National and Kapodistrian University of Athens
Deep reinforcement learning algorithms for intelligent agents in virtual environments
Abstract
dc:descriptionThe objective of the present dissertation is the construction of intelligent agents, based on deep reinforcement learning methods, and their application on virtual environments spanning visual, auditory and tabular domains. In the visual domain, we applied deep reinforcement learning in conjunction with neural state abstraction to create an agent capable of navigating a maze environment based on Minecraft using only raw visual information. Furthermore, we devised an improved version of ε-greedy exploration based on adversarial neural networks and showed that a deep reinforcement learning agent using this exploration scheme can more efficiently navigate Minecraft maze environments and play a two-dimensional side-scrolling game called Super Mario Bros. In the auditory domain, we created an environment, using the Unity game engine, where a number of human speakers are present in various places within a room, and created a deep reinforcement learning agent capable of navigating the room and guiding itself towards a given speaker using only auditory information. In the tabular domain, we employed deep reinforcement learning along with a variable batch training technique to create an agent capable of achieving perfect play on a board game called L-Game, given an array of the state of the board after each player move, while requiring less data to train. Combining the audio and tabular domains, we expressed the problem of optimizing the parameters of the post-processing stack applied to the predictions of an Audio Event Detection system as a Markov Decision Process. We employed deep reinforcement learning as a way of exploring the parameter space and finding the values for each parameter which maximize the final audio event classification performance of the detector system, given an array of the detector system predictions.
Degree
thesis:*- Grantor dc:publisher
- National and Kapodistrian University of Athens
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Giannakopoulos, Petros
- Γιαννακόπουλος, Πέτρος
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53576