Klink, Pascal (2023)
Generalization and Transferability in Reinforcement Learning.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00024717
Masterarbeit, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
Reinforcement learning has proven capable of extending the applicability of machine learning to domains in which knowledge cannot be acquired from labeled examples but only via trial-and-error. Being able to solve problems with such characteristics is a crucial requirement for autonomous agents that can accomplish tasks without human intervention. However, most reinforcement learning algorithms are designed to solve exactly one task, not offering means to systematically reuse previous knowledge acquired in other problems. Motivated by insights from homotopic continuation methods, in this work we investigate approaches based on optimization- and concurrent systems theory to gain an understanding of conceptual and technical challenges of knowledge transfer in reinforcement learning domains. Building upon these findings, we present an algorithm based on contextual relative entropy policy search that allows an agent to generate a structured sequence of learning tasks that guide its learning towards a target distribution of tasks by giving it control over an otherwise hidden context distribution. The presented algorithm is evaluated on a number of robotic tasks, in which a desired system state needs to be reached, demonstrating that the proposed learning scheme helps to increase and stabilize learning performance.
Typ des Eintrags: | Masterarbeit | ||||
---|---|---|---|---|---|
Erschienen: | 2023 | ||||
Autor(en): | Klink, Pascal | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Generalization and Transferability in Reinforcement Learning | ||||
Sprache: | Englisch | ||||
Publikationsjahr: | 17 Oktober 2023 | ||||
Ort: | Darmstadt | ||||
Kollation: | iii, 54 Seiten | ||||
DOI: | 10.26083/tuprints-00024717 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/24717 | ||||
Kurzbeschreibung (Abstract): | Reinforcement learning has proven capable of extending the applicability of machine learning to domains in which knowledge cannot be acquired from labeled examples but only via trial-and-error. Being able to solve problems with such characteristics is a crucial requirement for autonomous agents that can accomplish tasks without human intervention. However, most reinforcement learning algorithms are designed to solve exactly one task, not offering means to systematically reuse previous knowledge acquired in other problems. Motivated by insights from homotopic continuation methods, in this work we investigate approaches based on optimization- and concurrent systems theory to gain an understanding of conceptual and technical challenges of knowledge transfer in reinforcement learning domains. Building upon these findings, we present an algorithm based on contextual relative entropy policy search that allows an agent to generate a structured sequence of learning tasks that guide its learning towards a target distribution of tasks by giving it control over an otherwise hidden context distribution. The presented algorithm is evaluated on a number of robotic tasks, in which a desired system state needs to be reached, demonstrating that the proposed learning scheme helps to increase and stabilize learning performance. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
Status: | Verlagsversion | ||||
URN: | urn:nbn:de:tuda-tuprints-247171 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Intelligente Autonome Systeme |
||||
TU-Projekte: | EC/H2020|640554|SKILLS4ROBOTS | ||||
Hinterlegungsdatum: | 17 Okt 2023 11:39 | ||||
Letzte Änderung: | 18 Okt 2023 08:07 | ||||
PPN: | |||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |