T3. Probabilistic Temporal Planning

Open slides

Presenters

Abstract

Most real world domains feature significant uncertainty that cannot be ignored when planning for a reliable, autonomous agent. Until relatively recently, methods for planning under uncertainty were limited to sequential decision-making for instantaneous actions with discrete uncertain outcomes. In this tutorial we review new advances in dealing with stochastic domains involving concurrency and durative actions. Some of these approaches extend the traditional MDP model and algorithms (leading to e.g. time-dependent MDPs, concurrent MDPs, the DUR family of algorithms, or factored policy gradient), while others (Prottle, Paragraph, Hybrid AO*, incremental contingency planning) build upon the classical planning framework. We also address the practical considerations when applying these algorithms and describe approaches that avoid modeling full-blown stochastic uncertainty to gain speed and scalability.

This tutorial is intended both for planning researchers as well as other AI researchers who are interested in applying automated planning in domains involving uncertainty. We do not assume any knowledge of planning under uncertainty and will review the basics before discussing the advanced techniques.