Landmark-Based Heuristics and Search Control for Automated Planning

Abstract

A key characteristic of intelligence is the use of ecient problem-solving strategies when faced with unfamiliar tasks. Enabling machines to do autonomous problem-solving is thus a major milestone on the path to developing intelligent systems. Automated planning is a discipline in artificial intelligence research that studies this topic, specifically the process of automatically computing strategies for using actions to achieve a desired outcome. Given a declarative description of a task, a planning system finds an action sequence (a plan) that leads from a given initial state to a state that satisfies a specified goal description. The quality of a plan is measured via its length or, in cost-based planning, via associated costs of the actions it comprises. While the planning problem in general is computationally intractable, many planning tasks can be solved eciently due to some inherent structure of the task. Knowledge about such structure or certain properties of a planning task, so-called control knowledge, can often be extracted automatically from the problem description. This thesis makes several contributions to improve the eciency of automated planning. We focus on forward-chaining heuristic search in the state space of a planning task, currently the most widely used approach to planning. In the first part of this thesis, we detail novel methods for extracting landmarks, a particular type of control knowledge, from planning tasks. We then propose a way of using these landmarks as a heuristic estimator for judging progress during planning, and show empirically that this leads to shorter plans and allows solving more tasks in unit-cost planning. We furthermore analyse the performance gain achieved via landmarks in cost-based planning and find that landmarks can be particularly helpful in this setting, making up for the bad performance of other (cost-sensitive) heuristics. In the second part of this thesis, we focus on improving the underlying search algorithms to increase coverage (the number of tasks solved) and solution quality in planning. We conduct a detailed study of two popular search-control techniques, preferred operators and deferred evaluation, and demonstrate their respective usefulness for improving coverage and solution quality under various conditions. We also consider anytime planning to find high-quality plans given limited time. In anytime planning, the aim is to compute an initial solution quickly, and then iteratively improve on this solution while time remains. We demonstrate that the greediness that is necessary to find an initial plan quickly can impede the planning system in finding better solutions later, unless the system abandons previous eort and restarts the search. We then combine the methods analysed in the previous chapters and incorporate them into one planning system. The resulting planner LAMA, winner of the 2008 International Planning Competition, is presented in detail and compared with other state-of-the art planners. We study the interactions of various techniques employed in the system and show how much each feature contributes to the overall performance. We find that both landmarks and restarting anytime search contribute to the good performance on the set of benchmark tasks considered. Furthermore, the two techniques interact beneficially in some cases. Lastly, we provide an outlook on possible extensions of our work by investigating more complex types of landmarks. We show that using higher-order landmarks can significantly improve the heuristic estimates obtained from a landmark heuristic. However, the additional eort required for finding and using such landmarks does not necessarily pay off.