Back to search

Aristotle University Of Thessaloniki (AUTH)

Knowledge transfer in reinforcement learning

Abstract

dc:description

This thesis presents an original research in the field of machine learning and more specifically in the fields of reinforcement learning and transfer learning. Reinforcement learning addresses the problem of how an agent can learn an optimal behaviour through trial-and-error interactions with a dynamic environment. In the context of reinforcement learning,transfer learning refers to the process of using knowledge that has been acquired in one or more previously learned tasks, in order to enhance the learning performance in a new and usually more complex task.We specifically focus on the following research problems concerning transfer learning in reinforcement learning: a) improving data efficiency and reducing sample complexity in reinforcement learning, b) advancing transfer learning methodologies for specific types of transferred knowledge, c) transferring knowledge for significantly different tasks, d) transferring knowledge between heterogeneous agents and, e) applying reinforcement learning and transfer learning in complex and real-world problems.First, we present a novel method for transferring models to reinforcement learning agents.The models of the transition and reward functions of a task, are transferred to another relevant but different task. The proposed method introduces a hybrid agent architecture with two learning mechanisms, one for model-free learning and one for model-based learning in order to fully exploit the presence of a source task model.We then focus on policy transfer and specifically on action advice. In this setting, a teacher agent suggests actions to a student agent as the later learns a specific task. The interesting and realistic characteristics of this setting are that the teacher and the student agent may be significantly different (heterogeneous) and that the teacher may only give advice a limited number of times. The methods presented in this chapter allow such a teacher agent to produce quality advice and distribute its limited advice budget in the most effective way.We then continue exploring novel transfer learning methods for different types of transferred knowledge but we now extend the scope of the methods to that of automatic cross-domain transfer, i.e, transferring between significantly different tasks. One of the ways to enable cross-domain transfer is the use of inter-task mappings. To this end, we focus on methods that automatically discover suitable mappings in domains where human intuition cannot be applied to select them, i.e., complex domains with no domain expert, finally allowing autonomous transfer between different tasks.Finally, we present a number of novel and interesting applications of this thesis’ methods in complex and real-world problems. From transferring learning to humans using pol-icy advice to transferring knowledge between robots with different degrees of freedom and transferring between agents in complex video games, this part of the thesis demonstrates the ability of the presented methods to solve more complex TL problems.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2016

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Fachantidis, Anestis
  • Φαχαντίδης, Ανέστης

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/38511

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Fachantidis, Anestis; Φαχαντίδης, Ανέστης. Knowledge transfer in reinforcement learning. Aristotle University Of Thessaloniki (AUTH), 2016. http://hdl.handle.net/10442/hedi/38511