Automated planning with induced qualitative models in dynamic robotic domains

Domen Šoberl (Author), Ivan Bratko (Mentor), Jure Žabkar (Co-mentor)

Abstract

Qualitative modeling enables autonomous learning agents to quickly learn generalized models from small samples of numerical data. By abstracting away certain numerical information, qualitative models provide better insights into the operating principles of a robotic system in comparison to traditional numerical models. This is of special interest to the areas of AI where experts can benefit from insights discovered by AI, or need to check whether AI's decisions comply with human common sense. Planning with qualitative models is challenging in the sense that little or no numerical information is given with the model. Generalized plans or control strategies can be devised through the means of qualitative simulation. While these can provide further insights into the possible system's behaviors, no known method is able to execute such plans without some additional trial-and-error type of numerical training. This dissertation proposes a general-purpose framework for qualitative planning in robotics domains, with a novel method to execute qualitative plans without the need for additional training. The execution adapts to the specific numerical properties of the system in real-time, and is usually successful on the first run, while its performance significantly improves on the second run. This way a working, although typically a suboptimal, solution can quickly be provided. The proposed methods are demonstrated on problem domains that have previously not been attempted qualitatively, or to a more limited extent.

Keywords

learning qualitative models;qualitative reasoning;qualitative simulation;qualitative planning;explainable control strategies;computer and information science;doctoral dissertations;

Data

Language:	English
Year of publishing:	2021
Typology:	2.08 - Doctoral Dissertation
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[D. Šoberl]
UDC:	004.8(043.3)
COBISS:	59657731
Views:	436
Downloads:	96
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	Slovenian
Secondary title:	Avtomatizirano planiranje z induciranimi kvalitativnimi modeli v dinamičnih robotskih domenah
Secondary abstract:	Kvalitativno modeliranje omogoča avtonomnim agentom hitro učenje posplošenih modelov iz majhnih naborov numeričnih podatkov. V primerjavi s klasičnimi numeričnimi modeli, kvalitativni modeli nudijo bolj jasen vpogled v principe upravljanja robotskega sistema, in sicer tako, da abstrahirajo določene numerične informacije. To je posebej zanimivo na področjih, kjer ugotovitve umetne inteligence lahko pripomorejo k razumevanju nekega probleme oz. kjer je potrebno preveriti skladnost odločitev umetne inteligence s praktičnim razumevanjem problema. Planiranje z uporabo kvalitativih modelov predstavlja poseben izziv, saj ti modeli vsebujejo malo ali nič numeričnih informacij. Z uporabo kvalitativne simulacije lahko izpeljemo posplošene plane ali strategije vodenja, ki omogočajo nadaljnje vpoglede v možna obnašanje sistema. Vendar pa ni znane metode, ki bi takšne plane lahko izvedla brez dodatnega numeričnega učenja s poskušanjem. V tej disertaciji predlagamo splošnonamenski sklop metod kvalitativnega planiranja ter izvajanja kvalitativnih planov v robotskih domenah, ki za izvajanje ne zahteva dodatnega učenja. Izvajanje se prilagodi na specifične numerične lastnosti sistema v realnem času in je običajno uspešno že prvič, ob ponovitvi pa se rezultat bistveno izboljša. Na ta način lahko hitro pridemo do delujoče, čeprav neoptimalne, rešitve. Predlagane metode demonstriramo na problemskih domenah, ki še niso bile obravnavane kvalitativno, ali pa v bolj omejenem obsegu.
Secondary keywords:	učenje kvalitativnih modelov;kvalitativno sklepanje;kvalitativna simulacija;kvalitativno planiranje;razložljive strategije vodenja;računalništvo;računalništvo in informatika;Umetna inteligenca;Disertacije;
Type (COBISS):	Doctoral dissertation
Study programme:	1000474
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	XVI, 193 str.
ID:	12795252