Learning Bad Action Choices in Automatic Planning

Bc. Michal Krajňanský

Master's thesis

Abstract:

Řada moderních metod automatického plánování je založna na prohledávání stavového prostoru vedeném doménově nezávislou heuristickou funkcí. Často je ale snadné nalézt plánovací domény, ve kterých heuristická funkce není dostatečně informativní. Cílem práce je navrhnout a vyhodnotit metodu učení špatných akcí v automatickém plánování.

Abstract:

Abstract A number of current state-of-the-art methods of automatic planning are based on forward state-space search driven by domain-independent heuristic functions. There are, however, domains where these heuristics provide insufficient guidance. Abstract Given a planning instance from some domain, classical planning methods treats it as if the domain was seen for the first time. By contrast, in practice planners are typically required to solve instances from the same domain over and over again. Hence it is natural to attempt to learn from that experience. One of the most studied approaches is learning domain control knowledge in the form of reactive policies, which define `good' action choices. It turns out that good action choices are quite hard to learn and relying on the learned policies often leads to planning failure. Abstract In contrast, some of the action choices are obviously bad. The goal of the work is to evaluate the potential of using pruning rules to augment the heuristic forward state-space search planning.

Keywords

automatic planning, heuristic search, search space pruning, inductive learning, bad action choices

Language used: English

Date on which the thesis was submitted / produced: 8. 1. 2014

Thesis defence

• Date of defence: 10. 2. 2014
• Supervisor: doc. RNDr. Aleš Horák, Ph.D.
• Reader: Dr. Olivier Buffet

Masaryk University

Faculty of Informatics

Master programme / field:
Informatics / Artificial Intelligence and Natural Language Processing