Macro-Actions for PDDL
A Dynamic Approach
Pallabi Sree Sarker (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Sebastijan Dumančić – Mentor (TU Delft - Algorithmics)
Issa K. Hanou – Mentor (TU Delft - Algorithmics)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Automated Planning, also known as Artificial Intelligence (AI) planning is a branch of AI focused on automated decision-making and scheduling. A sub-problem within AI Planning is domain-independent planning, where we want to develop methods that are generalisable for solving planning problems in many domains. A popular modelling language for domain-independent planning is PDDL. In PDDL we model our problems as having some start state and some goal state; these states are defined by the truth-values of a set of defined predicates applied to a set of objects with corresponding types. In this work we explore the concept of dynamic macro-actions for PDDL, which are macro-actions whose utility are re-evaluated as we solve more problems, and does not require prior training. We find that dynamic macro-actions are a promising method, showing average improvements in the number of nodes explored in the search space of up to 84\% depending on the domain.