2) learn from user feedback on the decisions from those policies.