Learning to Plan in the Real World – Roni Stern, Ben Gurion University

Prof. Roni Stern, Ben Gurion University
Learning to Plan in the Real World

Planning is often referred to as the art of thinking before acting. As such, automated planning is a long-term goal of Artificial Intelligence (AI). Most planning algorithms in the literature are model-based, in the sense that they assume a compositional model of the environment is given, specifying when actions can be applied (preconditions) and what their possible outcome will be (effects). However, creating such a planning domain model for a real-world planning problem is notoriously difficult and time-consuming even for planning experts. Moreover, real-world applications are often notoriously dynamic, and thus manually creating planning domain models is not sustainable. To address these challenges, we introduce the Safe Action Model (SAM) learning family of algorithms. The input to a SAM learning algorithm are observations collected from successfully executed plans. The output is a composable domain action models that can be used by off-the-shelf automated planners. SAM learning algorithms are unique in that they output planning action models that are safe with respect to the real environment. That is, plans generated with the learned model are guaranteed to achieve their intended goal in the real environment. In this talk, I will describe several SAM learning algorithms that differ in the complexity and richness of environments they are suitable for. Notably, All SAM learning algorithms I will present run in polynomial time and have a relatively low sample complexity. Finally, I will describe how SAM learning fits into a more general agent design that enables acting in open worlds, detecting, and adapting to novelties when they are introduced. We have implemented this agent in three diverse domains: Angry Birds, Minecraft, and the balancing cartpole problem.
Bio:
Roni Stern is an Associate Professor at the department of Software and Information Systems Engineering in Ben Gurion University of the Negev. He recently returned from a prolonged sabbatical as a Principal Scientist at the Palo Alto Research Center (PARC). Roni is currently an Associate Editor for the Journal of Artificial Intelligence Research (JAIR) and the co-editor of JAIR award paper track. He also served as the president of the Symposium on Combinatorial Search (SoCS), and held various roles in the program committees of AAAI, IJCAI, ICAPS, and AAMAS.