Inherently interpretable Q-Learning

Koumentis, Ioannis; Κουμέντης, Ιωάννης

dc.contributor.advisor	Vouros, George
dc.contributor.advisor	Βούρος, Γεώργιος
dc.contributor.author	Koumentis, Ioannis
dc.contributor.author	Κουμέντης, Ιωάννης
dc.date.accessioned	2022-12-02T10:44:11Z
dc.date.available	2022-12-02T10:44:11Z
dc.date.issued	2022-06
dc.identifier.uri	https://dione.lib.unipi.gr/xmlui/handle/unipi/14867
dc.identifier.uri	http://dx.doi.org/10.26267/unipi_dione/2289
dc.description	Not available until 01/07/2023	en
dc.format.extent	79	el
dc.language.iso	en	el
dc.publisher	Πανεπιστήμιο Πειραιώς	el
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.title	Inherently interpretable Q-Learning	el
dc.type	Master Thesis	el
dc.contributor.department	Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων	el
dc.description.abstractEN	Reinforcement Learning algorithms, especially those that utilize Deep Neural Networks (DNN), have achieved significant and many times impressive results at solving problems within a broad range of applications. Since most implementations and model architectures are based on Neural Networks (NNs), which are non-interpretable by design, there is a growing desire for Interpretable Reinforcement Learning methods development, towards improving the algorithm’s decisions tracking and increase trust, as well as cooperation between intelligent agents and human users. A promising approach towards interpretable methods includes utilizing inherently interpretable methods such as Decision Trees. This thesis investigates interpretability in Reinforcement Learning by introducing the Stochastic Gradient Trees algorithm as the baseline for developing intelligent agents. To that end, we propose model designs and training methods that utilize agents based on Stochastic Gradient Trees to perform Q-Learning and learn effective policies on several virtual environments. Moreover, a comparison of the interpretable and their counter non-interpretable methods is made under similar settings to study comparatively their efficacy in problem solving. Additionally, experiments have been conducted in a Human - AI collaboration setting, towards creating a transparent method that utilizes visual signals to improve human-agent collaboration in problem solving.	el
dc.corporate.name	National Centre for Scientific Research "Demokritos"	el
dc.contributor.master	Τεχνητή Νοημοσύνη - Artificial Intelligence	el
dc.subject.keyword	Q-Learning	el
dc.subject.keyword	Interpretability	el
dc.subject.keyword	Transparency	el
dc.subject.keyword	Human-AI collaboration	el
dc.subject.keyword	Reinforcement learning	el
dc.subject.keyword	Explainability	el
dc.subject.keyword	Stochastic Gradient Trees	el
dc.date.defense	2022-07-21

Files in this item

Files	Size	Format	View
There are no files associated with this item.

This item appears in the following Collection(s)

Τμήμα Ψηφιακών Συστημάτων
Department of Digital Systems

Show simple item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα