AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


Focused Meta Actions

Authors

;

Term

4. term

Publication year

2024

Submitted on

Pages

17

Abstract

In classical planning (automated planning), the task is to find a sequence of actions that leads from a start state to a goal. To speed up this search, researchers often abstract the problem. A common tool is macros—fixed sequences of actions—but macros can be rigid and inflexible. Meta actions offer a more flexible alternative. A meta action specifies the intended effect but not the exact sequence of primitive actions that will achieve it. Earlier work has shown how to validate that a meta action can be replaced by some sequence of primitive actions. However, automatically generating good meta actions remains challenging. This thesis addresses that gap by first constructing candidate meta actions that satisfy mutex constraints (mutual-exclusion consistency) and then adding preconditions derived from exploring the state space. Experiments on benchmark domains show that the resulting meta actions are valid and can reduce search time.

I klassisk planlægning (automatisk planlægning) er opgaven at finde en rækkefølge af handlinger, der fører fra en starttilstand til et mål. For at gøre søgningen hurtigere abstraherer man ofte problemet. Et udbredt værktøj er makroer—faste handlingssekvenser—men de kan være stive og ufleksible. Meta-handlinger giver et mere fleksibelt alternativ. En meta-handling angiver den ønskede effekt, men ikke den præcise rækkefølge af primitive handlinger, der opnår den. Tidligere arbejde har vist, hvordan man kan validere, at en meta-handling kan erstattes af en sekvens af primitive handlinger. Alligevel er det svært automatisk at generere gode meta-handlinger. Denne afhandling tager fat på problemet ved først at konstruere kandidater, der overholder mutex-begrænsninger (gensidig udelukkelse), og derefter tilføje forudsætninger ud fra udforskning af tilstandsrummet. Forsøg i benchmark-domæner viser, at de genererede meta-handlinger er gyldige og kan reducere søgetiden.

[This apstract has been rewritten with the help of AI based on the project's original abstract]