Abstract:

Planning in AI represents the model-based approach to control where  actions are selected to achieve goals given a predictive model of the actions and sensors. The main challenges in planning are computational, as all the models, whether accommodating  feedback and uncertainty or not, are intractable in the general case. In this talk, I'll review some the models considered in planning,  and some recent ideas that have turned out to be useful computationally.

I'll focus mainly on three types of methods from our own recent work: 1) width-based search, where we show results on a number of video games from the ALE and GVG-AI environments, 2) translations for planning with partial observability  using classical planners, and 3) multiagent planning with nested beliefs. In all cases, we seek methods that can be both general and effective. This is joint work with students and collaborators, including Blai Bonet, Nir Lipovetzky, Miquel Ramirez, Filippos Kominis, and Tomas Geffner.