Automated decision making with rich representations

点击翻译按钮获取中文摘要

Computers promise to change the world in the future, even more than they have done in the past, by being able to make decisions.  What is a good decision for an agent (computer, human, or organization) depends on the agent's beliefs, its preferences and its abilities. One of the great results of 20th century science is Bayesian decision analysis: rational agents can make decisions measuring belief by probabilities and preferences by utilities. Another strand of artificial intelligence research considers rich representations of objects that are described at multiple levels of abstraction and detail, using complex ontologies to describe the vocabulary. The aim of this proposal is to allow probabilistic reasoning and decision making with rich representations of multiple objects. Thus we expect an agent to have beliefs and preferences and be able to observe and condition on multiple objects with complex relationships. This includes reasoning about the existence of objects and reasoning efficiently about sets of objects we don't need to distinguish. The challenges in this project are to find representations that are epistemically adequate (are rich enough to solve the problem), can be learned from data and/or acquired from people, and can be reasoned with efficiently. I plan to leverage the advances in probabilistic reasoning that have been made over the last 20 years, and more recent work showing, for example, how to do probabilistic reasoning without grounding. This is important as it's difficult to imagine an application that does not involve reasoning and decision making about multiple individuals.