• View
• Edit
• History
• Print

Non-classical Heuristics For Classical Planning

Speaker: Erez Karpas http://people.csail.mit.edu/karpase/

Abstract: Domain-independent planning is one of the foundational areas in the field of Artificial Intelligence (AI). A planning task consists of an initial world state, a goal, and a set of actions for modifying the world state, with the objective of finding a plan that transforms the initial world state into a goal state. In cost-optimal planning, we are interested in finding not just any valid plan, but a cheapest such plan. One of the most prominent approaches to cost-optimal planning these days is heuristic state-space search, guided by a heuristic which estimates the distance from each state to the goal. Most heuristics for domain-independent planning are what we call classical --- they estimate the distance from some given state to the goal using only properties of the given state. In this work, we explore non-classical heuristics --- heuristics which exploit additional information gathered during search. We propose a mathematical model which allows us to formally define non-classical heuristics, as well as a useful taxonomy of heuristics along several dimensions. We then describe two different classes of non-classical heuristics: landmark-based heuristics, and machine-learning based heuristics. Our empirical evaluation shows that non-classical heuristics are not just an interesting theoretical possibility, but rather state of the art tools in heuristic search planning.